Optimization method for cognitive disease nursing area

Abstract

The invention discloses an optimization method for a cognitive disease nursing area, relates to the technical field of nursing and intelligent perception for the aged, and particularly discloses an optimization method for constructing a special nursing area for the old with cognitive diseases in a nursing institution for the aged. According to the method, a closed-loop information acquisition and output system consisting of five parts, namely a context embedded lattice method (CAGM), a multi-domain sensing-sensing coupling body (PCN), a heterogeneous sensing common-standard system (HSCM), a cognitive pressure load distribution spectrum (CLFS) and a care structure sub-element set (CSES), is constructed, wherein the five parts are as follows: the context embedded lattice method (CAGM), the multi-domain sensing-sensing coupling body (PCN), the heterogeneous sensing common-standard system (HSCM), the cognitive pressure load distribution spectrum (CLFS) and the care structure sub-element set (CSES). Semantic gridding division is carried out on a nursing space, multi-modal sensing nodes are deployed to collect environment dynamic data, and semantic tags of nursing personnel, experts and an AI model are fused, so that recognition of susceptible areas of the cognitive old people and quantitative modeling of cognitive risks are realized. And finally, outputting a structured design suggestion sub-element set, and directly supporting scientific layout and intelligent transformation of the nursing special area.

Description

Technical Field

[0001] This invention relates to the field of elderly care and intelligent sensing technology, and specifically discloses an optimized method for constructing a care area for elderly people with dementia in elderly care institutions. Background Technology

[0002] With the increasing global trend of population aging, the care needs of elderly people with dementia (such as Alzheimer's disease) are growing. When constructing dementia care areas, elderly care institutions need to not only meet basic living care functions, but also create a safe, clear, stable, and cognitively low-burden care environment that addresses the perceptual impairment, spatial disorientation, and behavioral disturbances of elderly people with dementia. Therefore, the scientific and systematic collection and analysis of on-site hardware and software information in elderly care institutions has become an important foundation for the design of care spaces.

[0003] Currently, commonly used design research methods in the industry mainly include: recording physical structure information through manual site visits, analyzing functional layouts using 2D CAD drawings, collecting opinions from caregivers through questionnaires or interviews, or conducting simple video sampling based on fixed cameras. These methods have several shortcomings in practical application: First, the information collection process relies on manual experience and lacks standardized, intelligent tools; second, the collected data is mainly static physical parameters, unable to dynamically capture environmental disturbances, behavioral conflicts, or changes in perceived load; third, current information collection is isolated from risk assessment, functional adaptation, and design output, failing to form an effective linkage, resulting in unstable design outcomes and poor age-friendliness.

[0004] The closest technical solution to this invention is the "intelligent monitoring + expert evaluation" approach introduced in some high-end health and wellness institutions. This involves deploying a small number of sensors (such as infrared sensors or cameras) in key areas to collect trajectory information, and then adjusting the layout based on expert on-site diagnostic suggestions. However, such solutions mostly focus on single-point behavior collection and risk identification, and have not yet formed a complete closed-loop system of multi-source perception, multi-dimensional fusion, and structured output. They also lack semantic understanding capabilities, spatial cognitive modeling capabilities, and direct guidance capabilities for care design.

[0005] Therefore, there is an urgent need for a systematic, intelligent, and multi-source integrated information collection method that can cover the information needs of the entire process of building a dementia care zone, and improve the scientific nature, accuracy, and feasibility of the design.

[0006] How can we establish an intelligent data support system capable of comprehensively, efficiently, and systematically collecting on-site hardware and software information and quantitatively assessing cognitive risks during the spatial planning and design of dementia care areas in actual elderly care facilities? Current technologies often rely solely on architectural drawings, staff experience, and localized sensory data, failing to systematically perceive and quantify multi-dimensional factors such as behavioral flow, environmental disturbances, visual stimuli, path guidance, and semantic cognitive adaptability within the care area. Consequently, they cannot create a cognitively friendly care environment for dementia patients. Therefore, in actual renovation processes, problems frequently arise such as "functional compliance but elderly people easily get lost," "smooth paths but frequent behavioral conflicts," and "aesthetically pleasing design but excessive cognitive load." Summary of the Invention

[0007] This invention addresses the aforementioned problems by providing an optimized method for dementia care areas.

[0008] This method constructs a closed-loop information acquisition and output system consisting of five parts: Contextual Embedding Grid Method (CAGM), Multi-Domain Perceptual Coupled Module (PCN), Heterogeneous Perceptual Common Labeling System (HSCM), Cognitive Load Distribution Spectrum (CLFS), and Care Structure Subset (CSES). By semantically gridding the care space, deploying multimodal sensing nodes to collect dynamic environmental data, and integrating semantic labels from caregivers, experts, and AI models, it achieves the identification of vulnerable areas for dementia patients and the quantitative modeling of cognitive risks. Finally, it outputs a structured design suggestion subset, directly supporting the scientific layout and intelligent transformation of care areas. This invention effectively improves the data foundation and intelligent decision-making capabilities of cognitive care space design, possessing advantages such as high intelligence, strong adaptability, and good visualization, and is suitable for the design and planning of dementia care areas in newly built, renovated, and mixed-use elderly care institutions.

[0009] To achieve the above objectives, the present invention adopts the following technical solution, which includes the following steps:

[0010] Step 1: Construct the spatial coordinates of the care area and adapt the spatial coordinate grid to care behaviors;

[0011] Step 2: Deploy multi-mode sensing nodes in the high-adaptability coordinate grid area of ​​Step 1 to collect environmental disturbance factors in real time using multi-mode methods, and build a sensing model and a heterogeneous sensing co-calibration system (HSCM).

[0012] Step 3: Generate a "cognitive load distribution spectrum" for the spatial coordinate grid;

[0013] Step 4: Integrate spatial coordinates, perception model and heterogeneous perception co-calibration system, and "cognitive load distribution spectrum" into "care structure sub-element".

[0014] As a preferred embodiment, in step 1 of this invention, the care area is a two-dimensional closed space S. The space S is divided into M rows and N columns of grid cells with a fixed resolution, denoted as G(i,j), where i∈[1,M] and j∈[1,N]. Each grid cell represents a minimum evaluation unit, used to carry out subsequent behavioral adaptation and environmental assessment calculations.

[0015] The set of dementia care behaviors is B = {B_1, B_2, ..., B_m}, and each behavior is pre-defined to occur at a location O_k = (x_k, y_k).

[0016] For each grid G(i,j), calculate its fit score for all care behaviors, defined as follows:

[0017] Formula 1: Caregiver Behavior Fit Function

[0018] Q_ij=∑{k=1 to m}[(P_k(i,j)*W_k) / (1+D_k(i,j)^2)]

[0019] The explanation is as follows:

[0020] Q_ij: represents the overall fitness score of the (i,j)th grid for all behaviors;

[0021] m: Total number of caregiving behaviors;

[0022] P_k(i,j): Potential fitness score of behavior B_k performed in grid (i,j);

[0023] W_k: Importance weight of behavior B_k;

[0024] D_k(i,j): The Euclidean distance between the preset occurrence position O_k of behavior B_k and the grid (i,j), calculated as follows:

[0025] Formula 2: Euclidean distance calculation

[0026] D_k(i,j)=sqrt((i-x_k)^2+(j-y_k)^2)

[0027] in:

[0028] x_k, y_k: The center of the preset position of the k-th line;

[0029] i,j: The coordinate indices of the current grid in the spatial partitioning;

[0030] The Q_ij values ​​of all grids are visualized and output; the color intensity represents the behavioral load and adaptation density of the area, which is used to indicate which areas need to be replanned.

[0031] As another preferred embodiment, the present invention pre-determines the occurrence location O_k = (x_k, y_k) for each behavior. By constructing a spatial coordinate grid and a scoring model for behavior adaptability, the adaptability of each spatial unit to different behaviors is quantified. The adaptation function is calculated based on environmental factors, behavioral requirements, and accessibility indicators, and the region with the highest score is selected as O_k. At the same time, combined with the data collected by the perception system in real-world scenarios, the pre-determined occurrence location is further verified or adjusted using historical behavior trajectories and behavior hotspot distributions to make it closer to actual needs.

[0032] As another preferred embodiment, the present invention describes:

[0033] B_1: Bathing assistance

[0034] B_2: Toileting care

[0035] B_3: Dining Guidance

[0036] B_4: Night Patrol ...

[0038] B_m: Other customized behaviors (such as group activities).

[0039] As another preferred embodiment, the potential adaptability score of the present invention ranges from 0 to 1.

[0040] As another preferred solution, the method for obtaining P_k(i,j) in this invention is as follows: AI training is performed through historical care path trajectories, infrared records, and nurse check-in data to generate a behavior heatmap, and the high-frequency behavior areas are mapped to P_k(i,j); and P_k(i,j) is dynamically adjusted through environmental resistance factors.

[0041] As another preferred embodiment, the specific method for generating behavior heatmaps through AI training according to the present invention is as follows:

[0042] a. Input data construction

[0043] Collect the following multi-source historical care data:

[0044] Care path trajectory data: a sequence of time coordinates from devices worn by nurses (such as RFID, Bluetooth tags, and indoor positioning);

[0045] Infrared sensor image sequences: used to sense nighttime activities and behavioral distribution in low-light areas;

[0046] Nurse attendance data includes attendance time, attendance location coordinates, and behavior annotations (such as "turning over", "taking a bath", etc.).

[0047] b. Feature extraction and encoding methods

[0048] All raw behavioral events are transformed into a "spatial grid activity frequency matrix", which is the statistical frequency of behavior B_k corresponding to each grid (i,j);

[0049] Using the time sliding window method, N heat map samples are generated, each of which is a grid distribution map of row B_k within a time period;

[0050] Use CNN (Convolutional Neural Network) to extract spatial pattern features from spatial heatmaps;

[0051] We combined LSTM modules to model the temporal transition patterns of care behaviors.

[0052] c. Model training methods

[0053] Construct a multi-class behavior prediction model, with labels representing the regions where the actual behavior occurred;

[0054] The loss function is either multi-class cross-entropy or spatial clustering similarity index;

[0055] The confidence level of the region where the behavior occurs is used as the output behavior heatmap, and the predicted value on each grid (i,j) is the initial P_k(i,j) value;

[0056] All P_k(i,j) values ​​are normalized to the interval [0,1].

[0057] As another preferred embodiment, the dynamic adjustment mechanism of the environmental resistance factor described in this invention is as follows:

[0058] a. Acquisition of environmental factors

[0059] The following environmental metrics are collected at each grid-deployed or associated sensing node:

[0060] Noise intensity (N_ij)

[0061] Temperature and humidity levels (T_ij)

[0062] Air circulation (A_ij)

[0063] Traffic congestion level (C_ij)

[0064] b. Calculation of resistance factor

[0065] Define a weighted environmental resistance function:

[0066] R_ij=α1*N_ij+α2*|T_ij-T_opt|+α3*(1-A_ij)+α4*C_ij

[0067] T_opt represents the optimal temperature for the behavior (e.g., 26℃~28℃ for bathing).

[0068] Each α coefficient is set based on experience or expert knowledge.

[0069] R_ij∈[0,1]

[0070] c. P_k(i,j) dynamic adjustment method

[0071] The final score is calculated as follows:

[0072] P_k(i,j)_adjusted=P_k(i,j)*(1-R_ij)

[0073] The larger the resistance factor, the less suitable the current environment is for the execution of this behavior, and the corresponding P_k(i,j) will be suppressed;

[0074] The system automatically updates environmental data and refreshes the P_k(i,j) value at regular intervals (e.g., every 10 minutes).

[0075] As another preferred embodiment, the setting method of W_k in this invention is as follows: it is set according to the complexity of care and the risk level (e.g., if the risk of bathing is high, W_k = 0.9; if dining can be guided, W_k = 0.3); the configuration strategy matrix W = [W_1,...,W_m] is dynamically adjusted on the front-end interface.

[0076] As another preferred embodiment, the correspondence between the care complexity and risk level and W_k described in this invention is as follows:

[0077] Care complexity C_k∈[0,1]: represents the skill requirements, time consumption, and human resources required to perform the action;

[0078] Risk level R_k∈[0,1]: indicates the degree of harm that may be caused if the behavior is not performed properly or the environment is unfavorable (such as falls, burns, etc.);

[0079] Formula for calculating weight W_k:

[0080] W_k=α*C_k+β*R_k

[0081] Here, α and β are weighting coefficients, commonly set to α = 0.4 and β = 0.6, i.e., risk-prioritizing. These can be adjusted flexibly according to the system strategy.

