Method for predicting indoor decoration formaldehyde concentration based on fusion neural network

Abstract

The invention provides an indoor decoration formaldehyde concentration prediction method based on a fusion neural network, and the method comprises the steps: carrying out the quantification and construction of a multi-dimensional static feature and a one-dimensional time sequence feature, and obtaining a static feature matrix and a time sequence feature sequence; extracting features of the static feature matrix through a first convolutional layer and a first pooling layer to obtain key coupling features; performing feature extraction on the key coupling features through a second convolutional layer and a second pooling layer to obtain high-order coupling features; flattening the high-order coupling feature to obtain a coupling feature vector; carrying out feature fusion on the coupling feature vector and the time sequence feature sequence to obtain a fused feature; extracting and processing short-term time sequence dependence and long-term time sequence dependence of the fused features through a long short-term memory network and an attention mechanism to obtain a concentration prediction value; according to the maximum value of the formaldehyde concentration and the minimum value of the formaldehyde concentration, converting the normalized predicted value to obtain an actual value of the formaldehyde concentration; according to the invention, high-precision prediction of indoor formaldehyde concentration in specific days after decoration is realized.

Description

Technical Field

[0001] This invention relates to the intersection of indoor air quality prediction and deep learning technology, and in particular to a method for predicting formaldehyde concentration in indoor decoration based on a fused neural network. Background Technology

[0002] Formaldehyde is the most significant volatile organic pollutant emitted from indoor decoration materials. It has a long-term release characteristic, with a release period of 3-15 years, posing a serious threat to human health. The national standard GB / T 18883-2022 clearly requires that the average indoor formaldehyde concentration over 1 hour be ≤0.1 mg / m³. 3 However, the formaldehyde concentration in newly renovated houses is affected by multiple factors such as material characteristics, usage, temperature and humidity, and ventilation conditions. At the same time, there are complex coupling effects such as "material superposition release" and "temperature and humidity synergistic effect". For example, the formaldehyde release rate under high temperature and high humidity conditions can be 3-5 times higher than under standard conditions.

[0003] Traditional physical models, such as mass balance models, require simplification by assuming that each factor acts independently, and cannot capture the nonlinear coupling relationship between temperature and humidity, material release, and ventilation. Single neural network models, such as pure LSTM, can fit nonlinear relationships, but have limited ability to extract interactive features between multiple factors, resulting in prediction errors exceeding 30% in coupled scenarios.

[0004] The temporal dynamics of formaldehyde release, characterized by "rapid decay followed by stable release," and its spatial correlation with multiple factors, such as the influence of material distribution and room volume on concentration, require collaborative modeling. However, existing models often handle these two types of features separately, making it difficult to effectively integrate spatiotemporal features.

[0005] Existing models often directly input raw features, such as temperature and humidity, without constructing specific features for key influencing factors such as "material superposition effect" and "release-dilution coupling," resulting in models that cannot reflect the complex mechanisms of action in actual decoration scenarios. Summary of the Invention

[0006] This invention provides a method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network, which solves the defects of insufficient coupling effect processing capability and separation of temporal and spatial features in the existing technology. It realizes the fusion prediction technology of extracting multi-factor coupling features through a dedicated network structure and capturing and releasing dynamics by combining a temporal model, and is adapted to formaldehyde concentration prediction in different decoration scenarios.

[0007] This invention provides a method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network, comprising the following steps.

[0008] The collected multidimensional static features and one-dimensional temporal features are quantized and constructed to obtain a static feature matrix and a temporal feature sequence, respectively. The static feature matrix is ​​extracted using the first convolutional layer and the first pooling layer of the convolutional neural network to obtain key coupling features; the key coupling features are then extracted using the second convolutional layer and the second pooling layer of the convolutional neural network to obtain higher-order coupling features; and the higher-order coupling features are then flattened using a fully connected layer to obtain the coupling feature vector. Feature fusion is performed on the coupled feature vector and the temporal feature sequence to obtain the fused features; short-term and long-term temporal dependencies of the fused features are extracted and processed through a long short-term memory network and an attention mechanism to obtain the concentration prediction value; Based on the maximum and minimum formaldehyde concentrations, the normalized predicted values ​​are converted to obtain the actual formaldehyde concentration values.

