Building fireproof safety risk assessment method

Abstract

The invention belongs to the field of safety risk assessment, and particularly relates to a building fireproof safety risk assessment method. The invention provides a building fireproof safety risk assessment method, and aims to construct a building fireproof safety risk assessment model FSRAM, extract a long-term trend from time series data by introducing a long-mode extractor, and isolate short-term fluctuation and noise, so that the building fireproof safety risk assessment model FSRAM is obtained. A multi-scale decomposer is introduced to decompose time series data into a multi-scale approximate function and a multi-scale detailed function to capture different frequency components and detail features in the data, and then a feature enhancement module converts the time series data from a time domain to a frequency domain through Fourier transform. The method comprises the following steps of: establishing an improved U-net network model, processing time sequence data by using a multi-layer encoder and a decoder, and reducing information loss by using residual connection, so as to better capture periodicity, trend and frequency characteristics in the data, and finally constructing the improved U-net network model, and processing the time sequence data by using a multi-layer encoder and a multi-layer decoder.

Description

Technical Field

[0001] The invention belongs to the field of safety risk assessment, and in particular relates to a method for assessing building fire safety risks. Background Art

[0002] With the acceleration of urbanization and the increase in the number of buildings, the threat of fire to life and property has become increasingly serious. Fire safety risk assessment aims to identify potential fire hazards in the design, construction, use and maintenance of buildings through systematic analysis and evaluation, predict the possibility and severity of fire, and propose corresponding prevention and control measures. Traditional assessment methods usually include inspection and analysis of building structural materials, fire protection facilities, escape routes, fire loads and other factors, and use fire dynamic simulation technology and fire risk assessment models to evaluate the spread and degree of harm of fire. In recent years, with the development of artificial intelligence and big data technology, intelligent fire risk assessment methods have gradually emerged. Through deep learning and data mining of historical fire data and real-time monitoring data, it can more accurately predict the risk of fire and put forward personalized prevention and control suggestions. This method not only improves the accuracy and efficiency of the assessment, but also provides a scientific basis for the design and management of buildings, and further improves the overall fire safety level of buildings. Summary of the invention

[0003] The main purpose of the present invention is to provide a method for fire safety risk assessment of buildings, aiming to construct a building fire safety risk assessment model FSRAM, by introducing a long pattern extractor to extract long-term trends from time series data and isolate short-term fluctuations and noise, introducing a multi-scale decomposer to decompose the time series data into multi-scale approximate functions and detailed functions to capture different frequency components and detail features in the data, and then a feature enhancement module converts the time series data from the time domain to the frequency domain through Fourier transform, so as to better capture the periodicity, trend and frequency characteristics in the data, and finally construct an improved U-net network model, use multi-layer encoders and decoders to process the time series data, and use residual connections to reduce information loss.

[0004] To achieve the above object, the technical solution of the present invention is: a method for assessing building fire safety risk, the method comprising: S1. Collect building fire risk assessment data. The collected data types include basic building information data, fire protection facilities and equipment information, number of escape and evacuation routes, fire history data and monitored environmental information; S2. Perform feature engineering on the collected data, including missing value processing and normalization operations, to improve the quality of the data set and make it dimensionless, and divide the processed data into training set, validation set and test set; S3. Construct a building fire safety risk assessment model FSRAM. The specific steps are as follows: S31. The long mode extractor LME is proposed, which mainly extracts long-term trends from time series data and isolates short-term fluctuations and noise; S32. A multi-scale decomposer Multi-scaleD is proposed, which mainly decomposes time series data into multi-scale main functions and auxiliary functions to capture different frequency components and detailed features in the data; S33, proposed feature enhancement module FEM, the main function of which is to transform time series data from time domain to frequency domain through Fourier transform, so as to better capture the periodicity, trend and frequency characteristics in the data; S34. Build an improved U-net network model, use multi-layer encoders and decoders to process time series data, and use residual connections to reduce information loss; S4. Use building fire risk assessment data to train a building fire safety risk assessment model so that the model has a higher prediction accuracy; S5. Use the building fire safety risk assessment model for real-time prediction, deploy the trained risk assessment model, and use the test set for testing.