[0082] As another preferred embodiment, the strategy matrix W = [W_1,...,W_m] described in this invention is dynamically adjusted on the front-end interface as follows:

[0083] a. Interface Settings

[0084] On the system's front-end management interface (Web or touchscreen), the following fields are displayed for each care behavior B_k:

[0085] Name of the action (e.g., "taking a bath", "using the toilet");

[0086] The current complexity is C_k slider (range 0.0 to 1.0);

[0087] Current risk level R_k slider (range 0.0~1.0);

[0088] The result W_k is automatically calculated and displayed;

[0089] Whenever the user adjusts C_k or R_k, the system automatically updates the displayed value of W_k in real time according to the formula;

[0090] b. Dynamic interaction implementation method (technical implementation)

[0091] Front-end language: Use HTML + JavaScript (such as Vue.js or React) to build slider components;

[0092] Data binding: The slider component is bound to the calculation function in real time, and adjustments trigger an update of W_k.

[0093] Calculation logic:

[0094] function calculateWk(Ck, Rk){

[0095] return(0.4*Ck+0.6*Rk).toFixed(2);

[0096] }

[0097] Synchronize with the backend:

[0098] When the user clicks the "Save" button, the current W array is sent to the server via AJAX;

[0099] The backend updates the strategy matrix to the configuration database (such as SQLite or MongoDB);

[0100] The system will automatically use the updated W array for the next behavior adaptation calculation.

[0101] As another preferred embodiment, the environmental resistance factor of the present invention includes noise and obstacles.

[0102] As another preferred embodiment, step 2 of the present invention deploys multi-mode sensing nodes (PCNs) in the high Q_ij value region. Each PCN periodically collects the following five types of data streams within its deployment area: image and video streams, audio sound field intensity and frequency distribution, air composition and odor changes, dynamic changes in illumination, and temperature and humidity environmental changes.

[0103] As another preferred embodiment, the Q_ij threshold setting in this invention adopts the normalized threshold method or the quantile threshold method;

[0104] Normalized threshold method:

[0105] The calculation result of Q_ij is normalized and the value range is [0,1], with a fixed threshold set;

[0106] The region where Q_ij ≥ 0.7 is designated as the "high Q_ij value region";

[0107] The range 0.4 ≤ Q_ij < 0.7 is considered a "medium fit region".

[0108] Q_ij < 0.4 indicates a "low-fitness region" and deployments should be ignored.

[0109] Quantile-Based Thresholding:

[0110] To adapt to the differences in structure and behavior distribution in different care areas, a statistical distribution method is used to dynamically determine the "high Q_ij area";

[0111] Calculate the 80th percentile of all grid Q_ij values, denoted as Q_thr;

[0112] The grids that satisfy Q_ij≥Q_thr are designated as "high Q_ij regions";

[0113] Adjust the system configuration parameters to the percentile (e.g., 75%, 85%, 90%) to control the number of deployments.

[0114] As another preferred embodiment, the PCN node k in this invention collects a set of perceptual feature vectors at time t:

[0115] PCN_k(t)={V_k(t),A_k(t),G_k(t),E_k(t),T_k(t)}

[0116] in:

[0117] V_k(t): Image motion perturbation feature vector;

[0118] A_k(t): Audio spectrum feature vector;

[0119] G_k(t): Illumination distribution characteristic value;

[0120] E_k(t): Estimated odor intensity (ppm);

[0121] T_k(t): Environmental values ​​in both temperature and humidity dimensions.

[0122] As another preferred embodiment, in step 2 of the present invention, the perception model includes image perturbation complexity entropy and composite stimulus load function;

[0123] Image perturbation complexity entropy:

[0124] S_k(t)=-∑{n=1 to N}[(F_n(t) / ∑F_n(t))*log2(F_n(t) / ∑F_n(t))]

[0125] in:

[0126] N: The number of moving object categories detected in the image (e.g., people, curtains, lamp shadows, etc.);

[0127] F_n(t): The magnitude of the perturbation characteristic change of the nth type of moving target at time t, which can be obtained by continuous frame difference;

[0128] ∑F_n(t): Total disturbance intensity of all target categories at time t;

[0129] S_k(t): Visual perturbation entropy of node k at time t. The larger the value, the more complex the visual stimulus, which is more likely to cause excessive cognitive load.

[0130] Composite stimulus loading function:

[0131]

[0132] in:

[0133] R_k(t): The comprehensive environmental load score of node k at time t;

[0134] S_k(t): Visual perturbation entropy (see above);

[0135] σ_a(t): Standard deviation of the audio power spectrum at time t (indicating the degree of non-uniformity of the sound field);

[0136] ΔG_k(t): The difference in light intensity over one minute (representing the light variation frequency);

[0137] E_k(t): Odor concentration value (e.g., volatile matter);

[0138] Temperature gradient (e.g., the cold air zone in an air conditioner);

[0139] γ1~γ5: These are weighting factors that can be adjusted using training data to reflect the importance of each sensory channel.

[0140] As another preferred embodiment, in step 2 of this invention, constructing a heterogeneous sensing and common standardization system (HSCM) specifically involves:

[0141] The state-aware label for region (i,j) at time t is:

[0142] HSCM(i,j,t)={L_h,L_e,L_m}

[0143] in:

[0144] L_h: Labels added by caregivers, such as "difficult to pass through" or "noisy";

[0145] L_e: Caregivers label their patients, such as “discontinuous pathways” or “poor behavior orientation”;

[0146] L_m: AI model identifies tags, such as "chaotic crowd flow" or "abnormal lingering area".

[0147] All the above tags are standardized into vector form, in the following format:

[0148] L_x = {l_1,l_2,...,l_n}

[0149] Where l_k∈{0,1} indicates whether the k-th type of risk or warning label is included.

[0150] The semantic similarity sim(a,b) between label vectors is:

[0151] sim(a,b)=(a·b) / (||a||*||b||)

[0152] in:

[0153] a, b: Two label vectors;

[0154] a·b: the vector dot product of a and b, i.e. the number of common labels between them;

[0155] ||a||: the Euclidean norm of a, i.e., sqrt(∑a_k^2);

[0156] sim(a,b): The value range is [0,1]. The closer it is to 1, the more similar the two are.

[0157] After obtaining the three-source similarity, the fusion confidence score F_conf is calculated to generate a unified label output, as shown in the following formula:

[0158] Three-source collaborative fusion confidence function

[0159] F_conf=tanh[(λ1*sim(L_h,L_e)+λ2*sim(L_h,L_m)+λ3*sim(L_e,L_m)) / (1+θ*E_dev)]

[0160] in:

[0161] F_conf: Confidence value of the final fused label;

[0162] sim(L_h,L_e), sim(L_h,L_m), sim(L_e,L_m): Three-party label similarity;

[0163] λ1, λ2, λ3: Adjustable weighting coefficients, representing the contributions of different subjects to the fusion result;

[0164] E_dev: The degree of coordination bias of the three-party labels is as follows;

[0165] θ: Penalty intensity factor;

[0166] tanh(): Hyperbolic tangent normalization function, which compresses the output value to (-1,1), and is often used to sense and suppress extreme value fluctuations.

[0167] Formula for calculating the degree of coordination deviation E_dev

[0168] E_dev=(1-sim(L_h,L_e))+(1-sim(L_h,L_m))+(1-sim(L_e,L_m))

[0169] explain:

[0170] E_dev reflects the degree of label divergence among the three parties; a larger value indicates greater divergence.

[0171] In F_conf, it is used as an adjustment factor in the denominator of the penalty term to ensure that high divergence will significantly suppress the fusion confidence value.

[0172] As another preferred embodiment, in step 3 of this invention, generating the "cognitive load distribution spectrum" specifically involves:

[0173] The cognitive stress score for each grid cell (i,j) is denoted as:

[0174] Formula 1: Cognitive stress rating function

[0175] CLFS(i,j)=β1*B_dens(i,j)+β2*E_disp(i,j)+β3*P_blur(i,j)+β4*V_discord(i,j)

[0176] in:

[0177] CLFS(i,j): The overall cognitive load value of region (i,j);

[0178] B_dens(i,j): The overlap density of the behavior trajectories in this grid;

[0179] E_disp(i,j): Standard deviation of noise disturbance distribution in this region;

[0180] P_blur(i,j): The inverse ratio of path sharpness;

[0181] V_discord(i,j): Visual stimulus consistency deviation index;

[0182] β1, β2, β3, β4: These are the weight coefficients of the four indicators, which are set by data training.

[0183] Formula 2: Calculation of Behavioral Overlap Density

[0184] B_dens(i,j)=∑{k=1 to m}F_k(i,j) / T_window

[0185] in:

[0186] F_k(i,j): The number of times the k-th type of care behavior occurs in the (i,j) grid within the T_window time;

[0187] m: Total number of behavior types (e.g., bathing, patrolling, taking medication, eating, etc.);

[0188] T_window: Observation time window (in seconds or minutes);

[0189] Formula 3: Calculation of environmental noise disturbance

[0190] E_disp(i,j)=sqrt(∑{t=1 to N}(A_t(i,j)-μ_A(i,j))^2 / N)

[0191] in:

[0192] A_t(i,j): The sound pressure level of grid (i,j) at time t;

[0193] μ_A(i,j): The average sound pressure level of this grid;

[0194] N: Number of sampling times;

[0195] E_disp(i,j): Noise disturbance fluctuation in this region;

[0196] Formula 4: Path Ambiguity Calculation

[0197]

[0198] in:

[0199] The average value of the gradient changes along all path directions within the grid region (i,j);

[0200] P_blur(i,j): Path blur score, the higher the value, the more blurry the path;

[0201] Formula 5: Calculation of Visual Stimulus Bias

[0202] V_discord(i,j)=σ_color(i,j)+σ_texture(i,j)+|ΔL(i,j)|

[0203] in:

[0204] σ_color(i,j): The standard deviation of the color histogram in this region;

[0205] σ_texture(i,j): Standard deviation of the texture orientation histogram;

[0206] ΔL(i,j): The absolute value of the change in light intensity between two consecutive frames.

[0207] As another preferred embodiment, step 4 of the present invention specifically comprises:

[0208] For each functional area Z_n, its output structure is defined as follows:

[0209] CSES(Z_n) = {

[0210] F_phys(Z_n),

[0211] F_env(Z_n),

[0212] F_cog(Z_n),

[0213] F_risk(Z_n),

[0214] F_layout(Z_n)

[0215] }

[0216] The meanings of each field are as follows:

[0217] F_phys(Z_n): Physical space structure parameters (geometry, dimensions, structural boundaries, etc.);

[0218] F_env(Z_n): Environmental state parameters (mean and fluctuation of temperature, humidity, light intensity, noise, etc.);

[0219] F_cog(Z_n): Cognitive risk parameter (from CLFS);

[0220] F_risk(Z_n): Historical records and annotations of anomalous events (such as falls, wandering, etc.);

[0221] F_layout(Z_n): The system's recommended layout adjustment suggestions and interpretability score.

[0222] Layout_opt(Z_n)=ψ1*CLFS_score(Z_n)+ψ2*HSCM_disagreement_rate(Z_n)+ψ3*E_dev_zone(Z_n)

[0223] in:

[0224] Layout_opt(Z_n): Layout optimization suggestions for the nth functional area;

[0225] CLFS_score(Z_n): Average cognitive load score for zone n;

[0226] HSCM_disagreement_rate(Z_n): The proportion of disagreement among the three-party labels, reflecting information conflict;

[0227] E_dev_zone(Z_n): Abnormal fluctuation of sensing nodes within this region;

[0228] ψ1, ψ2, ψ3: Empirical weighting adjustment factors, used to balance the contributions of various factors.

[0229] CLFS_score(Z_n)=∑{i,j∈Z_n}CLFS(i,j) / N_grid(Z_n)

[0230] in:

[0231] CLFS(i,j): Cognitive load score for the grid;

[0232] N_grid(Z_n): The number of grids contained in region Z_n;

[0233] HSCM_disagreement_rate(Z_n)=N_disagree(Z_n) / N_total(Z_n)

[0234] in:

[0235] N_disagree(Z_n): The number of tag entries where the maximum similarity among the three source tags is less than δ;

[0236] N_total(Z_n): The total number of label entries for all samples collected in this area;

[0237] δ: Set a threshold (e.g., 0.5); values ​​below this threshold are considered significant divergences.

[0238] E_dev_zone(Z_n)=(σ_temp+σ_hum+σ_noise+σ_light) / 4

[0239] in:

[0240] σ_temp, σ_hum, σ_noise, and σ_light represent the standard deviations of temperature, humidity, noise, and light intensity in the region, respectively.

[0241] The beneficial effects of this invention are as follows: This invention constructs a five-step closed-loop method of "semantic space modeling - multi-modal perception acquisition - three-source label fusion - cognitive load modeling - structured sub-element output", which realizes intelligent support for the entire process from data acquisition, semantic understanding, risk assessment to the generation of executable suggestions. It solves the key bottlenecks of "data gaps, cognitive gaps, and unsolvable structures" in traditional methods, and significantly improves the scientificity, adaptability and intelligence level of cognitive care area construction.