[0009] According to the method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network provided by the present invention, before quantifying and constructing the collected multidimensional static features and one-dimensional time-series features respectively, the method further includes: Outliers in multidimensional static features are identified using box plots, and the median of the same material type is used to replace outliers; the multidimensional static features are then normalized. Linear interpolation is used to fill in short-term missing data in one-dimensional time series features.

[0010] According to the present invention, a method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network quantifies and constructs the collected multidimensional static features and one-dimensional temporal features respectively, to obtain a static feature matrix and a temporal feature sequence, including: The collected multidimensional static features are reconstructed by mapping multidimensional static features to a two-dimensional matrix to obtain a static feature matrix; Multiple time series samples are generated by using a sliding window to extract one-dimensional time series features through a preset window; the time series samples are then used to construct a time series feature sequence.

[0011] According to the present invention, a method for predicting formaldehyde concentration in indoor decoration based on a fused neural network is provided. Feature extraction is performed on a static feature matrix through the first convolutional layer and the first pooling layer of the convolutional neural network to obtain key coupling features; feature extraction is then performed on the key coupling features through the second convolutional layer and the second pooling layer of the convolutional neural network to obtain higher-order coupling features, including: By traversing the static feature matrix through the first convolutional layer of the convolutional neural network according to the first kernel size and the first step size, a local coupled feature map is obtained. The key coupling features are obtained by max pooling the local coupling feature map through the first pooling layer of the convolutional neural network. By using the second convolutional layer of the convolutional neural network, the key coupling features are traversed according to the second kernel size and the second stride to obtain a high-order coupling feature map; The higher-order coupled feature map is obtained by average pooling through the second pooling layer of the convolutional neural network.

[0012] According to the present invention, a method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network extracts and processes short-term and long-term temporal dependencies of the fused features through a long short-term memory network and an attention mechanism to obtain a concentration prediction value, including: The fused features are extracted by using a long short-term memory network to extract short-term and long-term temporal dependencies, thus obtaining short-term and long-term features. By using the dropout layer, short-term and long-term features are processed according to the first dropout rate to prevent overfitting, resulting in a fused feature sequence. The weights of each time step in the fused feature sequence are calculated using an attention mechanism, and a weighted composite feature vector is obtained based on the weights of each time step. The concentration prediction value is obtained by processing the weighted integrated feature vector through the linear activation function of the output layer.

[0013] According to the present invention, a method for predicting formaldehyde concentration in indoor decoration based on a fused neural network is provided. This method extracts short-term and long-term temporal dependencies from the fused features using a long short-term memory network, resulting in short-term and long-term features, including: The first temporal feature extraction layer of the Long Short-Term Memory Network extracts short-term temporal dependencies from the fused features according to the first number of neurons, thus obtaining short-term features. The long-term features are obtained by extracting long-term temporal dependencies from the fused features through the second temporal feature extraction layer of the long short-term memory network according to the second number of neurons.

[0014] The present invention also provides a device for predicting formaldehyde concentration in indoor decoration based on a fusion neural network, comprising the following modules: The feature input module is used to quantize and construct the collected multidimensional static features and one-dimensional temporal features respectively, to obtain the static feature matrix and the temporal feature sequence; The CNN coupling feature extraction module is used to extract features from the static feature matrix through the first convolutional layer and the first pooling layer of the convolutional neural network to obtain key coupling features; to extract features from the key coupling features through the second convolutional layer and the second pooling layer of the convolutional neural network to obtain higher-order coupling features; and to flatten the higher-order coupling features through a fully connected layer to obtain a coupling feature vector. The LSTM temporal prediction module is used to fuse coupled feature vectors and temporal feature sequences to obtain fused features. The fused features are then processed by a long short-term memory network and an attention mechanism to extract and process short-term and long-term temporal dependencies, resulting in concentration prediction values. The inverse normalization module is used to convert the normalized predicted value based on the maximum and minimum formaldehyde concentration values ​​to obtain the actual formaldehyde concentration value.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the indoor formaldehyde concentration prediction method based on a fused neural network as described above.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the indoor formaldehyde concentration prediction method based on a fused neural network as described above.

[0017] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the indoor formaldehyde concentration prediction method based on a fused neural network as described above.