[0005] Furthermore, in step S3, the long mode extractor LME described in S31 is used to extract the long-term trend component X from the time series data. T and the residual component X R , the original multivariate time series data is X = {x1, x2, ..., x T}, input the original data X into the long pattern extractor, and finally get the long-term trend component X T ={X T1 , x T2 , ..., x TT} and the residual component X R ={x R1 , x R2 , ..., x RT}; The specific steps are as follows: First, the input time series data is smoothed using the filling average technique to extract the trend component X T , the formula used by the fill-average technique is as follows: X T =Avg Pool(Padding(X)) k ; In the formula, AvgPool is the average pooling operation, which is used to calculate the average value over a period of time. Padding is the padding operation, which is used to fill the data before the pooling operation to ensure that the pooling window can completely cover the data. k is the step size of the moving average, which determines the size of the time window considered when calculating the trend. In getting the trend component X T After that, calculate the residual component X R , that is, the difference between the original data and the trend component, the residual calculation formula is as follows: X R =XX T ; In the formula, X is the original input data, X T is the trend component obtained by moving average pooling; for each time step t, calculate the original data x t With the trend component x Tt The difference between Rt :x Rt =x t -x Tt ; In this way, the residual component X R Captures short-term fluctuations and noise in the raw data that are not captured by trend decomposition.

[0006] Furthermore, in step S3, the multi-scale decomposer Multi-scaleD described in S32 convolves the input data through a low-pass filter and a high-pass filter to obtain a main coefficient cA and an auxiliary coefficient cD, and the low-pass filtering operation formula is: Where h is the filter coefficient of the low-pass filter, and n is the index after downsampling; the high-pass filter operation formula is: Wherein, g is the filter coefficient of the high-pass filter, and n is the index after downsampling; the final main coefficient can be expressed as cA = {a1, a2, a3, ...}, and the auxiliary coefficient can be expressed as cD = {d1, d2, d3, ...}; After filtering, the signal is downsampled to reduce the signal length by half for multi-scale decomposition. The downsampling formula is: cA n =cA[::2]; cD n =cD[::2]; Where::2 takes a value for every other sample, which is a downsampling operation. This operation helps to reduce the amount of data while retaining important frequency components so that features can be extracted more effectively in multi-scale decomposition.

[0007] Furthermore, in step S3, the feature enhancement module FEM in step S33 combines the main coefficient cA and the auxiliary coefficient cD in step S3 and divides them into odd-indexed and even-indexed sequences, and the odd-indexed sequence can be expressed as: X odd ={a1,a3,a5,...,d1,d3,d5,...}; The even index sequence can be expressed as: X even ={a2,a4,a6,...,d2,d4,d6,...}; Subsequently, the separated odd and even sequences are interactively learned to extract richer features; the feature extraction of the odd index sequence is calculated as follows: In the formula, ⊙ is the element-wise multiplication operation, exp is the exponential operation, is a feedforward neural network, X odd is an odd-indexed sequence, X even is an even-indexed sequence, is the result of feature extraction for odd index sequences; The dual index sequence is used for feature extraction, and the calculation formula is: In the formula, is the result of feature extraction of dual index sequence, ψ is a feedforward neural network, and The difference is that their weights are different; Finally, after interactive learning, and Fourier transform is used to further enhance the features. The formula for Fourier enhancement operation is as follows: Where P is the Fourier enhancement function, X′ odd is the odd-order sequence feature after Fourier enhancement; right The formula for Fourier enhancement operation is as follows: Where U is the Fourier enhancement function, X′ even is the characteristic of the odd-order subsequence after Fourier enhancement.