[0242] Step 1 of this invention constructs the spatial coordinates of the care area and adapts the spatial coordinate grid to care behaviors; this forms the basic geometric and functional framework of the entire care space, dividing the space into quantifiable and identifiable grid units. The grid is the basic building block of the "care structure sub-element," defining the spatial location and boundary conditions of the sub-element.

[0243] Step 2 of this invention involves deploying multimodal sensing nodes in a highly adaptable area to construct an HSCM system; and deploying multimodal sensors (such as infrared, video, and sound sensors) in key areas of the care area to collect behavioral and environmental data in real time, and establishing a perception model and heterogeneous co-labeling system (HSCM). The HSCM system provides semantic labels for behavioral perception and environmental disturbances, providing semantic dimensions and data-driven basis for the functional attribute labeling and dynamic adjustment of the "care structure sub-set".

[0244] Step 3 of this invention: Generate a cognitive stress distribution spectrum; for each sub-unit in the spatial grid, calculate its "cognitive stress value" based on perceived data and care behavior intensity, and form a spatial distribution spectrum. This spectrum provides key indicators for the functional classification and priority ranking of the "care structure sub-unit set," guiding which areas should be designated as high-frequency care units, buffer transition units, quiet recovery units, etc.

[0245] Step 4 of this invention: Output a set of care structure sub-elements; divide the entire care area into several structural sub-elements with specific functional attributes, spatial locations, and behavioral adaptability, forming a composite unit set of space-behavior-perception. This is a comprehensive result of the spatial division (Step 1), perceptual annotation (Step 2), and cognitive load index (Step 3) generated in the preceding steps. Step 4 is an integrated output step that integrates the spatial coordinate system of Step 1, the perceptual and semantic information of Step 2, and the cognitive load index of Step 3 into a set of functionally defined and dynamically adaptable "care structure sub-elements," thereby achieving scientific layout, behavioral adaptation, and dynamic optimization of the care area for dementia patients. Therefore, Step 4 is not an independent action, but rather a summary and structural expression of the output of the first three steps.

[0246] To address the lack of systematic and intelligent information collection and cognitive risk assessment support during the construction of dementia care zones, this invention constructs a closed-loop, data-driven design support system:

[0247] Innovation Point 1: Contextual Embedded Grid Method (CAGM). This method provides the foundation for spatial semantic structure, serving as the spatial coordinate and semantic anchor framework for the entire system. Breaking away from traditional physical grid division, it introduces for the first time a dual-dimensional anchor model of "behavioral semantics + spatial distribution," achieving highly adaptable annotation between spatial regions and care behaviors, thus providing a semantic basis for information collection and behavioral analysis.

[0248] Innovation Point Two: Multi-Domain Sensing Coupler (PCN). Based on the output positioning of CAGM, sensor nodes are deployed to achieve real-time multi-modal acquisition of environmental disturbance factors; by constructing an integrated multi-modal sensing node, environmental factors such as images, sounds, light, odors, temperature and humidity are fused to form a dynamic coupling sensing mechanism, solving the problem that traditional static assessment cannot capture cognitive load triggers.

[0249] Innovation Point 3: Heterogeneous Perception and Co-labeling System (HSCM). This system fuses three-source semantic annotations into PCN-collected results, outputting consistent and reliable labels. It proposes a three-source fusion annotation mechanism involving nursing staff, experts, and AI models, and introduces semantic similarity and difference penalty functions to achieve consistent and reliable semantic label judgment and fusion, avoiding bias caused by single subjective judgments.

[0250] Innovation Point 4: Cognitive Load Distribution Spectrum (CLFS). Based on HSCM and PCN data, a spatial cognitive risk heatmap is calculated; a multi-factor cognitive load scoring model for dementia patients is constructed, which integrates factors such as behavioral density, path ambiguity, visual consistency, and environmental disturbance, and for the first time achieves a quantitative assessment of the cognitive burden in caregiver space.

[0251] Innovation Point Five: Care Structured Subset (CSES) Output Mechanism. This mechanism integrates all results from the previous four steps to generate structured output, directly supporting design optimization and implementation. All collected and calculated results are structured into multi-dimensional care design elements, outputting standardized design suggestions and layout scores. This truly connects the entire "collection-analysis-output-execution" chain, solving the problem of data and design disconnect.

[0252] Compared to the more advanced "intelligent monitoring + expert evaluation" model in the industry, which involves deploying infrared or cameras in key areas to collect behavioral data and combining it with expert subjective diagnosis to identify risks and make suggestions for space modification, this invention has achieved substantial breakthroughs and system-level improvements at several key levels.

[0253] First, in terms of information acquisition, existing technologies rely solely on video behavior detection, resulting in a single data dimension and a lack of ability to identify non-visual factors such as noise, lighting, and odor. In contrast, the "Multi-Domain Sensory Coupler (PCN)" constructed in this invention fuses multi-source data, including visual, auditory, and environmental disturbance data, to achieve comprehensive perception of cognitive stimuli.

[0254] Secondly, regarding label judgment mechanisms, existing solutions heavily rely on expert experience, are highly subjective, and are easily affected by human limitations. This invention proposes a "Heterogeneous Comparison and Labeling System (HSCM)," which significantly improves label consistency and decision credibility through the fusion of labels from caregivers, experts, and AI models, along with a similarity reconciliation algorithm.

[0255] In terms of risk modeling methods, most existing technologies remain at the level of risk "warnings" and are difficult to quantify and assess cognitive load. However, the "Cognitive Load Distribution Spectrum (CLFS)" proposed in this invention constructs a scoring model with multidimensional factors, and for the first time realizes a quantitative expression of the load for elderly people with dementia, with clear assessment indicators and risk level output.

[0256] Finally, at the implementation level, the "Care Structure Subset (CSES)" constructed by this invention can output the collected and analyzed results as structured design parameters, which are compatible with CAD / BIM systems and can be used directly, filling the industry gap of the "data-design gap" in traditional systems.

[0257] In summary, this invention not only surpasses existing best-in-class technical solutions in terms of perception breadth, evaluation depth, semantic credibility, and structural executability, but also achieves a complete closed loop of data-driven, risk quantification, and structural output for the first time, demonstrating extremely high industry applicability and promotional value. Attached Figure Description

[0258] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0259] As shown in the figure, the present invention includes the following steps:

[0260] Step 1: Construct the spatial coordinates of the care area and adapt the spatial coordinate grid to care behaviors;

[0261] Step 2: Deploy multi-mode sensing nodes in the high-adaptability coordinate grid area of ​​Step 1 to collect environmental disturbance factors in real time using multi-mode methods, and build a sensing model and a heterogeneous sensing co-calibration system (HSCM).

[0262] Step 3: Generate a "cognitive load distribution spectrum" for the spatial coordinate grid;

[0263] Step 4: Integrate spatial coordinates, perception model and heterogeneous perception co-calibration system, and "cognitive load distribution spectrum" into "care structure sub-element".

[0264] In step 1, the care area is a two-dimensional closed space S. The space S is divided into M rows and N columns of grid cells with a fixed resolution, denoted as G(i,j), where i∈[1,M] and j∈[1,N]. Each grid cell represents a minimum evaluation unit, used to carry out subsequent behavioral adaptation and environmental assessment calculations.

[0265] The set of dementia care behaviors is B = {B_1, B_2, ..., B_m}, and each behavior is pre-defined to occur at a location O_k = (x_k, y_k).

[0266] For each grid G(i,j), calculate its fit score for all care behaviors, defined as follows:

[0267] Formula 1: Caregiver Behavior Fit Function

[0268] Q_ij=∑{k=1 to m}[(P_k(i,j)*W_k) / (1+D_k(i,j)^2)]

[0269] The explanation is as follows:

[0270] Q_ij: represents the overall fitness score of the (i,j)th grid for all behaviors;

[0271] m: Total number of caregiving behaviors;

[0272] P_k(i,j): Potential fitness score of behavior B_k performed in grid (i,j);

[0273] W_k: Importance weight of behavior B_k;

[0274] D_k(i,j): The Euclidean distance between the preset occurrence position O_k of behavior B_k and the grid (i,j), calculated as follows:

[0275] Formula 2: Euclidean distance calculation

[0276] D_k(i,j)=sqrt((i-x_k)^2+(j-y_k)^2)

[0277] in:

[0278] x_k, y_k: The center of the preset position of the k-th line;

[0279] i,j: The coordinate indices of the current grid in the spatial partitioning;

[0280] The structural characteristics of this function are: the higher the behavior priority fit degree and the closer to the ideal position, the larger the fit degree Q_ij value, indicating that the area is more suitable for configuring this type of care task.

[0281] The Q_ij values ​​of all grids are visualized and output; the color intensity represents the behavioral load and adaptation density of the area, which is used to indicate which areas need to be replanned.

[0282] Description of practical application scenarios:

[0283] Before renovating the dementia-specific ward at a nursing home, this method was used to conduct Q_ij mapping analysis on the entire floor. The results showed that the original toilet area was concentrated in the southeast corner, while the Q_ij heatmap indicated that the high-fit areas for the care behavior "toilet assistance" were concentrated near the center of the living activity area. The original night patrol route had several grids with significantly low Q_ij values, suggesting insufficient support for patrol activities, and a redesign of emergency lighting and quiet passageways was recommended. Through the "contextual embedding grid method" fit modeling, the designers not only optimized the behavioral flow but also quantified "behavioral-environment consistency" digitally, significantly improving the professionalism of the care space.

[0284] Each behavior has a preset occurrence location O_k = (x_k, y_k). A spatial coordinate grid and behavior adaptability scoring model are constructed to quantify the adaptability of each spatial unit to different behaviors. An adaptation function is calculated based on environmental factors, behavioral requirements, and accessibility indicators, and the region with the highest score is selected as O_k. At the same time, data collected by the perception system in real-world scenarios are combined with historical behavior trajectories and behavior hotspot distributions to further verify or adjust the preset occurrence location to make it closer to actual needs.

[0285] Each behavior has a pre-defined location O_k = (x_k, y_k), which is the coordinate of the area considered most suitable for its occurrence in the care process planning. In the dementia care area optimization method, setting a pre-defined location O_k = (x_k, y_k) for each care behavior B_k is a systematic standard based on multi-dimensional criteria and methods. First, from the perspective of care practice and ergonomics, different types of behaviors should be spatially reasonable and feasible. For example, toileting should be near the restroom, eating should be in the dining area, and relaxation activities should be in quiet areas. These arrangements are derived from the experience and norms of the elderly care industry and ergonomic design principles. Second, at the technical implementation level, the adaptability of each spatial unit to different behaviors can be quantified by constructing a spatial coordinate grid and a behavioral adaptability scoring model. Based on environmental factors (such as light and noise), behavioral needs (such as privacy and safety), and accessibility indicators, an adaptation function is calculated, and then the area with the highest score is selected as O_k. Furthermore, by combining data collected by the perception system in real-world scenarios, historical behavioral trajectories and behavioral hotspot distributions can be used to further verify or adjust these preset locations, making them more closely aligned with actual needs. Expert assessments and feedback from caregivers can also serve as supplementary inputs, enhancing the scientific rigor and human-centeredness of site selection. Ultimately, through multi-source information fusion and model optimization, a care area layout with a high degree of alignment between behavior and space can be achieved, providing a solid foundation for subsequent cognitive load modeling and structural sub-element division. This approach effectively improves space utilization efficiency and care response accuracy, demonstrating a fusion of systematic and intelligent applications.

[0286] The following is stated:

[0287] B_1: Bathing assistance

[0288] B_2: Toileting care

[0289] B_3: Dining Guidance

[0290] B_4: Night Patrol ...

[0292] B_m: Other customized behaviors (such as group activities).

[0293] The potential adaptability score ranges from 0 to 1.

[0294] In the dementia care area optimization method described in this invention, the "potential adaptability score" refers to the feasibility and suitability of a certain care behavior within a specific spatial grid unit. Its value ranges from 0 to 1, with a higher value indicating that the behavior is more suitable for occurring within that grid. This score is not the actual frequency of occurrence, but rather an indicator of "potential adaptability" assessed based on multiple spatial and behavioral elements.

[0295] The definition basis of potential fitness score

[0296] The potential fitness score P_k(i,j) is typically evaluated based on a combination of the following dimensions:

[0297] 1. Spatial Function Matching Degree: Does the grid have the basic functional facilities required for this behavior, such as the need to be close to the toilet, have a floor drain, and have anti-slip design for bathing behavior?

[0298] 2. Environmental suitability: Are the lighting, noise, temperature, and humidity conditions suitable for the occurrence of this behavior?

[0299] 3. Behavioral accessibility and pathway patency: How easily can the patient reach the grid from the living area to perform the behavior?

[0300] 4. Requirements for privacy or safety: For example, a certain degree of privacy is required when using the toilet or taking a shower, and whether there are partitions or safety protection facilities;

[0301] 5. Historical behavior data support: Does the system already have records of behaviors that frequently occur in this area?