[0018] The present invention provides a method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network. This method quantifies and constructs collected multidimensional static features and one-dimensional temporal features to obtain a static feature matrix and a temporal feature sequence. Features are extracted from the static feature matrix using the first convolutional layer and the first pooling layer of a convolutional neural network to obtain key coupling features. Further feature extraction is performed on the key coupling features using the second convolutional layer and the second pooling layer of the convolutional neural network to obtain higher-order coupling features. The higher-order coupling features are flattened using a fully connected layer to obtain a coupling feature vector. Finally, the coupling feature vector and the temporal feature sequence are fused to obtain the fused features. Long Short-Term Memory (LSTM) networks and attention mechanisms are used to extract and process the short-term and long-term temporal dependencies of the fused features to obtain the concentration prediction value. Based on the maximum and minimum formaldehyde concentration values, the normalized prediction value is transformed to obtain the actual formaldehyde concentration value. Compared with the existing technology, which has insufficient coupling effect processing capability and separation of temporal and spatial features, this invention uses a convolutional neural network to specifically extract multi-factor coupling features and combines a LTM network to capture the temporal dynamic law of formaldehyde release. By constructing a hybrid network structure of multi-dimensional coupling features, designing convolutional feature extraction and temporal dynamic modeling, high-precision prediction of indoor formaldehyde concentration on specific days after decoration is achieved. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the indoor formaldehyde concentration prediction method based on a fused neural network provided by the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of the indoor formaldehyde concentration prediction device based on a fusion neural network provided by the present invention.

[0022] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] The following is combined Figures 1-3 This invention is described.

[0025] Figure 1 This is a flowchart illustrating the indoor formaldehyde concentration prediction method based on a fused neural network provided by the present invention. Figure 1 As shown, the method includes the following: Step 101: Quantize and construct the collected multidimensional static features and one-dimensional temporal features respectively to obtain the static feature matrix and the temporal feature sequence.

[0026] In step 101 above, the present invention specifically relates to an indoor formaldehyde concentration prediction model based on CNN-LSTM fusion neural network, which is particularly suitable for comprehensively processing the coupling effect of multiple factors and accurately predicting the formaldehyde concentration at any time point after decoration. It can be applied to scenarios such as building decoration quality assessment, decoration scheme optimization, indoor environmental monitoring, and health risk early warning.

[0027] This invention adopts a two-level fusion architecture of "CNN coupled feature extraction + LSTM temporal prediction", which specifically includes the following modules: feature input module, feature preprocessing module, CNN coupled feature extraction module, LSTM temporal prediction module, and loss function and inverse normalization module.

[0028] The feature input module is used to collect and quantify three-dimensional static features of "material-environment-space" and one-dimensional temporal features of "time-dynamic parameters" to construct a multi-dimensional input matrix with a total of 11 core parameters. These include static features and temporal features. Static features are used for coupled extraction, covering material type, used area, initial release rate, superposition coefficient, average temperature and humidity, room volume, and furniture shading coefficient. Temporal features are used for dynamic prediction, covering temperature and humidity, air exchange rate, and number of days after renovation for the 7 days prior to the prediction date.

[0029] Static characteristics include: material properties, environmental benchmarks, and spatial parameters.

[0030] Material properties include: material type; area of ​​a single material used, in square meters (m²). 2 Initial release rate of the material, in mg / (m³). 2.h), tested according to GB 18580-2017 standard; material superposition coefficient. Material type is determined using One-Hot coding. This includes, but is not limited to, five common material categories such as particleboard and latex paint.

[0031] Environmental benchmarks include: daily average temperature in °C; and daily average relative humidity in %RH.

[0032] Spatial parameters include: net room volume, in meters (m²). 3 Furniture occlusion coefficient, with a value of 0-1, where 1 is for no occlusion and 0 is for complete occlusion.

[0033] The temporal characteristics include: dynamic environment, ventilation parameters, and time stamps.

[0034] The dynamic environment includes: the daily average temperature for the 7 days prior to the forecast, in °C; and the daily average relative humidity, in %RH.

[0035] Ventilation parameters include: the average daily air exchange rate for the 7 days prior to the forecast, in hours (h). -1 .

[0036] The time stamp includes: the number of days after the renovation was completed, in days (d).

[0037] Static features are constructed as a 2D feature matrix [number of samples, static feature dimension, 1], with dimension = 8; temporal features are constructed as a 1D sequence [number of samples, temporal length, temporal feature dimension], with temporal length = 7 and dimension = 3.

[0038] Optionally, step 101 includes steps A1 to A2: Step A1: Reconstruct the collected multidimensional static features by mapping the multidimensional static features to a two-dimensional matrix to obtain the static feature matrix.