[0008] Furthermore, in step S3, the improved U-net network model described in S34 is an important component of feature extraction and prediction, which combines the output after Fourier enhancement and uses a multi-level encoder and decoder structure to perform feature extraction, fusion and prediction. The improved U-net structure consists of an encoder, a decoder and a residual connection; First, the encoder part is used to extract high-level features of the input data. The first layer of the encoder operates as follows: In the formula, is a concatenation operation, Conv1 is the convolution operation of the first layer encoder, ReLU is the activation function, which adds nonlinearity, and E1 is the output of the first layer encoder; The second layer encoder operates as follows: E2=ReLU(Conv2(MaxPool(E1))); Where Conv2 is the convolution operation of the second-layer encoder, MaxPool is the maximum pooling operation, and E2 is the output of the second-layer encoder; The third layer encoder operates as follows: E3=ReLU(Conv3(MaxPool(E2))); Where Conv3 is the convolution operation of the third-layer encoder, and E3 is the output of the third-layer encoder; The decoder part is then used to restore the high-level features extracted by the encoder to the size of the original data. The decoder part operates in a residual connection manner, so that the output of the encoder layer can be directly passed to the corresponding decoder layer, effectively retaining the original feature information; The operation of the first layer decoder is as follows: Where ConvTrans1 is the deconvolution operation of the first layer decoder, UpSample is the upsampling operation, and D1 is the output of the first decoder; The operation of the second layer decoder is as follows: Where ConvTrans2 is the deconvolution operation of the second layer decoder, and D2 is the output of the second decoder; The operation of the third layer decoder is as follows: D3=ReLU(ConvTrans3(UpSample(D2))); Where ConvTrans3 is the deconvolution operation of the third layer decoder, and D3 is the output of the third decoder.

[0009] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: In the present invention, a building fire safety risk assessment model FSRAM is constructed. A long pattern extractor is introduced to extract long-term trends from time series data and isolate short-term fluctuations and noise. A multi-scale decomposer is introduced to decompose the time series data into multi-scale approximate functions and detailed functions to capture different frequency components and detail features in the data. Subsequently, a feature enhancement module converts the time series data from the time domain to the frequency domain through Fourier transform, so as to better capture the periodicity, trend and frequency characteristics in the data. Finally, an improved U-net network model is constructed, and multi-layer encoders and decoders are used to process the time series data, and residual connections are used to reduce information loss. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A flowchart of the steps of a method for risk assessment of building fire safety is shown.

[0011] Figure 2 The present invention is a flow chart of a building fire safety risk assessment model used in a building fire safety risk assessment method.

[0012] Figure 3 This is a structural diagram of a feature enhancement module FEM used in a building fire safety risk assessment method.

[0013] Figure 4 This is a structural diagram of an improved U-net network model used in building fire safety risk assessment methods.

[0014] Figure 5 It is a fitting graph of the predicted value and the true value of a building fire safety risk assessment model used in a building fire safety risk assessment method. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0016] See also Figure 1-Figure 5The present invention provides a technical solution: a method for assessing building fire safety risk, which constructs a building fire safety risk assessment model FSRAM, introduces a long pattern extractor to extract long-term trends from time series data, and isolates short-term fluctuations and noise, and introduces a multi-scale decomposer to decompose the time series data into multi-scale approximate functions and detailed functions to capture different frequency components and detail features in the data, and then a feature enhancement module converts the time series data from the time domain to the frequency domain through Fourier transform, so as to better capture the periodicity, trend and frequency features in the data, and finally constructs an improved U-net network model, uses multi-layer encoders and decoders to process the time series data, and uses residual connections to reduce information loss.

[0017] Please refer to Figure 1 As shown, a method for assessing building fire safety risk in an embodiment of the present application is as follows:

[0018] S1. Collect building fire risk assessment data. The collected data types include basic building information data, fire protection facilities and equipment information, number of escape and evacuation routes, fire history data and monitored environmental information.

[0019] Furthermore, in the step S1, among the data types collected, basic building information data includes: building type, building height and area, building structure, among which building types include residential, commercial, industrial, and public buildings, and building structures include steel structures, concrete structures, and brick-concrete structures; fire-fighting facilities and equipment information includes fire-fighting systems, fire-fighting equipment, and daily maintenance records, among which fire-fighting systems include sprinkler systems, fire extinguishing systems, fire alarm systems, and smoke exhaust system configurations, and fire-fighting equipment includes the number and distribution of fire extinguishers, fire hydrants, fire pumps, and fire alarm equipment; fire history data includes the time, location, and cause of the fire; and monitored environmental information includes the temperature, humidity, and air pressure of the internal and external environment of the building.

[0020] S2. Perform feature engineering on the collected data, including missing value processing and normalization operations, to improve the quality of the data set and make it dimensionless.