[0302] Example description

[0303] Care Behavior B1: Bathing

[0304] Suppose a care area has a grid G(5,6). When calculating its potential adaptive score P_1(5,6) for bathing behavior, the evaluation is as follows:

[0305] Spatial Function: This area contains a restroom and a drain (score 0.9)

[0306] Environmental conditions: Moderate temperature and humidity, but slightly insufficient ventilation (score 0.7)

[0307] Accessibility: Shortest walking distance from the elderly person's room (score 0.8)

[0308] Privacy: Partitioned and used reasonably (score 0.9)

[0309] After comprehensive weighted scoring: P_1(5,6)=0.84

[0310] Therefore, the potential fitness score of the grid G(5,6) for bathing behavior B1 is 0.84, indicating that it is a relatively ideal spatial candidate unit.

[0311] The potential adaptability score P_k(i,j) is a predictive evaluation of behavior-space relationships. It combines spatial physical characteristics, environmental perception parameters, behavioral demand logic, and empirical data to reflect the "potential rationality of occurrence" of a certain behavior in a grid cell. In this invention, this score can be used as a key parameter of the fitness function Q_ij to participate in spatial planning and optimization, further promoting the scientific output of care structure sub-sets.

[0312] The method for obtaining P_k(i,j) is as follows: AI training is performed through historical care path trajectories, infrared records, and nurse check-in data to generate a behavior heatmap, and the high-frequency behavior areas are mapped to P_k(i,j); and P_k(i,j) is dynamically adjusted through environmental resistance factors (such as noise and obstacles).

[0313] The specific method for generating behavior heatmaps through AI training is as follows:

[0314] a. Input data construction

[0315] Collect the following multi-source historical care data:

[0316] Care path trajectory data: a sequence of time coordinates from devices worn by nurses (such as RFID, Bluetooth tags, and indoor positioning);

[0317] Infrared sensor image sequences: used to sense nighttime activities and behavioral distribution in low-light areas;

[0318] Nurse attendance data includes attendance time, attendance location coordinates, and behavior annotations (such as "turning over", "taking a bath", etc.).

[0319] b. Feature extraction and encoding methods

[0320] All raw behavioral events are transformed into a "spatial grid activity frequency matrix", which is the statistical frequency of behavior B_k corresponding to each grid (i,j);

[0321] Using the time sliding window method, N heat map samples are generated, each of which is a grid distribution map of row B_k within a time period;

[0322] Use CNN (Convolutional Neural Network) to extract spatial pattern features from spatial heatmaps;

[0323] We combined LSTM modules to model the temporal transition patterns of care behaviors.

[0324] c. Model training methods

[0325] Construct a multi-class behavior prediction model, with labels representing the regions where the actual behavior occurred;

[0326] The loss function is either multi-class cross-entropy or spatial clustering similarity index;

[0327] The confidence level of the region where the behavior occurs is used as the output behavior heatmap, and the predicted value on each grid (i,j) is the initial P_k(i,j) value;

[0328] All P_k(i,j) values ​​are normalized to the interval [0,1].

[0329] The dynamic adjustment mechanism of the environmental resistance factor is as follows:

[0330] a. Acquisition of environmental factors

[0331] The following environmental metrics are collected at each grid-deployed or associated sensing node:

[0332] Noise intensity (N_ij)

[0333] Temperature and humidity levels (T_ij)

[0334] Air circulation (A_ij)

[0335] Traffic congestion level (C_ij)

[0336] b. Calculation of resistance factor

[0337] Define a weighted environmental resistance function:

[0338] R_ij=α1*N_ij+α2*|T_ij-T_opt|+α3*(1-A_ij)+α4*C_ij

[0339] T_opt represents the optimal temperature for the behavior (e.g., 26℃~28℃ for bathing).

[0340] Each α coefficient is set based on experience or expert knowledge.

[0341] R_ij∈[0,1]

[0342] c. P_k(i,j) dynamic adjustment method

[0343] The final score is calculated as follows:

[0344] P_k(i,j)_adjusted=P_k(i,j)*(1-R_ij)

[0345] The larger the resistance factor, the less suitable the current environment is for the execution of this behavior, and the corresponding P_k(i,j) will be suppressed;

[0346] The system automatically updates environmental data and refreshes the P_k(i,j) value at regular intervals (e.g., every 10 minutes).

[0347] This method generates a behavior heatmap by training an AI model based on historical multi-modal behavioral data, thus forming the initial value of P_k(i,j). This initial value is then dynamically adjusted in real-time using the environmental resistance factor R_ij, ultimately yielding the dynamic potential adaptability score of each spatial grid for behavior B_k. This approach combines the accuracy of data-driven learning with the real-time nature and personalized adjustment capabilities of sensory feedback, effectively enhancing the responsiveness of care space configuration to real-world behavioral scenarios.

[0348] The W_k setting method is as follows: it is set according to the complexity of care and the risk level (e.g., if the risk of bathing is high, W_k = 0.9; if dining can be guided, W_k = 0.3); the configuration strategy matrix W = [W_1,...,W_m] is dynamically adjusted on the front-end interface.

[0349] The correspondence between the care complexity and risk level and W_k is as follows:

[0350] Care complexity C_k∈[0,1]: represents the skill requirements, time consumption, and human resources required to perform the action;

[0351] Risk level R_k∈[0,1]: indicates the degree of harm that may be caused if the behavior is not performed properly or the environment is unfavorable (such as falls, burns, etc.);

[0352] Formula for calculating weight W_k:

[0353] W_k=α*C_k+β*R_k

[0354] Here, α and β are weighting coefficients, commonly set to α = 0.4 and β = 0.6, i.e., risk-prioritizing. These can be adjusted flexibly according to the system strategy.

[0355] The strategy matrix W = [W_1,...,W_m] is dynamically adjusted in the front-end interface as follows:

[0356] a. Interface Settings

[0357] On the system's front-end management interface (Web or touchscreen), the following fields are displayed for each care behavior B_k:

[0358] Name of the action (e.g., "taking a bath", "using the toilet");

[0359] The current complexity is C_k slider (range 0.0 to 1.0);

[0360] Current risk level R_k slider (range 0.0~1.0);

[0361] The result W_k is automatically calculated and displayed;

[0362] Whenever the user adjusts C_k or R_k, the system automatically updates the displayed value of W_k in real time according to the formula;

[0363] b. Dynamic interaction implementation method (technical implementation)

[0364] Front-end language: Use HTML + JavaScript (such as Vue.js or React) to build slider components;

[0365] Data binding: The slider component is bound to the calculation function in real time, and adjustments trigger an update of W_k.

[0366] Calculation logic:

[0367] function calculateWk(Ck, Rk){

[0368] return(0.4*Ck+0.6*Rk).toFixed(2);

[0369] }

[0370] Synchronize with the backend:

[0371] When the user clicks the "Save" button, the current W array is sent to the server via AJAX;

[0372] The backend updates the strategy matrix to the configuration database (such as SQLite or MongoDB);

[0373] The system will automatically use the updated W array for the next behavior adaptation calculation.

[0374] The weighting of W_k in this invention can be calculated by linear weighting of care complexity C_k and risk level R_k, which is a standardized and scalable scoring mechanism. The front-end dynamic adjustment strategy matrix W is input through a slider interface, and W_k is calculated in real time by binding the calculation function. It also supports synchronization to the back-end system configuration, which can be flexibly adjusted according to individual needs in actual care scenarios.

[0375] In dementia care area optimization methods, the importance weight W_k of care behaviors is a core parameter affecting the accuracy and scientific validity of the spatial fit score Q_ij. Its assignment should be based on clear standards rather than subjective settings. First, the weight can be comprehensively scored based on multiple dimensions such as the risk level of the behavior itself, execution frequency, care dependency, time sensitivity, and cognitive challenge. This dimensional scoring method can form a structured weight evaluation matrix, helping to systematically determine the spatial sensitivity and priority of each type of behavior. Second, if the system has multimodal perception capabilities, it can also utilize the correlation between various care behaviors and abnormal events (such as falls, getting lost, delayed medication, etc.) in historical data. A data-driven weight calculation model can be constructed based on the event frequency F_k and severity S_k, for example, using the formula W_k = α*F_k + β*S_k, where α and β are adjustment coefficients, thereby achieving continuous optimization and feedback adjustment. Furthermore, the Delphi method can be used to invite multiple care experts and rehabilitation personnel to rank the importance of various behaviors, reaching consensus through multiple rounds of anonymous scoring, thus forming an expert knowledge-driven weight baseline. These three methods can be used individually or combined in proportion to construct a more scientific and adjustable hybrid weighted model. Finally, through normalization, it is ensured that all W_k values ​​are within a certain range (e.g., 0 to 1) and can directly participate in the fit calculation. This allows care behaviors to reflect their inherent risk sensitivity and care value in spatial configuration optimization, providing solid data support and decision-making basis for the structural division and dynamic planning of dementia care areas.

[0376] The environmental resistance factor includes noise and obstacles.

[0377] Step 2 involves deploying multi-mode sensing nodes (PCNs) in high Q_ij value regions. Each PCN periodically collects the following five types of data streams within its deployment area: image and video streams, audio sound field intensity and frequency distribution, changes in air composition and odor, dynamic changes in illumination, and changes in temperature and humidity.

[0378] The Q_ij threshold setting adopts the normalized threshold method or the quantile threshold method;

[0379] Normalized threshold method:

[0380] The calculation result of Q_ij is normalized and the value range is [0,1], with a fixed threshold set;

[0381] The region where Q_ij ≥ 0.7 is designated as the "high Q_ij value region";

[0382] The range 0.4 ≤ Q_ij < 0.7 is considered a "medium fit region".

[0383] Q_ij < 0.4 indicates a "low-fitness region" and deployments should be ignored.

[0384] Quantile-Based Thresholding:

[0385] To adapt to the differences in structure and behavior distribution in different care areas, a statistical distribution method is used to dynamically determine the "high Q_ij area";

[0386] Calculate the 80th percentile of all grid Q_ij values, denoted as Q_thr;

[0387] The grids that satisfy Q_ij≥Q_thr are designated as "high Q_ij regions";

[0388] Adjust the system configuration parameters to the percentile (e.g., 75%, 85%, 90%) to control the number of deployments.

[0389] For example:

[0390] If a care area has 100 grid cells, and after sorting, the 80th highest Q_ij value is found to be 0.68, then Q_thr = 0.68, and the 20 grid cells with Q_ij ≥ 0.68 are considered as high-fit regions.

[0391] The normalized threshold method is simple, clear, and easy to implement; it is easy to combine with color heatmaps for intuitive display; and it can stably control the number of sensing nodes and deployment costs.

[0392] The quantile threshold method is more adaptable and data-driven; it can dynamically control the number of deployments and distribution density in large-scale deployments; and it can be combined with the density of care behaviors to adjust the perception deployment strategy.

[0393] The normalized threshold method can be used to define regions with Q_ij ≥ 0.7 as high Q_ij value regions, or the quantile method can be used to identify the top 20% of all grid Q_ij values ​​as high-fit regions, thereby determining the target area for deploying multimodal sensing nodes (PCNs). Specific threshold or quantile parameters can be preset or dynamically adjusted in the system interface to adapt to differences in care space layouts and care behavior density distributions.

[0394] The "high Q_ij value region" is a key spatial selection criterion for deploying multimodal sensing nodes (PCNs). Q_ij essentially represents the comprehensive adaptability score of a grid (i,j) to the protection behavior. The above-mentioned "high Q_ij region" division criteria take into account the interpretability of the system, the efficiency of environmental resource allocation, and the technical feasibility of sensing deployment.

[0395] The PCN node k collects a set of perceptual feature vectors at time t:

[0396] PCN_k(t)={V_k(t),A_k(t),G_k(t),E_k(t),T_k(t)}

[0397] in:

[0398] V_k(t): Image motion perturbation feature vector;

[0399] A_k(t): Audio spectrum feature vector;

[0400] G_k(t): Illumination distribution characteristic value;

[0401] E__k(t): Estimated odor intensity (ppm);

[0402] T_k(t): Environmental values ​​in both temperature and humidity dimensions.

[0403] Each multimodal sensing node (PCN node) can adopt a modular integrated edge intelligence structure, consisting of an edge computing unit and multimodal sensor components, possessing the ability to acquire and preliminarily process multi-source image, audio, and environmental data. Specifically, the edge computing unit can use embedded AI platforms such as Raspberry Pi 4B, Jetson Nano, and Google Coral Dev Board, equipped with processors, GPU / TPU coprocessor units, and rich peripheral interfaces, capable of supporting the operation of local AI models. Image motion perturbation features V_k(t) are acquired through a camera and local motion information is extracted using methods such as background modeling, optical flow, or frame difference; audio spectral features A_k(t) are obtained through I... 2 The data is acquired by a digital MEMS microphone with an S-interface and converted into a spectral vector using algorithms such as Fast Fourier Transform (FFT); the illumination characteristic G_k(t) is measured by a photosensor (such as BH1750, TSL2591); the odor intensity E_k(t) is estimated at the ppm level using a common gas sensor (such as MQ-135, SGP30); and the temperature and humidity environmental value T_k(t) can be synchronously acquired by a high-precision environmental sensor such as SHT31, BME280. All of the above modules can be accessed via I... 2 C, SPI, UART, or GPIO communicate with the edge master controller to achieve periodic multi-mode data acquisition. Although this configuration is not an off-the-shelf commercial integrated device, it is a widely used hardware and software combination in current intelligent sensing systems, possessing high scalability, low cost, and engineering feasibility. It can support the high-frequency dynamic sensing requirements of the local environmental state of the care area in this invention and provide high-quality data support for subsequent cognitive stress analysis and care structure sub-element generation.