[0039] Step A2: Use a sliding window to extract one-dimensional time series features through a preset window to generate multiple time series samples; construct the multiple time series samples to obtain a time series feature sequence.

[0040] In steps A1 to A2 above, the 8-dimensional static features are reconstructed into a 2D matrix, where rows represent material-environment / space, and columns represent specific parameters. This simulates the spatial correlation structure of multiple factors, adapting to the convolutional operations of CNNs.

[0041] Sliding window extraction: A 7-day window is used to generate time-series samples, corresponding to the concentration on day 1 after the prediction window ends. Standardization: Z-score standardization. , where μ is the mean and σ is the standard deviation. This eliminates the dimensional effects of time-series fluctuations.

[0042] Optionally, before step 101, steps B1 to B2 are also included: Step B1: Identify outliers in the multidimensional static features using the box plot method, and replace outliers with the median of the same material type; normalize the multidimensional static features.

[0043] Step B2: Fill in the short-term missing data in the one-dimensional time series features using linear interpolation.

[0044] In steps B1 and B2 above, different characteristics of static and temporal features are processed separately to ensure input quality. Outlier removal: Outliers such as material release rate and area are identified using box plots and replaced with the median of the same material type; Normalization: Min-Max normalization is performed to the [0,1] interval, using the formula... .

[0045] Step 102: Extract key coupling features from the static feature matrix using the first convolutional layer and the first pooling layer of the convolutional neural network; extract higher-order coupling features from the key coupling features using the second convolutional layer and the second pooling layer of the convolutional neural network; flatten the higher-order coupling features using a fully connected layer to obtain the coupling feature vector.

[0046] Optionally, step 102 includes steps C1 to C4: Step C1: Through the first convolutional layer of the convolutional neural network, the static feature matrix is ​​traversed according to the first kernel size and the first step length to obtain the local coupled feature map.

[0047] Step C2: Max pooling is performed on the local coupled feature map through the first pooling layer of the convolutional neural network to obtain the key coupled features.

[0048] Step C3: Through the second convolutional layer of the convolutional neural network, the key coupling features are traversed according to the second kernel size and the second stride to obtain a high-order coupling feature map.

[0049] Step C4: Perform average pooling on the high-order coupled feature map through the second pooling layer of the convolutional neural network to obtain the high-order coupled features.

[0050] In steps C1 to C4 above, a 1D convolutional network is used to specifically extract the multi-factor coupling relationship in static features. The interaction between parameters is captured through local receptive fields. The structure is as follows: Convolutional layer 1: Set 16 convolutional kernels, kernel size = 3, covering the interaction of 3 adjacent features, stride = 1, activation function is ReLU, input is 2D static feature matrix, output is 16-dimensional local coupling feature map.

[0051] Pooling layer 1: Max pooling, pooling size = 2, stride = 1, retains key coupling features, and reduces dimensionality.

[0052] Convolutional layer 2: Set 32 ​​convolutional kernels, kernel size = 2, stride = 1, activation function is ReLU, to further extract higher-order coupling features, such as the ternary coupling of "temperature-material release-ventilation rate".

[0053] Pooling layer 2: Average pooling, pooling size = 2, step size = 1; Fully connected layer: Flattens the pooled feature map into a 64-dimensional coupled feature vector, which serves as one of the inputs to the LSTM module.

[0054] This module automatically learns the nonlinear coupling patterns between multiple factors through multi-layer convolution operations, such as the differential effects of high temperature and high humidity on the release rates of different materials, without the need for manual design of coupling terms.

[0055] Step 103: Perform feature fusion on the coupled feature vector and the temporal feature sequence to obtain the fused features; extract and process the short-term and long-term temporal dependencies of the fused features through a long short-term memory network and attention mechanism to obtain the concentration prediction value.

[0056] Optionally, step 103 includes steps D1 to D4: Step D1: Extract short-term and long-term temporal dependencies from the fused features using a long short-term memory network to obtain short-term and long-term features.

[0057] Step D2: Using the dropout layer, short-term and long-term features are processed to prevent overfitting according to the first dropout rate, resulting in a fused feature sequence.

[0058] Step D3: Calculate the weights of each time step in the fused feature sequence using an attention mechanism, and obtain the weighted comprehensive feature vector based on the weights of each time step.

[0059] Step D4: Process the weighted integrated feature vector using the linear activation function of the output layer to obtain the concentration prediction value.