[0021] Furthermore, in step S2, the data set is processed for missing values ​​and normalized, and the data set is tested for missing values. Since enough data is collected, the present invention deletes data samples with missing values. Normalization is to process data according to the rows of the feature matrix, and its purpose is to map the values ​​in the sample to between 0 and 1, thereby eliminating the dimensional influence between the indicators. The data normalization formula is as follows: In the formula, x iis the sample data value to be normalized, min(x) is the minimum value in the attribute, max(x) is the maximum value in the attribute, x′ i is the data value after normalization.

[0022] Furthermore, in step S3, Figure 2 As shown in the figure, the model consists of a long pattern extractor LME, a multi-scale decomposer Multi-scaleD, a feature enhancement module FEM and an improved U-net module.

[0023] Furthermore, in step S3, the long mode extractor LME described in S31 is used to extract the long-term trend component X from the time series data. T and the residual component X R , the original multivariate time series data is X = {x1, x2, ..., x T}, input the original data X into the long pattern extractor, and finally get the long-term trend component X T ={x T1 , x T2 , ..., x TT} and the residual component X R ={x R1 , x R2 , ..., x RT}; The specific steps are as follows: First, the input time series data is smoothed using the filling average technique to extract the trend component X T , the formula used by the fill-average technique is as follows: X T =AvgPool(Padding(X)) k ; In the formula, AvgPool is the average pooling operation, which is used to calculate the average value over a period of time. Padding is the padding operation, which is used to fill the data before the pooling operation to ensure that the pooling window can completely cover the data. k is the step size of the moving average, which determines the size of the time window considered when calculating the trend. In getting the trend component X T After that, calculate the residual component X R , that is, the difference between the original data and the trend component, the residual calculation formula is as follows: X R =XX T ; In the formula, X is the original input data, X T is the trend component obtained by moving average pooling; for each time step t, calculate the original data x t With the trend component x Tt The difference between Rt :xRt =x t -x Tt ; In this way, the residual component X R Captures short-term fluctuations and noise in the raw data that are not captured by trend decomposition.

[0024] Furthermore, in steps S3 and S31, the long pattern extractor can extract long-term trends from the time series data and isolate short-term fluctuations and noise. By removing noise and short-term fluctuations in the data, the model can more accurately learn and predict fire risks. At the same time, the data insights provided can help managers make more informed decisions and improve the overall fire safety level of the building.

[0025] Furthermore, in step S3, the multi-scale decomposer Multi-scaleD described in S32 convolves the input data through a low-pass filter and a high-pass filter to obtain a main coefficient cA and an auxiliary coefficient cD, and the low-pass filtering operation formula is: Where h is the filter coefficient of the low-pass filter, and n is the index after downsampling; the high-pass filter operation formula is: Wherein, g is the filter coefficient of the high-pass filter, and n is the index after downsampling; the final main coefficient can be expressed as cA = {a1, a2, a3, ...}, and the auxiliary coefficient can be expressed as cD = {d1, d2, d3, ...}; After filtering, the signal is downsampled to reduce the signal length by half for multi-scale decomposition. The downsampling formula is: cA n =cA[::2]; cD n =cD[::2]; Where::2 takes a value for every other sample, which is a downsampling operation. This operation helps to reduce the amount of data while retaining important frequency components so that features can be extracted more effectively in multi-scale decomposition.

[0026] Furthermore, in steps S3 and S32, the multi-scale decomposer decomposes the time series data into multi-scale main functions and auxiliary functions to capture different frequency components and detail features in the data. By decomposing the data into components of different scales, important features and details in the data can be effectively extracted, making the model more accurate and robust when processing complex time series data, while providing multi-level data information for prediction.