[0404] In step 2, the perception model includes image perturbation complexity entropy and composite stimulus loading function;

[0405] Image perturbation complexity entropy:

[0406] S_k(t)=-∑{n=1 to N}[(F_n(t) / ∑F_n(t))*log2(F_n(t) / ∑F_n(t))]

[0407] in:

[0408] N: The number of moving object categories detected in the image (e.g., people, curtains, lamp shadows, etc.);

[0409] F_n(t): The magnitude of the perturbation characteristic change of the nth type of moving target at time t, which can be obtained by continuous frame difference;

[0410] ∑F_n(t): Total disturbance intensity of all target categories at time t;

[0411] S_k(t): Visual perturbation entropy of node k at time t. The larger the value, the more complex the visual stimulus, which is more likely to cause excessive cognitive load.

[0412] Visual perturbation entropy is an indicator that measures the degree of visual perception "chaos" in a certain area at a specific moment. When a large number of objects move simultaneously in an irregular manner in the field of vision, and the lighting changes frequently, the distribution of F_n(t) becomes uneven, and the value of S_k(t) increases, suggesting that the area may become a high-risk point for cognitive impairment.

[0413] Composite stimulus loading function (environmental burden assessment value):

[0414]

[0415] in:

[0416] R_k(t): The comprehensive environmental load score of node k at time t;

[0417] S_k(t): Visual perturbation entropy (see above);

[0418] σ_a(t): Standard deviation of the audio power spectrum at time t (indicating the degree of non-uniformity of the sound field);

[0419] ΔG_k(t): The difference in light intensity over one minute (representing the light variation frequency);

[0420] E_k(t): Odor concentration value (e.g., volatile matter);

[0421] Temperature gradient (e.g., the cold air zone in an air conditioner);

[0422] γ1~γ5: These are weighting factors that can be adjusted using training data to reflect the importance of each sensory channel.

[0423] R_k(t) is a multi-channel stimulus load estimation model used to determine whether there is a risk of overstimulation in the current area. For example, if an area has high-intensity stimuli in terms of vision, hearing, and smell, it is not suitable as the core activity area for elderly people with dementia, even if the structural layout is reasonable.

[0424] Deployment method: Use portable module integration devices (such as Raspberry Pi + Camera + environmental sensor set); each PCN node contains an AI edge module to realize local identification and raw data encoding; power supply can be achieved by wireless charging board or low-voltage power supply wiring; data is transmitted back to the central computing platform via Wi-Fi or LoRa communication.

[0425] Sampling frequency: 1 frame per second (1fps) for image stream; 1 spectral compression packet for sound sampling every 10 seconds; 1 complete feature vector for environmental parameters every 30 seconds.

[0426] Dynamic adjustment mechanism: Combining the Q_ij heatmap changes in step 1 with the real-time load function R_k(t), the system supports: dynamic activation / deactivation of nodes (saving computing power); device self-check and reporting of data collection quality indicators.

[0427] Description of practical application scenarios

[0428] In a pilot project for a cognitive care zone in a senior care facility, the following issues were identified by deploying PCN nodes in six high-behavioral-density areas:

[0429] 1. In a certain dining area, the visual disturbance entropy S_k(t) has been consistently higher than 0.75 due to the frequent movement of the curtains.

[0430] 2. In another area, the air conditioner vents blow directly into the passageway, causing a temperature gradient. Unstable, with R_k(t) consistently exceeding the set threshold;

[0431] After adjustments, the curtains were fixed and the ventilation angle was modified, resulting in a 35% decrease in regional R_k(t). Experts assessed that the frequency of mood fluctuations in elderly people with dementia decreased significantly.

[0432] The weighting factors γ1 to γ5 in the composite stimulus loading function are used to measure the relative influence of different sensory channels (visual, audio, light, odor, and temperature) on cognitive load, and their setting has clear standard basis. Initial weights can be empirically assigned based on research findings in behavioral psychology and cognitive neuroscience. For example, visual disturbances and changes in the sound field generally have a greater impact on cognitive load, and their weights can be set to 0.35 and 0.30 respectively, while the weights for sudden changes in light, odor concentration, and temperature gradient are appropriately lower, such as 0.15, 0.10, and 0.10, with a total weight of 1. In practical applications, these weights can also be trained and fine-tuned based on historical care event data through regression analysis or gradient optimization methods to better suit the environmental characteristics and usage scenarios of specific locations. Furthermore, at the system operation level, the multi-modal sensing nodes support a dynamic activation and deactivation mechanism. They can intelligently schedule node acquisition status based on real-time evaluation results of spatial adaptability scores Q_ij or load functions R_k(t), thereby saving computational and energy resources while ensuring data integrity. Simultaneously, each node has a periodic self-checking function, capable of monitoring quality indicators such as camera frame rate, sensor response status, and audio input integrity, and reporting these to the central system in real-time via a network interface. This achieves an intelligent operation mechanism with controllable acquisition quality and transparent node status. The above technologies are mature methods in edge computing and IoT systems, possessing feasibility and replicability in existing engineering projects.

[0433] In step 2, the heterogeneous perception and common standardization system (HSCM) is constructed, specifically as follows:

[0434] Three-source label vector definition: The state-aware label for region (i,j) at time t is:

[0435] HSCM(i,j,t)={L_h,L_e,L_m}

[0436] in:

[0437] L_h: Labels added by caregivers, such as "difficult to pass through" or "noisy";

[0438] L_e: Caregivers label their patients, such as “discontinuous pathways” or “poor behavior orientation”;

[0439] L_m: AI model identifies tags, such as "chaotic crowd flow" or "abnormal lingering area".

[0440] All the above tags are standardized into vector form, in the following format:

[0441] L_x = {l_1,l_2,...,l_n}

[0442] Where l_k∈{0,1} indicates whether the k-th type of risk or warning label is included.

[0443] The formula for calculating tag similarity is: The semantic similarity sim(a,b) between tag vectors is:

[0444] sim(a,b)=(a·b) / (||a||*||b||)

[0445] in:

[0446] a, b: Two label vectors;

[0447] a·b: the vector dot product of a and b, i.e. the number of common labels between them;

[0448] ||a||: the Euclidean norm of a, i.e., sqrt(∑a_k^2);

[0449] sim(a,b): The value range is [0,1]. The closer it is to 1, the more similar the two are.

[0450] Three-way fusion confidence function: After obtaining the three-source similarity, calculate the fusion confidence score F_conf, which is used to generate unified label output. The formula is as follows:

[0451] Three-source collaborative fusion confidence function

[0452] F_conf=tanh[(λ1*sim(L_h,L_e)+λ2*sim(L_h,L_m)+λ3*sim(L_e,L_m)) / (1+θ*E_dev)]

[0453] in:

[0454] F_conf: Confidence value of the final fused label;

[0455] sim(L_h,L_e), sim(L_h,Lm), sim(L_e,L_m): Three-party label similarity;

[0456] λ1, λ2, λ3: Adjustable weighting coefficients, representing the contributions of different subjects to the fusion result;

[0457] E_dev: The degree of coordination bias of the three-party labels is as follows;

[0458] θ: Penalty intensity factor;

[0459] tanh(): Hyperbolic tangent normalization function, which compresses the output value to (-1,1), and is often used to sense and suppress extreme value fluctuations.

[0460] Formula for calculating the degree of coordination deviation E_dev

[0461] E_dev=(1-sim(L_h,L_e))+(1-sim(L_h,L_m))+(1-sim(L_e,L_m))

[0462] explain:

[0463] E_dev reflects the degree of label divergence among the three parties; a larger value indicates greater divergence.

[0464] In F_conf, it is used as an adjustment factor in the denominator of the penalty term to ensure that high divergence will significantly suppress the fusion confidence value.

[0465] System Implementation Details

[0466] 1. Methods for obtaining third-party annotations:

[0467] L_h: Nursing staff can quickly select tags during daily rounds via a mobile app;

[0468] L_e: Annotated through the expert system backend, combining historical images and environmental sampling revisit;

[0469] L_m: Automatically generated from the inference results of the AI ​​model deployed on the front-end camera.

[0470] 2. Tag body definition:

[0471] The system comes pre-installed with a unified semantic label ontology library LabelSet, which contains 30 common care risk and auxiliary suggestion labels; all L_h, L_e, and L_m are vector-mapped through LabelSet to ensure structural consistency.

[0472] 3. Tag fusion output format:

[0473] The system outputs the following structure for each grid:

[0474] Unified_Label(i,j,t)={

[0475] Label_fused:L_f,

[0476] Confidence:F_conf,

[0477] Agreement_Level:1-E_dev / 3

[0478] }

[0479] in:

[0480] L_f: The fused label vector, based on a majority voting mechanism (if 2 / 3 are the same, it is retained);

[0481] Confidence: Fusion confidence level;

[0482] Agreement_Level: Tripartite Consistency Index. A higher value indicates convergence of judgments from multiple sources.

[0483] Description of practical application scenarios

[0484] In the preliminary survey for the renovation of corridor care in a certain integrated medical and elderly care institution, a HSCM system was deployed in a section of the corridor. The results are as follows:

[0485] Caregiver L_h was labeled as "insufficient lighting, slippery"; expert L_e was labeled as "lack of behavioral guidance"; model L_m was identified as "abnormal passage delays, repeated U-turns";

[0486] Similarity calculated as follows:

[0487] sim(L_h,L_e)=0.5;

[0488] sim(L_h,L_m)=0.6;

[0489] sim(L_e,L_m)=0.4;

[0490] E_dev = 1.0 + 0.4 + 0.6 = 2.0

[0491] F_conf≈tanh[(0.5+1.2+0.8) / (1+1.5*2.0)]=tanh(2.5 / 4.0)≈0.6

[0492] Final fusion output of the system:

[0493] Tags: Traffic congestion, unclear environment

[0494] Confidence level: 0.6

[0495] Consistency level: 1 - 2.0 / 3 ≈ 0.33 (low to medium consistency)

[0496] Recommended renovation: Add sensor lighting, apply anti-slip treatment to the floor, and install dynamic guide lights; the effect will be significant.

[0497] The weight parameters λ1, λ2, and λ3 in the three-source collaborative fusion confidence function are used to measure the contribution of different label sources to the final fusion decision, forming a trade-off mechanism based on multi-dimensional evaluation criteria. Typically, the three label sources include manual annotation (L_h), edge AI node identification results (L_e), and central model or historical statistical output (L_m). In practical applications, evaluation can be based on the accuracy of each subject in historical data (e.g., consistency rate with validation samples), output stability (label fluctuation), data integrity (whether there are missing or anomalies), and importance in the task context. For example, in emergency response, edge node identification has a faster response speed, so λ2 can be increased; while in daily behavior pattern recognition, the central model has global regularity due to long-term accumulation, so λ3 can be appropriately increased. Although manual annotation has an empirical basis, it is usually assigned a moderate weight λ1 due to subjective bias or operational errors, such as an initial suggestion of setting λ1 = 0.3, λ2 = 0.4, and λ3 = 0.3. Furthermore, this system also supports dynamic adjustment of weight parameters through data-driven methods. For example, by calculating the accuracy and consistency of each label source in historical samples in real time, a linear combination model is used to optimize weight allocation, thereby improving the credibility and environmental adaptability of the final fused label. This fusion weight strategy takes into account human experience, edge response capability, and central model accuracy, ensuring the scientific rigor and robustness of label judgment in care scenarios.