[0060] Optionally, step D1 includes steps D11 to D12: Step D11: Through the first temporal feature extraction layer of the Long Short-Term Memory Network, extract the short-term temporal dependence of the fused features according to the first number of neurons to obtain short-term features.

[0061] Step D12: Through the second temporal feature extraction layer of the Long Short-Term Memory Network, long-term temporal dependencies are extracted from the fused features according to the second number of neurons to obtain long-term features.

[0062] In steps D1 to D4 above, the coupled features extracted by the CNN are fused with the preprocessed temporal features, and the temporal dynamics of formaldehyde release are captured by LSTM, as shown in the following structure: Feature fusion layer: The 64-dimensional coupled feature vector is concatenated with the 7×3-dimensional temporal features to construct a fused feature sequence [number of samples, 7, 67], where the temporal length = 7 and the fused dimension = 64+3=67.

[0063] LSTM layer 1: 64 neurons, activation function is tanh, forget gate threshold = 0.5, extracting short-term temporal dependencies, such as the effect of temperature and humidity fluctuations on concentration over the past 3 days.

[0064] LSTM layer 2: 128 neurons, activation function is tanh, extracting long-term temporal dependencies, such as the release decay trend from 1 to 180 days after renovation.

[0065] Dropout layer: Dropout rate = 0.2, to prevent overfitting.

[0066] Output layer: Linear activation function, outputting the formaldehyde concentration for the predicted day, i.e., the normalized value.

[0067] To further improve accuracy, an attention mechanism is introduced: a temporal attention layer is added between the LSTM layer and the output layer to calculate the weights of each time step in the fused feature sequence, using the following formula: .

[0068] Where st is the feature importance score at time step t, which is calculated through a fully connected layer. The weighted feature vector is the sum of the products of the features and weights at each time step, enhancing the model's attention to key time nodes, such as days of sudden ventilation changes.

[0069] The LSTM temporal prediction module first concatenates the 64-dimensional coupled features output by the CNN with the 7×3-dimensional temporal features to form a 7×67-dimensional fused sequence. Then, it extracts the temporal dependencies through two LSTM layers (64+128 neurons) and calculates the weights of each time step using an attention mechanism.

[0070] Step 104: Based on the maximum and minimum formaldehyde concentration values, convert the normalized predicted values ​​to obtain the actual formaldehyde concentration values.

[0071] In step 104 above, the loss function is designed as follows: a composite loss function of "mean squared error (MSE) + coupling feature regularization" is used to constrain the model to learn effective coupling features. Where y is the actual concentration. The predicted values ​​are represented by Wconv, where Wconv represents the weights of the CNN convolutional kernel. The regularization coefficient is 0.001-0.01, preferably 0.005.

[0072] The loss function is a weighted sum of MSE and L2 regularization of the CNN convolution kernel, with regularization coefficients ranging from 0.001 to 0.01.

[0073] Inverse normalization: Using the maximum and minimum concentration values ​​(Cmax and Cmin) saved in the static feature preprocessing stage, the predicted normalized values ​​are converted into actual concentrations. .

[0074] This invention extracts multi-factor coupling features using CNN, eliminating the need for manual design of coupling terms. The prediction error for the "temperature and humidity-material-ventilation" coupling scenario is ≤12%, which is 35% lower than that of the pure LSTM model, and the coupling effect is handled more accurately.

[0075] This invention uses CNN to capture static coupling features and LSTM to capture dynamic temporal features. The fusion architecture takes into account both spatial correlation and temporal dynamics, significantly improving generalization and achieving a cross-household prediction error of ≤15%. It realizes spatiotemporal feature collaborative modeling.

[0076] This invention solves the problem that traditional models struggle to balance coupling effects and temporal characteristics, achieving a prediction relative error ≤12% and R0. 2 It has a value of ≥0.96 and the coupling contribution can be visualized through the convolution kernel weights. It is suitable for scenarios such as building decoration assessment and indoor environmental monitoring, and it is also suitable for formaldehyde concentration prediction for different types of houses.

[0077] This invention visualizes the convolutional kernel weights of a CNN, clarifying the coupling strength between various factors. For example, the weights of the second-layer convolutional kernel show that the coupling contribution of "temperature × particleboard release" is the largest, solving the "black box" problem of traditional neural networks. This enhances interpretability.

[0078] The input features of this invention can be obtained from a renovation list and conventional sensors. After model training, the prediction time is less than 0.05 seconds, which is convenient for engineering applications. It is highly user-friendly.