[0027] Furthermore, in step S3, the feature enhancement module FEM described in S33 is as follows: Figure 3As shown, the main coefficient cA and the auxiliary coefficient cD in step S3 are combined and divided into odd and even index sequences, and the odd index sequence can be expressed as: X odd ={a1,a3,a5,...,d1,d3,d5,...}; The even index sequence can be expressed as: X even ={a2,a4,a6,...,d2,d4,d6,...}; Subsequently, the separated odd and even sequences are interactively learned to extract richer features; the feature extraction of the odd index sequence is calculated as follows: In the formula, ⊙ is the element-wise multiplication operation, exp is the exponential operation, is a feedforward neural network, X odd is an odd-indexed sequence, X even is an even-indexed sequence, is the result of feature extraction for odd index sequences; The dual index sequence is used for feature extraction, and the calculation formula is: In the formula, is the result of feature extraction of dual index sequence, ψ is a feedforward neural network, and The difference is that their weights are different; Finally, after interactive learning, and Fourier transform is used to further enhance the features. The formula for Fourier enhancement operation is as follows: Where P is the Fourier enhancement function, is the odd-order sequence feature after Fourier enhancement; right The formula for Fourier enhancement operation is as follows: Where U is the Fourier enhancement function, X′ even is the characteristic of the odd-order subsequence after Fourier enhancement.

[0028] Furthermore, in steps S3 and S33, the feature enhancement module converts the time series data from the time domain to the frequency domain through Fourier transform, so as to better capture the periodicity, trend and frequency characteristics in the data, and convert the time domain information into frequency domain information. The model can more clearly display the important features in the data, help the model better understand and learn the inherent laws of the data, and thus perform better in complex time series prediction tasks.

[0029] Furthermore, in step S3, the improved U-net network model described in S34 is an important component of feature extraction and prediction, such as Figure 4 As shown in the figure, the improved U-net network model consists of an input layer, a three-layer encoder, a three-layer decoder and an output layer. It combines the output after Fourier enhancement and uses a multi-level encoder and decoder structure for feature extraction, fusion and prediction. The improved U-net structure consists of an encoder, a decoder and a residual connection. First, the encoder part is used to extract high-level features of the input data. The first layer of the encoder operates as follows: In the formula, is a concatenation operation, Conv1 is the convolution operation of the first layer encoder, ReLU is the activation function, which adds nonlinearity, and E1 is the output of the first layer encoder; The second layer encoder operates as follows: E2=ReLU(Conv2(MaxPool(E1))); Where Conv2 is the convolution operation of the second-layer encoder, MaxPool is the maximum pooling operation, and E2 is the output of the second-layer encoder; The third layer encoder operates as follows: E3=ReLU(Conv3(MaxPool(E2))); Where Conv3 is the convolution operation of the third-layer encoder, and E3 is the output of the third-layer encoder; The decoder part is then used to restore the high-level features extracted by the encoder to the size of the original data. The decoder part operates in a residual connection manner, so that the output of the encoder layer can be directly passed to the corresponding decoder layer, effectively retaining the original feature information; The operation of the first layer decoder is as follows: Where ConvTrans1 is the deconvolution operation of the first layer decoder, UpSample is the upsampling operation, and D1 is the output of the first decoder; The operation of the second layer decoder is as follows: Where ConvTrans2 is the deconvolution operation of the second layer decoder, and D2 is the output of the second decoder; The operation of the third layer decoder is as follows: D3=ReLU(ConvTrans3(UpSample(D2))); Where ConvTrans3 is the deconvolution operation of the third layer decoder, and D3 is the output of the third decoder.

[0030] Furthermore, in steps S3 and S34, an improved U-net structure is used to perform building fire safety risk assessment, due to its powerful multi-scale feature extraction and fusion capabilities, effective information retention mechanism, and enhanced model robustness. The residual connection ensures the integrity of feature information during the transmission process, enabling the model to more accurately capture the fire safety status of buildings in different time periods and areas.

[0031] Furthermore, in step S4, after the output D3 of the improved U-net network model is input into the feedforward neural network, the final prediction value f(x) is output. The specific formula is as follows: y=W f ·D3+b f ; Where W f is the weight matrix of the feedforward neural network, D3 is the output of the improved U-net network model, and b f is a paranoid term; the building fire risk assessment data is used to train the building fire safety risk assessment model, and the mean square error loss function MSE is used to adjust the weight of the network. The mean square error loss function formula is as follows: In the formula, f i (x) is the predicted value of the i-th sample, y i is the true value of the i-th sample, n is the number of batch_size, batch_size is set to 512, the optimizer uses Adam, and the number of epoch training is set to 1000 times.