[0498] In the three-source collaborative fusion confidence function described in this invention, the collaborative bias E_dev is used in the denominator of the fusion confidence value F_conf, serving as a penalty to dynamically adjust the confidence output. The core idea is that when there is significant divergence among the three labels, the confidence strength of the fusion result should be reduced to avoid erroneous merging or ambiguous label output. To this end, E_dev is introduced into the denominator and combined with the penalty strength factor θ to form the following adjustment term: (1+θ*E_dev). This mechanism achieves confidence contraction suppression. The specific adjustment rules are as follows: When the consistency of the three-party labels is high, i.e., the values ​​of sim(L_h,L_e), sim(L_h,L_m), and sim(L_e,L_m) are all close to 1, then E_dev approaches 0, resulting in a denominator of approximately 1. The output of F_conf is mainly affected by the weighted similarity, and the fusion confidence is high at this time. However, when there are significant differences among the three parties, such as one party deviating from the other two, resulting in one of them having a low similarity or even approaching 0, then E_dev will increase rapidly, significantly amplifying the denominator and thus suppressing the overall F_conf value. The penalty strength factor θ controls the sensitivity of this suppression effect. The larger θ is, the stronger the system's penalty for divergence, and the more obvious the F_conf contraction. Conversely, when θ approaches 0, the penalty effect weakens, and F_conf's tolerance for divergence increases. The default setting θ∈[1,3] is dynamically adjusted according to the different requirements of the system for fusion stability and error tolerance, ensuring that the system automatically reduces the confidence of the fusion output when the reliability and consistency of the collaborative labels are insufficient, thus ensuring the robustness and reliability of the label decision. This adjustment mechanism not only improves the discrimination accuracy of multi-source collaborative fusion, but also enhances the system's robustness to perceived noise, abnormal labels, and edge conflicts.

[0499] In step 3, generating the "cognitive load distribution spectrum" specifically involves:

[0500] The cognitive stress score for each grid cell (i,j) is denoted as:

[0501] Formula 1: Cognitive stress rating function

[0502] CLFS(i,j)=β1*B_dens(i,j)+β2*E_disp(i,j)+β3*P_blur(i,j)+β4*V_discord(i,j)

[0503] in:

[0504] CLFS(i,j): The overall cognitive load value of region (i,j);

[0505] B_dens(i,j): The overlap density of the behavior trajectories in this grid;

[0506] E_disp(i,j): Standard deviation of noise disturbance distribution in this region;

[0507] P_blur(i,j): The inverse ratio of path sharpness;

[0508] V_discord(i,j): Visual stimulus consistency deviation index;

[0509] β1, β2, β3, β4: These are the weight coefficients of the four indicators, which are set by data training.

[0510] This formula integrates multi-source information from both the behavioral and perceptual levels to provide subsequent behavioral perception fusion judgments.

[0511] Formula 2: Calculation of Behavioral Overlap Density

[0512] B_dens(i,j)=∑{k=1 to m}F_k(i,j) / T_window

[0513] in:

[0514] F_k(i,j): The number of times the k-th type of care behavior occurs in the (i,j) grid within the T_window time;

[0515] m: Total number of behavior types (e.g., bathing, patrolling, taking medication, eating, etc.);

[0516] T_window: Observation time window (in seconds or minutes);

[0517] The behavioral overlap density B_dens(i,j) index is used to measure the degree of clustering of different care behaviors at the same spatial point. A high value indicates a high potential for behavioral conflict or confusion.

[0518] Formula 3: Calculation of environmental noise disturbance

[0519] E_disp(i,j)=sqrt(∑{t=1 to N}(A_t(i,j)-μ_A(i,j))^2 / N)

[0520] in:

[0521] A_t(i,j): The sound pressure level of grid (i,j) at time t;

[0522] μ_A(i,j): The average sound pressure level of this grid;

[0523] N: Number of sampling times;

[0524] E_disp(i,j): Noise disturbance fluctuation in this area; describes the volatility of the acoustic environment in this area, calculated using the time-domain sliding standard deviation. A higher value indicates drastic sound changes in the area, which may easily cause panic or confusion among the elderly.

[0525] Formula 4: Path Ambiguity Calculation

[0526]

[0527] in:

[0528] The average value of the gradient changes along all path directions within the grid region (i,j);

[0529] P_blur(i,j): Path ambiguity score. A higher value indicates a more blurred path; it reflects the degree of ambiguity in the path direction of the grid. Areas with unclear directions and poor guidance may prevent elderly people from forming clear movement goals, leading to wandering behavior or staying in one place.

[0530] Formula 5: Calculation of Visual Stimulus Bias

[0531] V_discord(i,j)=σ_color(i,j)+σ_texture(i,j)+|ΔL(i,j)|

[0532] in:

[0533] σ_color(i,j): The standard deviation of the color histogram in this region;

[0534] σ_texture(i,j): Standard deviation of the texture orientation histogram;

[0535] ΔL(i,j): The absolute value of the change in light intensity between two consecutive frames.

[0536] Visual stimulus bias V_discord(i,j): describes the degree of consistency of visual elements (color, lighting, texture) in this region. Elderly people with dementia are often more sensitive to image edges and color differences; high inconsistency will interfere with their spatial recognition. A larger value indicates higher visual stimulus complexity in this region, making recognition more difficult.

[0537] System implementation: The data collection cycle for each (i,j) grid is 60 seconds; all data is collected uniformly through deployed PCN nodes; the CLFS heatmap is updated every minute and can be deployed on a local embedded device to run in real time; the system supports "hot zone warning": when CLFS(i,j)>T_thresh (preset threshold), it is automatically marked as a key area of ​​concern.

[0538] Real-world application scenarios

[0539] In a pilot test at a senior care center before renovation, the system detected that CLFS(i,j) consistently exceeded 0.85 in the library entrance area, mainly due to:

[0540] Frequent gatherings of people result in high B_dens values; simultaneous operation of air conditioners and automatic doors leads to large noise fluctuations, resulting in high E_disp values; complex wall textures and a lack of uniformity in color markings result in high V_discord values.

[0541] Adjustment measures:

[0542] Add directional signage and light strips, simplify wall decorations; adjust air conditioning fan speed and door opening mode; and relocate the "conversation area" away from the main passageway in terms of behavior planning.

[0543] After the rectification, the CLFS in the area dropped to 0.38, and the frequency of elderly people wandering around decreased significantly.

[0544] Step 4 specifically involves:

[0545] For each functional area Z_n, its output structure is defined as follows:

[0546] CSES(Z_n) = {

[0547] F_phys(Z_n),

[0548] F_env(Z_n),

[0549] F_cog(Z_n),

[0550] F_risk(Z_n),

[0551] F_layout(Z_n)

[0552] }

[0553] The meanings of each field are as follows:

[0554] F_phys(Z_n): Physical space structure parameters (geometry, dimensions, structural boundaries, etc.);

[0555] F_env(Z_n): Environmental state parameters (mean and fluctuation of temperature, humidity, light intensity, noise, etc.);

[0556] F_cog(Z_n): Cognitive risk parameter (from CLFS);

[0557] F_risk(Z_n): Historical records and annotations of anomalous events (such as falls, wandering, etc.);

[0558] F_layout(Z_n): The system's recommended layout adjustment suggestions and interpretability score.

[0559] Layout_opt(Z_n)=ψ1*CLFS_score(Z_n)+ψ2*HSCM_disagreement_rate(Z_n)+ψ3*E_dev_zone(Z_n)

[0560] in:

[0561] Layout_opt(Z_n): Layout optimization suggestions for the nth functional area;

[0562] CLFS_score(Z_n): Average cognitive load score for zone n;

[0563] HSCM_disagreement_rate(Z_n): The proportion of disagreement among the three-party labels, reflecting information conflict;

[0564] E_dev_zone(Z_n): Abnormal fluctuation of sensing nodes within this region;

[0565] ψ1, ψ2, ψ3: Empirical weighting adjustment factors, used to balance the contributions of various factors.

[0566] The core recommendation scoring function (Layout_opt) guides space designers to quickly identify key risk areas and optimization directions. The system outputs a "layout optimization suggestion score" for each area Z_n to assess whether it is necessary to restructure the space structure, adjust the functional configuration, or deploy additional auxiliary equipment.

[0567] Methods for interpreting and obtaining each variable

[0568] 1. CLFS_score(Z_n)

[0569] Defined as the average CLFS(i,j) of all grids in region Z_n:

[0570] CLFS_score(Z_n)=∑{i,j∈Z_n}CLFS(i,j) / N_grid(Z_n)

[0571] in:

[0572] CLFS(i,j): Cognitive load score for the grid;

[0573] N_grid(Z_n): The number of grids contained in region Z_n;

[0574] Meaning: A higher score indicates that the overall perceptual complexity of the area is high and the cognitive load is heavy, which may require reconstruction.

[0575] 2. HSCM_disagreement_rate(Z_n)

[0576] Defined as the percentage of tags whose third-party annotation discrepancy exceeds a set threshold δ:

[0577] HSCM_disagreement_rate(Z_n)=N_disagree(Z_n) / N_total(Z_n)

[0578] in:

[0579] N_disagree(Z_n): The number of tag entries where the maximum similarity among the three source tags is less than δ;

[0580] N_total(Z_n): The total number of label entries for all samples collected in this area;

[0581] δ: Set a threshold (e.g., 0.5); values ​​below this threshold are considered significant divergences.

[0582] Meaning: This indicates that there is significant controversy regarding the semantic understanding of this area, making it unsuitable for direct deployment of fixed care procedures.

[0583] 3. E_dev_zone(Z_n)

[0584] Abnormal fluctuations in sensor data are defined as the mean of the standard deviation of the environmental perception values:

[0585] E_dev_zone(Z_n)=(σ_temp+σ_hum+σ_noise+σ_light) / 4

[0586] in:

[0587] σ_temp, σ_hum, σ_noise, and σ_light represent the standard deviations of temperature, humidity, noise, and light intensity in this region, respectively.

[0588] Meaning: Used to determine whether there is systemic instability in a region, such as direct airflow from air conditioners or frequent noise interference.

[0589] System output methods and support

[0590] The system generates a CSES report for each area every hour, which includes: the structure number and spatial coordinates of area Z_n; the latest values ​​of the F_* parameters in each field; textual descriptions of layout adjustment suggestions and suggestion levels (high, medium, low); and can be exported to JSON / XML / CSV format for use by CAD systems, BIM platforms, or construction units.

[0591] Practical application cases

[0592] In the early design phase of a two-story nursing facility in a certain location, the following outputs were obtained through CSES: Z3 is the corridor connection area, CLFS_score = 0.81 (height), HSCM_disagreement_rate = 0.57, E_dev_zone = 0.33; the system evaluation Layout_opt(Z_3) = 0.76, marked as "recommendation for improvement" level;

[0593] Recommendations: Adjust the lighting layout, add partitions to reduce redundancy in the sensing channels, and standardize graphic symbols;

[0594] After construction was completed and reassessed, Layout_opt dropped to 0.22, and the rate of people getting lost decreased by 42%.

[0595] CLFS_score(Z_n)=∑{i,j∈Z_n}CLFS(i,j) / N_grid(Z_n)

[0596] in:

[0597] CLFS(i,j): Cognitive load score for the grid;

[0598] N_grid(Z_n): The number of grids contained in region Z_n;

[0599] HSCM_disagreement_rate(Z_n)=N_disagree(Z_n) / N_total(Z_n)

[0600] in:

[0601] N_disagree(Z_n): The number of tag entries where the maximum similarity among the three source tags is less than δ;

[0602] N_total(Z_n): The total number of label entries for all samples collected in this area;

[0603] δ: Set a threshold (e.g., 0.5); values ​​below this threshold are considered significant divergences.

[0604] E_dev_zone(Z_n)=(σ_temp+σ_hum+σ_noise+σ_light) / 4

[0605] in:

[0606] σ_temp, σ_hum, σ_noise, and σ_light represent the standard deviations of temperature, humidity, noise, and light intensity in the region, respectively.

[0607] The weight coefficients β1, β2, β3, and β4 used to construct the cognitive load scoring function correspond to four indicators: behavioral trajectory overlap density, noise disturbance distribution, path ambiguity, and visual stimulus bias, respectively, and are used to measure their relative influence in the overall cognitive load. These weight coefficients are set through a data-driven training method, providing clear evaluation criteria and feasibility. In data-driven training, training samples containing spatial grid feature indicators and corresponding cognitive load levels are collected. A regression model is constructed, with the load score as the target output. The best-fit weight combination is calculated using the least squares method or gradient optimization method, thereby achieving adaptive optimization of β1 to β4. To prevent overfitting and abnormal single-item weights, a regularization term is introduced, and the weight results are normalized to a sum of 1, meeting the interpretability and comparability requirements of the scoring function. Through the above setting method, the final generated cognitive load distribution spectrum can more realistically reflect the differences in perceived load in different areas of the care space, providing solid data support for subsequent care structure sub-division and intervention strategy formulation.

[0608] The layout optimization suggestion score Layout_opt(Z_n) is used to measure the overall layout optimization priority of each functional area Z_n in terms of cognitive load, information coordination consistency, and environmental disturbance stability. Its score range is set in the interval [0,1]. The higher the value, the more obvious the structural load risk and cognitive incompatibility of the area under the current layout, and therefore the higher the optimization priority. This score is composed of three components: the average cognitive load score CLFS_score(Z_n), the perception co-standard system disagreement rate HSCM_disagreement_rate(Z_n), and the perception node environmental volatility E_dev_zone(Z_n). The three are generated by weighted summation to generate Layout_opt(Z_n). Among them, CLFS(i,j) is the cognitive load score of each grid, with a value range of [0,1]. It is generated by weighted summation of factors such as behavioral density, noise disturbance, path ambiguity, and visual conflict (see step 3). It can be obtained by regression training to fit the load level label. The higher the standard value, the more likely the area is to cause cognitive pressure.