[0079] To further explain the present invention, the following specific examples are provided.

[0080] The model training process includes: sample collection, training parameter setting, and model evaluation.

[0081] Sample collection includes laboratory samples, field samples, and dataset partitioning.

[0082] Laboratory sample: 1m 3 Inside the climate chamber, five types of materials were selected: particleboard, latex paint, MDF, wallpaper, and adhesives; the temperature was set at 18-35℃, including seven gradients; the humidity was set at 30%-85%, including four gradients; and the air exchange rate was 0.5-3 hours. -1 It includes 6 gradients, with each working condition monitored for 180 days, and 1260 sets of samples collected.

[0083] On-site sample: 100+ different apartment types were selected, with room volumes ranging from 30-450m². 3 Actual renovated rooms with different material combinations were continuously monitored for 180 days, covering the period of rapid formaldehyde release; the daily average concentration was collected for 24 hours. Dataset partitioning: The preprocessed dataset is divided into a training set (70%), a validation set (20%), and a test set (10%) in a 7:2:1 ratio. The test set must include samples of "new material combination + new apartment type" that were not used in training to ensure the evaluation of the model's generalization ability.

[0084] Training parameter settings include optimizer and device.

[0085] Optimizer: Adamw, weight decay = 0.001, initial learning rate = 0.001, learning rate decay strategy: decays by 50% every 50 rounds. Batch size = 32, number of iterations = 200. Early stopping condition: validation set loss does not decrease for 8 consecutive rounds.

[0086] Equipment: NVIDIA Tesla V100 GPU, training time approximately 12 hours.

[0087] The test set, including laboratory and field samples, totaled 492 groups. The evaluation results are shown in Table 1.

[0088] Table 1 index This invention model Pure LSTM model Physical Model <![CDATA[R 2 ]]> 0.96 0.81 0.70 <![CDATA[MAE(mg / m 3 )]]> 0.06 0.11 0.15 relative error 11% 28% 36% The following is a prediction example.

[0089] At 120m 2 Three-bedroom apartment with a net floor area of ​​300 square meters. 3 For example, predict the formaldehyde concentration 45 days after renovation.

[0090] Static characteristics: Particleboard furniture 20m 2 Release rate 0.18 mg / (m³) 2 .h); Latex paint wall 60m 2 Release rate 0.06 mg / (m 2 .h), superposition coefficient 1.5; daily average temperature and humidity 26℃, 55% RH; furniture shading coefficient 0.8.

[0091] Temporal characteristics: Temperature and humidity for the first 7 days: 24-28℃, 50%-60% RH; air exchange rate: 1-1.5h. -1 45 days after renovation.

[0092] Static features are coupled to form a 64-dimensional feature vector using a CNN, which is then fused with temporal features to form a 7×67-dimensional sequence. An LSTM with an attention mechanism outputs a normalized predicted value of 0.58. Inverse normalization is performed on training set concentrations ranging from 0.02 to 0.22 mg / m³. 3 The actual concentration = 0.58 × (0.22 - 0.02) + 0.02 = 0.136 mg / m³ 3 The standard is exceeded.

[0093] The renovation optimization suggestions include: Based on CNN convolutional kernel weight analysis, the "temperature-particleboard release" coupling contributes the most, accounting for 28% of the weight. Recommendation: Lower the indoor temperature to 22℃ and increase the ventilation rate to 2 hours. -1 The re-predicted concentration is 0.09 mg / m³. 3 The standard has been met.

[0094] This invention provides a method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network. The method involves quantifying and constructing collected multidimensional static features and one-dimensional temporal features to obtain a static feature matrix and a temporal feature sequence. Features are extracted from the static feature matrix using the first convolutional layer and the first pooling layer of a convolutional neural network to obtain key coupling features. Further feature extraction is performed on the key coupling features using the second convolutional layer and the second pooling layer of the convolutional neural network to obtain higher-order coupling features. These higher-order coupling features are then flattened using a fully connected layer to obtain a coupling feature vector. Finally, the coupling feature vector and the temporal feature sequence are fused to obtain the fused features. By extracting and processing the short-term and long-term temporal dependencies of the fused features using a long short-term memory network and an attention mechanism, a concentration prediction value is obtained. Based on the maximum and minimum formaldehyde concentration values, the normalized prediction value is transformed to obtain the actual formaldehyde concentration value. Compared with the existing technology, which has insufficient coupling effect processing capability and separation of temporal and spatial features, this invention uses a convolutional neural network to specifically extract multi-factor coupling features and combines a long short-term memory network to capture the temporal dynamic law of formaldehyde release. By constructing a hybrid network structure of multi-dimensional coupling features, designing convolutional feature extraction, and temporal dynamic modeling, high-precision prediction of indoor formaldehyde concentration for a specific number of days after decoration is achieved.