[0032] Furthermore, in step S5, the trained building fire safety risk assessment model is tested using a test set. Figure 5 This is the fitting of the model's predicted value and the true value of the sample. It can be seen from the figure that the difference between the model's predicted value and the true value of the sample is small, which proves that the model can accurately evaluate the building fire safety risk factor.

[0033] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for assessing building fire safety risk, characterized in that: The following steps are involved: S1. Collect building fire risk assessment data. The collected data types include basic building information data, fire protection facilities and equipment information, number of escape and evacuation routes, fire history data and monitored environmental information; S2. Perform feature engineering on the collected data, including missing value processing and normalization operations, to improve the quality of the data set and make it dimensionless, and divide the processed data into training set, validation set and test set; S3. Construct a building fire safety risk assessment model FSRAM. The specific steps are as follows: S31. The long mode extractor LME is proposed, which mainly extracts long-term trends from time series data and isolates short-term fluctuations and noise; S32. A multi-scale decomposer Multi-scaleD is proposed, which mainly decomposes time series data into multi-scale main functions and auxiliary functions to capture different frequency components and detailed features in the data; S33, proposed feature enhancement module FEM, the main function of which is to transform time series data from time domain to frequency domain through Fourier transform, so as to better capture the periodicity, trend and frequency characteristics in the data; S34. Build an improved U-net network model, use multi-layer encoders and decoders to process time series data, and use residual connections to reduce information loss; S4. Use building fire risk assessment data to train a building fire safety risk assessment model so that the model has a higher prediction accuracy; S5. Use the building fire safety risk assessment model for real-time prediction, deploy the trained risk assessment model, and use the test set for testing.

2. A method for building fire safety risk assessment according to claim 1, characterized in that: In step S3, the long mode extractor LME described in S31 is used to extract the long-term trend component X from the time series data. T and the residual component X R , the original multivariate time series data is X = {x1, x2, ..., x T }, input the original data X into the long pattern extractor, and finally get the long-term trend component X T ={x T1 , x T2 , ..., x TT } and the residual component X R ={x R1 , x R2 , ..., x RT }; The specific steps are as follows: First, the input time series data is smoothed using the filling average technique to extract the trend component X T , the formula used by the fill-average technique is as follows: X T =AvgPool(Padding(X)) k ; In the formula, AvgPool is the average pooling operation, which is used to calculate the average value over a period of time. Padding is the padding operation, which is used to fill the data before the pooling operation to ensure that the pooling window can completely cover the data. k is the step size of the moving average, which determines the size of the time window considered when calculating the trend. In getting the trend component X T After that, calculate the residual component X R , that is, the difference between the original data and the trend component, the residual calculation formula is as follows: X R =X-X T ; In the formula, X is the original input data, X T is the trend component obtained by moving average pooling; for each time step t, calculate the original data x t With the trend component x Tt The difference between Rt :x Rt =x t -x Tt ; In this way, the residual component X R Captures short-term fluctuations and noise in the raw data that are not captured by trend decomposition.

3. A method for building fire safety risk assessment according to claim 2, characterized in that: In step S3, the multi-scale decomposer Multi-scaleD described in S32 convolves the input data through a low-pass filter and a high-pass filter to obtain a main coefficient cA and an auxiliary coefficient cD. The low-pass filter operation formula is: Where h is the filter coefficient of the low-pass filter, and n is the index after downsampling; the high-pass filter operation formula is: Wherein, g is the filter coefficient of the high-pass filter, and n is the index after downsampling; the final main coefficient can be expressed as cA = {a1, a2, a3, ...}, and the auxiliary coefficient can be expressed as cD = {d1, d2, d3, ...}; After filtering, the signal is downsampled to reduce the signal length by half for multi-scale decomposition. The downsampling formula is: that n =cA[::2]; CD n =cD[::2]; Where: ::2 means taking a value for every other sample, which is a downsampling operation. This operation helps to reduce the amount of data while retaining important frequency components so that features can be extracted more effectively in multi-scale decomposition.