[0609] For the setting of weighting factors ψ1, ψ2, and ψ3, a normalized empirical weighting strategy is adopted, which can be flexibly adjusted according to different scenario requirements. Generally, initial values ​​of ψ1 = 0.5, ψ2 = 0.3, and ψ3 = 0.2 are recommended. The reason is that cognitive load, as a core indicator, has the greatest impact on behavior, followed by coordination disagreement as a potential risk signal of cognitive confusion, while environmental disturbances, as chronic influencing factors, have a smaller weight. The system also supports dynamic adjustment of the weight combination through data backtracking, such as adaptive optimization based on the statistical coefficients of the correlation between each factor and historical abnormal events, so that the scoring results have good generalization and interpretability in various care spaces. The final output Layout_opt(Z_n) value can serve as an important reference indicator for subsequent spatial layout adjustment suggestions, providing basic support for achieving refined, data-driven intelligent optimization layout of dementia care areas.

[0610] CAGM is the core positioning foundation. Without CAGM, it is impossible to deploy PCN at the semantic level, and it is also impossible to accurately aggregate HSCM and CLFS.

[0611] PCN provides data support, and its multimodal output is a prerequisite for HSCM and CLFS operations;

[0612] HSCM performs cross-subject verification of perception results, improving data interpretability and providing semantic consistency assurance for CLFS;

[0613] The CLFS integrates three core dimensions of stress indicators—behavior, environment, and pathway—and is the only comprehensive way to reflect the cognitive burden of older adults.

[0614] CSES enables the implementation of results, transforming the intelligent analysis results of the system into specific design elements and forming a closed-loop feedback mechanism.

[0615] Therefore, each step is mutually supportive and indispensable. The overall process, from semantic anchoring to sensor acquisition, semantic fusion, risk calculation, and result output, forms a closed-loop system.

[0616] Although each step of this invention can be implemented independently, their combined use achieves synergistic enhancement at key levels, as demonstrated in the following aspects:

[0617] 1. The leap from two-dimensional physical structure to multi-dimensional cognitive environment

[0618] Traditional spatial design is based solely on two-dimensional CAD drawings. This method, by embedding semantic and behavioral layers into CAGM and supplementing it with PCN environmental perception and CLFS cognitive reasoning, achieves a leap from two-dimensional physical space to four-dimensional cognitive space (structure, behavior, perception, and semantics).

[0619] 2. Improving semantic credibility from a single data source to a three-source co-labeled system

[0620] HSCM breaks away from the traditional one-way label generation method that relies on expert experience or AI models. By introducing a three-source perception fusion of nurses, experts, and models, it uses difference constraints to achieve high-confidence labels, which greatly improves data interpretability.

[0621] 3. Innovation in indicators for cognitive stress assessment, moving from environmental parameter assessment.

[0622] CLFS goes beyond parameters such as environmental temperature and humidity, and integrates comprehensive calculations of behavioral density, path clarity, visual consistency, and environmental disturbance fluctuations, achieving for the first time a stress estimation model for people with cognitive impairment.

[0623] 4. A practical closed loop from data analysis to implementable design.

[0624] CSES packages all indicators into a structured set of sub-elements, making it easy for CAD / BIM systems to import them. It realizes a closed-loop path from data collection to analysis, output, design, and execution, completely eliminating the traditional interruptions.

[0625] This invention integrates five subsystems to construct a complete architecture, exhibiting strong complementarity. The combined effect far surpasses the linear summation of the individual components. This constitutes a systemic synergy innovation, embodying the continuous closed-loop value of the method combination in its "structure-perception-semantics-cognition-design" process.

[0626] Outstanding substantive features and remarkable progress

[0627] 1. Systematic breakthrough at the methodological level

[0628] Unlike existing methods based on manual surveys or single-point monitoring, this invention is the first to construct a five-dimensional integrated intelligent care design model that integrates semantic modeling, multi-modal perception, semantic collaboration, risk load, and structural output. The technical path is clear and the logical closed loop is well-defined.

[0629] 2. A New Paradigm for Data-Driven Cognitive Care Design

[0630] This method shifts from subjective experience to data-driven environmental cognitive analysis and transformation design, has strong scalability, and solves the industry pain point of "no data to rely on" in the current design of care areas.

[0631] 3. Strong scalability and platform accessibility

[0632] The CSES output format is compatible with multiple building information systems and has good engineering application value; at the same time, the five modules can be flexibly configured to adapt to the differences in different elderly care institution scenarios.

[0633] This invention is not a simple combination of traditional method modules, but a systematic integrated innovation encompassing spatial understanding, intelligent data acquisition, semantic cognition, risk assessment, and structural optimization. Its technical effects surpass the sum of existing methods, possessing outstanding substantive features and significant progress, fully meeting the innovation and inventiveness requirements for invention patent authorization.

[0634] The principle and workflow of this invention start with the semantic segmentation of care spaces, gradually introducing multimodal sensor deployment, three-party semantic fusion annotation, cognitive stress modeling, and structural output optimization, ultimately achieving quantifiable care zone design. The process includes the following steps:

[0635] 1. Input raw spatial data: CAD drawings or on-site survey maps;

[0636] 2. Execute Context-Based Mocking (CAGM): Construct grid anchors with caregiving semantics;

[0637] 3. Deploy multi-domain sensing couplers (PCNs): Arrange multi-mode sensing nodes based on highly adaptable regions;

[0638] 4. Acquire multi-modal sensing data streams: obtain dynamic sensing information such as images, sounds, smells, and lighting;

[0639] 5. Heterogeneous Sensory Co-labeling System (HSCM) Processing: Three-source semantic annotation and fusion are performed by nursing staff, experts, and the model;

[0640] 6. Generate Cognitive Load Distribution Spectrum (CLFS): Integrate features such as behavioral density, path ambiguity, visual bias, and environmental disturbances for scoring;

[0641] 7. Output Care Structure Subset (CSES): Standardizes and outputs the region structure, perceived state, and design recommendations;

[0642] 8. Support the design decisions of the care area: Provide a basis for subsequent layout optimization, wayfinding design, and equipment placement.

[0643] The present invention can be summarized as including the following steps:

[0644] Step 1: Constructing the "Context-Embedded Positional Method"

[0645] This step aims to provide a structure-semantic-behavioral coupled spatial mapping mechanism for the design of dementia care zones. Traditional architectural space divisions are mostly based on geometric dimensions or functional naming, which is difficult to directly support the matching design of behavioral paths and environmental needs for dementia care tasks. To achieve care task-driven spatial perception and layout guidance, this invention proposes the "Contextual Embedded Grid Method (CAGM)," which integrates structural areas, behavioral intentions, path accessibility, and environmental variables by constructing a multi-dimensional anchor grid. This method not only provides a logical reference for the subsequent deployment of information collection equipment but also directly provides quantitative input for spatial optimization and transformation.

[0646] Step 2: Deploy the "Multi-Domain Sensing Coupler"

[0647] The core objective of this step is to construct a multimodal sensor coupler (PCN) within the identified key spatial regions to achieve real-time correlation modeling between behavioral trajectories and environmental states. Due to the generally declining sensory and spatial awareness abilities of elderly individuals with dementia, they are more sensitive to visual, auditory, olfactory, and light stimuli in their environment. Therefore, the design of the care area must not only consider structural-functional fit but also grasp the actual distribution and time-varying characteristics of multidimensional sensory perturbation factors in the space. To this end, this method deploys aggregated multimodal sensor nodes (PCNs) in regions with high Q_ij values ​​in the previous CAGM study, forming a dynamic "sensory snapshot." Each PCN node is responsible for periodically collecting the following five types of data streams within its deployment area:

[0648] 1. Image and video stream (FOV motion sensing) 2. Audio sound field intensity and frequency distribution 3. Changes in air composition and odor 4. Dynamic changes in illumination 5. Changes in temperature and humidity environment.

[0649] Step 3: Construct a "Heterogeneous Detection and Standardization System"

[0650] In the design of dementia care areas, risk assessments or spatial labels from a single source often lack objectivity and stability due to subjectivity or model bias. For example, caregiver labeling focuses on practical experience; expert labeling leans more towards theoretical standards; and AI models may fail to identify accurately due to biased training data. To improve the uniformity, credibility, and behavioral adaptability of labels, this invention proposes a "Heterogeneous Perception Co-labeling System (HSCM)". By constructing a multi-subject label fusion mechanism, it unifies the semantic labels of caregivers (Human), experts (Expert), and models (Model) into a collaborative matrix, and generates a fused confidence label F_conf based on semantic similarity and a divergence penalty function.

[0651] Step 4: Generate the "Cognitive Load Distribution Spectrum"

[0652] The behavior of elderly people with dementia is greatly influenced by environmental stimuli. If a certain area contains multiple high-intensity, rapidly changing environmental factors (such as sudden sound changes, flickering light and shadow, and unclear paths), it can easily trigger panic, wandering, disorientation, or even falls in the elderly. Therefore, before constructing a care area, high cognitive stress areas should be identified in advance for proactive design adjustments. This step introduces a "cognitive stress distribution spectrum" mechanism to generate a quantitative score for each spatial grid (i,j), reflecting the potential risk of inducing abnormal cognitive load in that area. This score integrates multi-dimensional sensory features in a data-driven manner, ultimately outputting a spatial distribution heatmap to assist in structural modifications, guidance system setup, and dynamic monitoring deployment.

[0653] Step 5: Output the "Care Structure Subset"

[0654] After the first four steps of semantic structure partitioning of the spatial grid, multimodal perception deployment, multi-source semantic annotation, and cognitive load assessment, the system has acquired a large amount of structured and unstructured data. To support subsequent applications such as spatial layout optimization, care path planning, and intelligent perception system deployment, this step proposes a unified "Care Structure Subset (CSES)" data output mechanism. The essence of CSES is to decompose the care space into multiple functional regions Z_n and generate a set of structural, behavioral, perceptual, and risk parameters for each region, forming multidimensional parameter sub-elements usable for design and execution.

Claims

1. A method for optimizing a dementia care area, characterized in that... Includes the following steps: Step 1: Construct the spatial coordinates of the care area and adapt the spatial coordinate grid to care behaviors; Step 2: Deploy multi-mode sensing nodes in the highly adaptable coordinate grid area of ​​Step 1 to collect environmental disturbance factors in real time using multi-mode methods, and build a sensing model and a heterogeneous sensing co-calibration system; Step 3: Generate a "cognitive load distribution spectrum" for the spatial coordinate grid; Step 4: Integrate spatial coordinates, perception model and heterogeneous perception common standardization system, and "cognitive load distribution spectrum" into "care structure sub-element".

2. The method for optimizing a dementia care area according to claim 1, characterized in that... In step 1, the care area is a two-dimensional closed space S. The space S is divided into M rows and N columns of grid cells with a fixed resolution, denoted as G(i,j), i∈[1,M], j∈[1,N]. Each grid represents a minimum evaluation unit, which is used to carry out subsequent behavioral adaptation and environmental assessment calculations. The set of dementia care behaviors is B = {B_1, B_2, ..., B_m}, and each behavior is pre-defined to occur at a location O_k = (x_k, y_k). For each grid G(i,j), calculate its fit score for all care behaviors, defined as follows: Formula 1: Caregiver Behavior Fit Function Q_ij=∑{k=1to m}[(P_k(i,j)*W_k) / (1+D_k(i,j)^2)] The explanation is as follows: Q_ij: represents the overall fitness score of the (i,j)th grid for all behaviors; m: Total number of caregiving behaviors; P_k(i,j): Potential fitness score of behavior B_k performed in grid (i,j); W_k: Importance weight of behavior Bk; D_k(i,j): The Euclidean distance between the preset occurrence position O_k of behavior B_k and the grid (i,j), calculated as follows: Formula 2: Euclidean distance calculation D_k(i,j)=sqrt((i-x_k)^2+(j-y_k)^2) in: x_k, y_k: The center of the preset position of the k-th line; i,j: The coordinate indices of the current grid in the spatial partitioning; The Q_ij values ​​of all grids are visualized and output; the color intensity represents the behavioral load and adaptation density of the area, which is used to indicate which areas need to be replanned.

3. The method for optimizing a dementia care area according to claim 2, characterized in that... Each behavior has a preset occurrence location O_k = (x_k, y_k). A spatial coordinate grid and behavior adaptability scoring model are constructed to quantify the adaptability of each spatial unit to different behaviors. An adaptation function is calculated based on environmental factors, behavioral requirements, and accessibility indicators, and the region with the highest score is selected as O_k. At the same time, data collected by the perception system in real-world scenarios are combined with historical behavior trajectories and behavior hotspot distributions to further verify or adjust the preset occurrence location to make it closer to actual needs. The method for obtaining P_k(i,j) is as follows: AI training is performed using historical care path trajectories, infrared records, and nurse check-in data to generate a behavior heatmap, and the high-frequency behavior areas are mapped to P_k(i,j); and P_k(i,j) is dynamically adjusted using environmental resistance factors.