[0095] The following describes the indoor formaldehyde concentration prediction device based on fusion neural network provided by the present invention. The indoor formaldehyde concentration prediction device based on fusion neural network described below can be referred to in correspondence with the indoor formaldehyde concentration prediction method based on fusion neural network described above.

[0096] Figure 2 This is a schematic diagram of the process for predicting formaldehyde concentration in indoor decoration based on a fused neural network, as provided in this invention. Figure 2 As shown, the device includes the following: The feature input module 201 is used to quantize and construct the collected multidimensional static features and one-dimensional temporal features respectively, to obtain the static feature matrix and the temporal feature sequence.

[0097] The CNN coupling feature extraction module 202 is used to extract features from the static feature matrix through the first convolutional layer and the first pooling layer of the convolutional neural network to obtain key coupling features; to extract features from the key coupling features through the second convolutional layer and the second pooling layer of the convolutional neural network to obtain higher-order coupling features; and to flatten the higher-order coupling features through a fully connected layer to obtain a coupling feature vector.

[0098] The LSTM temporal prediction module 203 is used to fuse the coupled feature vector and the temporal feature sequence to obtain the fused features. The fused features are then processed by a long short-term memory network and an attention mechanism to extract and process the short-term and long-term temporal dependencies to obtain the concentration prediction value.

[0099] The inverse normalization module 204 is used to convert the normalized predicted value based on the maximum and minimum formaldehyde concentration values ​​to obtain the actual formaldehyde concentration value.

[0100] This invention provides a device for predicting formaldehyde concentration in indoor decoration based on a fusion neural network. The device quantifies and constructs collected multidimensional static features and one-dimensional temporal features to obtain a static feature matrix and a temporal feature sequence. It then extracts features from the static feature matrix using the first convolutional layer and the first pooling layer of a convolutional neural network to obtain key coupling features. Next, it extracts features from the key coupling features using the second convolutional layer and the second pooling layer of the convolutional neural network to obtain higher-order coupling features. Finally, it flattens the higher-order coupling features using a fully connected layer to obtain a coupling feature vector. Finally, it fuses the coupling feature vector and the temporal feature sequence to obtain the fused features. By extracting and processing the short-term and long-term temporal dependencies of the fused features using a long short-term memory network and an attention mechanism, a concentration prediction value is obtained. Based on the maximum and minimum formaldehyde concentration values, the normalized prediction value is transformed to obtain the actual formaldehyde concentration value. Compared with the existing technology, which has insufficient coupling effect processing capability and separation of temporal and spatial features, this invention uses a convolutional neural network to specifically extract multi-factor coupling features and combines a long short-term memory network to capture the temporal dynamic law of formaldehyde release. By constructing a hybrid network structure of multi-dimensional coupling features, designing convolutional feature extraction, and temporal dynamic modeling, high-precision prediction of indoor formaldehyde concentration for a specific number of days after decoration is achieved.

[0101] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840. The processor 810, communications interface 820, and memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions from the memory 830 to execute a method for predicting formaldehyde concentration in indoor decoration based on a fused neural network.

[0102] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0103] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the indoor decoration formaldehyde concentration prediction method based on fused neural networks provided by the above methods.

[0104] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the indoor formaldehyde concentration prediction method based on a fused neural network provided by the above methods.

[0105] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0106] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network, characterized in that, include: The collected multidimensional static features and one-dimensional temporal features are quantized and constructed to obtain a static feature matrix and a temporal feature sequence, respectively. The static feature matrix is ​​subjected to feature extraction through the first convolutional layer and the first pooling layer of the convolutional neural network to obtain key coupling features; the key coupling features are subjected to feature extraction through the second convolutional layer and the second pooling layer of the convolutional neural network to obtain higher-order coupling features; the higher-order coupling features are flattened through a fully connected layer to obtain a coupling feature vector. The coupled feature vector and the temporal feature sequence are fused to obtain fused features; the fused features are then processed using a long short-term memory network and an attention mechanism to extract and process short-term and long-term temporal dependencies, thereby obtaining concentration prediction values. Based on the maximum and minimum formaldehyde concentrations, the normalized predicted values ​​are converted to obtain the actual formaldehyde concentration values.