4. A method for building fire safety risk assessment according to claim 3, characterized in that: In step S3, the feature enhancement module FEM in step S33 combines the main coefficient cA and the auxiliary coefficient aD in step S3 and divides them into odd and even index sequences. The odd index sequence can be expressed as: <h2 style=";text-align:left;direction:ltr">X<h2 style=";text-align:left;direction:ltr"> odd <h2 style=";text-align:left;direction:ltr"> (a1, a3, a5,..., d1, d3, d5,...) Even index sequences can be represented as: <h2 style=";text-align:left;direction:ltr">X<h2 style=";text-align:left;direction:ltr"> even <h2 style=";text-align:left;direction:ltr"> (a2, a4, a6,..., d2, d4, d6,...) Subsequently, the separated odd and even sequences are interactively learned to extract richer features; The feature extraction of odd index sequence is calculated as follows: In the formula, ⊙ is the element-wise multiplication operation, exp is the exponential operation, is a feedforward neural network, X odd is an odd-indexed sequence, X even is an even-indexed sequence, is the result of feature extraction for odd index sequences; The dual index sequence is used for feature extraction, and the calculation formula is: In the formula, is the result of feature extraction of dual index sequence, ψ is a feedforward neural network, and The difference is that their weights are different; Finally, after interactive learning, and Fourier transform is used to further enhance the features. The formula for Fourier enhancement operation is as follows: Where P is the Fourier enhancement function, X′ odd is the odd-order sequence feature after Fourier enhancement; right The formula for Fourier enhancement operation is as follows: Where U is the Fourier enhancement function, X′ even is the characteristic of the odd-order subsequence after Fourier enhancement.

5. A method for building fire safety risk assessment according to claim 4, characterized in that: In step S3, the improved U-net network model described in S34 is an important component of feature extraction and prediction, which combines the output after Fourier enhancement and uses a multi-level encoder and decoder structure to perform feature extraction, fusion and prediction. The improved U-net structure consists of an encoder, a decoder and a residual connection; First, the encoder part is used to extract high-level features of the input data. The first layer of the encoder operates as follows: In the formula, is a concatenation operation, Conv1 is the convolution operation of the first layer encoder, ReLU is the activation function, which adds nonlinearity, and E1 is the output of the first layer encoder; The second layer encoder operates as follows: E2=ReLU(Conv2(MaxPoool(E1)))); Where Conv2 is the convolution operation of the second-layer encoder, MaxPool is the maximum pooling operation, and E2 is the output of the second-layer encoder; The third layer encoder operates as follows: E3=ReLU(Conv3(MaxPool((E2))); Where Conv3 is the convolution operation of the third-layer encoder, and E3 is the output of the third-layer encoder; The decoder part is then used to restore the high-level features extracted by the encoder to the size of the original data. The decoder part operates in a residual connection manner, so that the output of the encoder layer can be directly passed to the corresponding decoder layer, effectively retaining the original feature information; The operation of the first layer decoder is as follows: Where ConvTrans1 is the deconvolution operation of the first layer decoder, UpSample is the upsampling operation, and D1 is the output of the first decoder; The operation of the second layer decoder is as follows: Where ConvTrans2 is the deconvolution operation of the second layer decoder, and D2 is the output of the second decoder; The operation of the third layer decoder is as follows: D3=ReLU(ConvTrans3(UpSample(D2))); Where ConvTrans3 is the deconvolution operation of the third layer decoder, and D3 is the output of the third decoder.

6. A method for building fire safety risk assessment according to claim 5, characterized in that: In step S4, after the output D3 of the improved U-net network model is input into the feedforward neural network, the final predicted value f(x) is output. The specific formula is as follows: <h2 style=";text-align:left;direction:ltr">y=W<h2 style=";text-align:left;direction:ltr"> f <h2 style=";text-align:left;direction:ltr"> D3+b<h2 style=";text-align:left;direction:ltr"> f <h2 style=";text-align:left;direction:ltr"> ; Where W f is the weight matrix of the feedforward neural network, D3 is the output of the improved U-net network model, and b f is a paranoid term; the building fire risk assessment data is used to train the building fire safety risk assessment model, and the mean square error loss function MSE is used to adjust the weight of the network. The mean square error loss function formula is as follows: In the formula, f i (x) is the predicted value of the i-th sample, y i is the true value of the i-th sample.

Images