4. The method for optimizing a dementia care area according to claim 3, characterized in that... The specific method for generating behavior heatmaps through AI training is as follows: a. Input data construction Collect the following multi-source historical care data: Care pathway trajectory data: a time-series coordinate sequence from devices worn by nurses; Infrared sensor image sequences: used to sense nighttime activities and behavioral distribution in low-light areas; Nurse attendance data includes attendance time, attendance location coordinates, and behavior annotations; b. Feature extraction and encoding methods All raw behavioral events are transformed into a "spatial grid activity frequency matrix", which is the statistical frequency of behavior B_k corresponding to each grid (i,j); Using the time sliding window method, N heat map samples are generated, each of which is a grid distribution map of row B_k within a time period; Use CNN to extract spatial pattern features from spatial heatmaps; The LSTM module was used to model the temporal transition patterns of caregiving behaviors. c. Model training methods Construct a multi-class behavior prediction model, with labels representing the regions where the actual behavior occurred; The loss function is either multi-class cross-entropy or spatial clustering similarity index; The confidence level of the region where the behavior occurs is used as the output behavior heatmap, and the predicted value on each grid (i,j) is the initial P_k(i,j) value; All P_k(i,j) values ​​are normalized to the interval [0,1]. The dynamic adjustment mechanism of the environmental resistance factor is as follows: a. Acquisition of environmental factors The following environmental metrics are collected at each grid-deployed or associated sensing node: Noise intensity (N_ij) Temperature and humidity levels (T_ij) Air circulation (A_ij) Traffic congestion level (C_ij) b. Calculation of resistance factor Define a weighted environmental resistance function: R_ij=α1*N_ij+α2*|T_ij-T_opt|+α3*(1-A_ij)+α4*C_ij T_opt is the optimal temperature for the behavior. Each α coefficient is set based on experience or expert knowledge. R_ij∈[0,1] c. P_k(i,j) dynamic adjustment method The final score is calculated as follows: P_k(i,j)_adjusted=P_k(i,j)*(1-R_ij) The larger the resistance factor, the less suitable the current environment is for the execution of this behavior, and the corresponding P_k(i,j) will be suppressed; The system automatically updates environmental data and refreshes the P_k(i,j) value at regular intervals.

5. The method for optimizing a dementia care area according to claim 2, characterized in that... The W_k is set according to the complexity and risk level of care; the strategy matrix W = [W_1,...,W_m] is dynamically adjusted on the front-end interface. The correspondence between the care complexity and risk level and W_k is as follows: Care complexity C_k∈[0,1]: represents the skill requirements, time consumption, and human resources required to perform the action; Risk level R_k∈[0,1]: indicates the degree of harm that may be caused if the behavior is not performed properly or if the environment is unfavorable; Formula for calculating weight W_k: W_k=α*C_k+β*R_k Here, α and β are weighting coefficients, commonly set to α = 0.4 and β = 0.6, i.e., risk priority; these can be flexibly adjusted according to the system strategy. The strategy matrix W = [W_1,...,W_m] is dynamically adjusted in the front-end interface as follows: a. Interface Settings On the system's front-end management interface, the following fields are displayed for each care behavior B_k: Behavior name; The current complexity is C_k slider; Current risk level R_k slider; The result W_k is automatically calculated and displayed; Whenever the user adjusts C_k or R_k, the system automatically updates the displayed value of W_k in real time according to the formula; b. Dynamic interaction implementation method Front-end language: Using HTML + JavaScript to build the slider component; Data binding: The slider component is bound to the calculation function in real time, and adjustments trigger an update of W_k. Calculation logic: function calculateWk(Ck, Rk){ return(0.4*Ck+0.6*Rk).toFixed(2); } Synchronize with the backend: When the user clicks the "Save" button, the current W array is sent to the server via AJAX; The backend updates the strategy matrix to the configuration database; The system will automatically use the updated W array for the next behavior adaptation calculation.

6. The method for optimizing a dementia care area according to claim 1, characterized in that... Step 2 involves deploying multi-mode sensing nodes (PCNs) in high Q_ij value regions. Each PCN periodically collects the following five types of data streams within its deployment area: image and video streams, audio sound field intensity and frequency distribution, changes in air composition and odor, dynamic changes in illumination, and changes in temperature and humidity. The Q_ij threshold setting adopts the normalized threshold method or the quantile threshold method; Normalized threshold method: The calculation result of Q_ij is normalized and the value range is [0,1], with a fixed threshold set; The region where Q_ij ≥ 0.7 is designated as the "high Q_ij value region"; The range 0.4 ≤ Q_ij < 0.7 is considered a "medium fit region". Q_ij < 0.4 indicates a "low-fitness region", and deployment will be ignored. Quantile threshold method: To adapt to the differences in structure and behavior distribution in different care areas, a statistical distribution method is used to dynamically determine the "high Q_ij area"; Calculate the 80th percentile of all grid Q_ij values, denoted as Q_thr; The grids that satisfy Q_ij≥Q_thr are designated as "high Q_ij regions"; Adjust the quantiles of system configuration parameters to control the number of deployments; The PCN node k collects a set of perceptual feature vectors at time t: PCN_k(t)={V_k(t),A_k(t),G_k(t),E_k(t),T_k(t)} in: V_k(t): Image motion perturbation feature vector; A_k(t): Audio spectrum feature vector; G_k(t): Illumination distribution characteristic value; E_k(t): Estimate of odor intensity; T_k(t): Environmental values ​​in both temperature and humidity dimensions.

7. The method for optimizing a dementia care area according to claim 1, characterized in that... In step 2, the perception model includes image perturbation complexity entropy and composite stimulus loading function; Image perturbation complexity entropy: S_k(t)=-∑{n=1to_N}[(F_n(t) / ∑F_n(t))*log2(F_n(t) / ∑F_n(t))] in: N: The number of moving object categories detected in the image; F_n(t): The magnitude of the perturbation characteristic change of the nth type of moving target at time t, which can be obtained by continuous frame difference; ∑F_n(t): Total disturbance intensity of all target categories at time t; S_k(t): Visual perturbation entropy of node k at time t. The larger the value, the more complex the visual stimulus, which is more likely to cause excessive cognitive load. Composite stimulus loading function: in: R_k(t): The comprehensive environmental load score of node k at time t; S_k(t): Visual perturbation entropy; σ_a(t): Standard deviation of the audio power spectrum at time t; ΔG_k(t): The difference in light intensity over one minute; E_k(t): Odor concentration value; Temperature gradient; γ1~γ5: These are weighting factors that can be adjusted using training data to reflect the importance of each sensory channel.

8. The method for optimizing a dementia care area according to claim 1, characterized in that... In step 2, the heterogeneous perception and common standardization system is constructed as follows: The state-aware label for region (i,j) at time t is: HSCM(i,j,t)={L_h,L_e,L_m} in: L_h: Labels marked by caregivers; L_e: Caregiver label; L_m: Labels identified by the AI ​​model; All the above tags are standardized into vector form, in the following format: L_x = {l_1,l_2,...,l_n} Where l_k∈{0,1} indicates whether the k-th type of risk or warning label is included; The semantic similarity sim(a,b) between label vectors is: sim(a,b)=(a·b) / (||a||*||b||) in: a, b: Two label vectors; a·b: the vector dot product of a and b, i.e. the number of common labels between them; ||a||: the Euclidean norm of a, i.e., sqrt(∑a_k^2); sim(a,b): The value range is [0,1], and the closer it is to 1, the more similar the two are; After obtaining the three-source similarity, the fusion confidence score F_conf is calculated to generate a unified label output, as shown in the following formula: Three-source collaborative fusion confidence function F_conf=tanh[(λ1*sim(L_h,L_e)+λ2*sim(L_h,L_m)+λ3*sim(L_e,L_m)) / (1+θ*E_dev)] in: F_conf: Confidence value of the final fused label; sim(L_h,L_e), sim(L_h,L_m), sim(L_e,L_m): Three-party label similarity; λ1, λ2, λ3: Adjustable weighting coefficients, representing the contributions of different subjects to the fusion result; E_dev: The degree of coordination bias of the three-party labels is as follows; θ: Penalty intensity factor; tanh(): Hyperbolic tangent normalization function, which compresses the output value to (-1,1), and is often used to sense and suppress extreme value fluctuations; Formula for calculating the degree of coordination deviation E_dev E_dev=(1-sim(L_h,L_e))+(1-sim(L_h,L_m))+(1-sim(L_e,L_m)) explain: E_dev reflects the degree of label divergence among the three parties; a larger value indicates greater divergence. In F_conf, it is used as an adjustment factor in the denominator of the penalty term to ensure that high divergence will significantly suppress the fusion confidence value.

9. The method for optimizing a dementia care area according to claim 1, characterized in that... In step 3, generating the "cognitive load distribution spectrum" specifically involves: The cognitive stress score for each grid cell (i,j) is denoted as: Formula 1: Cognitive stress rating function CLFS(i,j)=β1*B_dens(i,j)+β2*E_disp(i,j)+β3*P_blur(i,j)+β4*V_discord(i,j) in: CLFS(i,j): The overall cognitive load value of region (i,j); B_dens(i,j): The overlap density of the behavior trajectories in this grid; E_disp(i,j): Standard deviation of noise disturbance distribution in this region; P_blur(i,j): The inverse ratio of path sharpness; V_discord(i,j): Visual stimulus consistency deviation index; β1, β2, β3, β4: These are the weight coefficients of the four indicators, which are set by data training. Formula 2: Calculation of Behavioral Overlap Density B_dens(i,j)=∑{k=1to_m}F_k(i,j) / T_window in: F_k(i,j): The number of times the k-th type of care behavior occurs in the (i,j) grid within the T_window time; m: Total number of behavior types; T_window: The observation time window; Formula 3: Calculation of environmental noise disturbance E_disp(i,j)=sqrt(∑{t=1to_N}(A_t(i,j)-μ_A(i,j))^2 / N) in: A_t(i,j): The sound pressure level of grid (i,j) at time t; μ_A(i,j): The average sound pressure level of this grid; N: Number of sampling times; E_disp(i,j): Noise disturbance fluctuation in this region; Formula 4: Path Ambiguity Calculation in: The average value of the gradient changes along all path directions within the grid region (i,j); P_blur(i,j): Path blur score, the higher the value, the more blurry the path; Formula 5: Calculation of Visual Stimulus Bias V_discord(i,j)=σ_color(i,j)+σ_texture(i,j)+|ΔL(i,j)| in: σ_color(i,j): The standard deviation of the color histogram in this region; σ_texture(i,j): Standard deviation of the texture orientation histogram; ΔL(i,j): The absolute value of the change in light intensity between two consecutive frames.

10. The method for optimizing a dementia care area according to claim 1, characterized in that... Step 4 specifically involves: For each functional area Z_n, its output structure is defined as follows: CSES(Z_n)={ F_phys(Z_n), F_env(Z_n), F_cog(Z_n), F_risk(Z_n), F_layout(Z_n) } The meanings of each field are as follows: F_phys(Z_n): Physical space structure parameters; F_env(Z_n): Environment state parameter; F_cog(Z_n): Cognitive risk parameter; F_risk(Z_n): Historical anomaly event records and annotations; F_layout(Z_n): System-recommended layout adjustment suggestions and interpretability score; Layout_opt(Z_n)=ψ1*CLFS_score(Z_n)+ψ2*HSCM_disagreement_rate(Z_n)+ψ3*E_dev_zone(Z_n) in: Layout_opt(Z_n): Layout optimization suggestions for the nth functional area; CLFS_score(Z_n): Average cognitive load score for zone n; HSCM_disagreement_rate(Z_n): The proportion of disagreement among the three-party labels, reflecting information conflict; E_dev_zone(Z_n): Abnormal fluctuation of sensing nodes within this region; ψ1, ψ2, ψ3: Empirical weighting adjustment factors, used to balance the contributions of various factors; CLFS_score(Z_n)=∑{i,j∈Z_n}CLFS(i,j) / N_grid(Z_n) in: CLFS(i,j): Cognitive load score for the grid; N_grid(Z_n): The number of grids contained in region Z_n; HSCM_disagreement_rate(Z_n)=N_disagree(Z_n) / N_total(Z_n) in: N_disagree(Z_n): The number of tag entries where the maximum similarity among the three source tags is less than δ; N_total(Z_n): The total number of label entries for all samples collected in this area; δ: Set a threshold (e.g., 0.5); values ​​below this threshold are considered significant divergences. E_dev_zone(Z_n)=(σ_temp+σ_hum+σ_noise+σ_light) / 4 in: σ_temp, σ_hum, σ_noise, and σ_light represent the standard deviations of temperature, humidity, noise, and light intensity in the region, respectively.

Images