2. The method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network according to claim 1, characterized in that, Before quantizing and constructing the collected multidimensional static features and one-dimensional temporal features respectively, the process also includes: Outliers in the multidimensional static features are identified using the box plot method, and the median of the same material type is used to replace the outliers; the multidimensional static features are then normalized. Linear interpolation is used to fill in short-term missing data in one-dimensional time series features.

3. The method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network according to claim 1, characterized in that, The process of quantizing and constructing the collected multidimensional static features and one-dimensional temporal features to obtain a static feature matrix and a temporal feature sequence includes: The collected multidimensional static features are reconstructed by mapping multidimensional static features to a two-dimensional matrix to obtain a static feature matrix; Multiple time series samples are generated by using a sliding window to extract the one-dimensional time series features through a preset window; the multiple time series samples are then used to construct a time series feature sequence.

4. The method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network according to claim 1, characterized in that, The static feature matrix is ​​subjected to feature extraction through the first convolutional layer and the first pooling layer of the convolutional neural network to obtain key coupling features; the key coupling features are then subjected to feature extraction through the second convolutional layer and the second pooling layer of the convolutional neural network to obtain higher-order coupling features, including: By traversing the static feature matrix through the first convolutional layer of the convolutional neural network according to the first kernel size and the first step size, a local coupled feature map is obtained. The key coupling features are obtained by max pooling the local coupling feature map through the first pooling layer of the convolutional neural network. By using the second convolutional layer of the convolutional neural network, the key coupling features are traversed according to the second kernel size and the second stride to obtain a high-order coupling feature map; The higher-order coupling feature map is obtained by average pooling through the second pooling layer of the convolutional neural network.

5. The method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network according to claim 1, characterized in that, The process of extracting and processing short-term and long-term temporal dependencies of the fused features using a long short-term memory network and attention mechanism to obtain concentration prediction values ​​includes: The fused features are extracted using a long short-term memory network to obtain short-term and long-term temporal dependencies. By using a dropout layer, the short-term features and the long-term features are processed to prevent overfitting according to a first dropout rate, resulting in a fused feature sequence. The weights of each time step in the fused feature sequence are calculated using an attention mechanism, and a weighted comprehensive feature vector is obtained based on the weights of each time step. The concentration prediction value is obtained by processing the weighted integrated feature vector through the linear activation function of the output layer.

6. The method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network according to claim 5, characterized in that, The step of extracting short-term and long-term temporal dependencies from the fused features using a Long Short-Term Memory (LSTM) network to obtain short-term and long-term features includes: The first temporal feature extraction layer of the Long Short-Term Memory network extracts short-term temporal dependencies from the fused features according to a first number of neurons, thereby obtaining short-term features. The long-term features are obtained by extracting long-term temporal dependencies from the fused features through the second temporal feature extraction layer of the long short-term memory network according to the second number of neurons.

7. A device for predicting formaldehyde concentration in indoor decoration based on a fusion neural network, characterized in that, include: The feature input module is used to quantize and construct the collected multidimensional static features and one-dimensional temporal features respectively, to obtain the static feature matrix and the temporal feature sequence; The CNN coupling feature extraction module is used to extract features from the static feature matrix through the first convolutional layer and the first pooling layer of the convolutional neural network to obtain key coupling features; to extract features from the key coupling features through the second convolutional layer and the second pooling layer of the convolutional neural network to obtain higher-order coupling features; and to flatten the higher-order coupling features through a fully connected layer to obtain a coupling feature vector. The LSTM temporal prediction module is used to fuse the coupled feature vector and the temporal feature sequence to obtain the fused features; the fused features are then processed by a long short-term memory network and an attention mechanism to extract and process short-term and long-term temporal dependencies to obtain the concentration prediction value. The inverse normalization module is used to convert the normalized predicted value based on the maximum and minimum formaldehyde concentration values ​​to obtain the actual formaldehyde concentration value.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method for predicting indoor formaldehyde concentration based on a fused neural network as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for predicting formaldehyde concentration in indoor decoration based on a fusion neural network as described in any one of claims 1 to 6.

Images