Tunnel fire source characteristic prediction and inversion method based on multi-subnet BP neural network

Abstract

The application discloses a tunnel fire source feature prediction and inversion method based on a multi-subnet BP neural network, and belongs to the technical field of tunnel fire safety, and comprises the following steps: S1, arranging temperature sensors to collect temperature monitoring data at each position; S2, inputting preprocessed data into a multi-subnet BP neural network fusion model which is pre-trained, and outputting a tunnel ceiling temperature distribution curve; S3, determining the longitudinal coordinate and the transverse coordinate of the fire source position according to the peak temperature position on the temperature distribution curve; and S4, substituting the tunnel ceiling temperature rise prediction value into an empirical correlation formula of the ceiling temperature rise, and inversely calculating the heat release rate of the fire source by using a numerical iteration solving method. The application adopts a multi-subnet parallel fusion structure, effectively suppresses the overfitting problem of a single network, realizes the integration of temperature field prediction, fire source positioning and heat release rate inversion in combination with the empirical correlation formula of the ceiling temperature rise, has high prediction precision, and the inversion result is reliable.

Description

Technical Field

[0001] This invention relates to the field of tunnel fire safety technology, and in particular to a method for predicting and inverting the fire source characteristics of tunnel fires based on a multi-subnet BP neural network. Background Technology

[0002] Tunnels are an important component of highway transportation. Once a tunnel fire occurs, its narrow, semi-enclosed space allows smoke to accumulate rapidly and temperatures to rise sharply, posing a serious threat to personnel safety and the integrity of the tunnel structure. Accurately predicting the temperature distribution of the tunnel ceiling, quickly locating the fire source, and quantitatively calculating the heat release rate of the fire source are of significant engineering importance for early fire warning, emergency decision-making, and fire rescue.

[0003] Currently, temperature monitoring in tunnel fires mainly relies on contact or single-point temperature measurement methods such as thermocouples, distributed optical fibers, and infrared thermal imagers. Regarding temperature prediction, existing research mostly employs single backpropagation (BP) neural networks or simple machine learning models for temperature fitting and prediction. In recent years, some researchers have attempted to combine particle swarm optimization (PSO) with a BP neural network (PSO-BPNN) for predicting ceiling temperature in tunnel fires with longitudinal ventilation, or to use a genetically optimized BP neural network (GA-BP) for predicting the temperature field in tunnel fires, as well as utilizing deep learning models such as LSTM and Transformer to predict the temperature field in tunnel fires.

[0004] However, the aforementioned existing technologies still have the following defects and shortcomings: First, single BP neural networks have weak generalization ability and are prone to overfitting. When faced with complex fire conditions, their temperature prediction errors are large and cannot meet engineering accuracy requirements. Although optimized algorithms such as PSO-BPNN and GA-BP have improved prediction accuracy to some extent, they still rely on a single network structure and have not fundamentally solved the problems of overfitting and insufficient generalization ability.

[0005] Second, existing methods rarely involve the quantitative inversion of the heat release rate of the fire source. The heat release rate of the fire source is a key parameter for assessing fire hazards and making emergency decisions, but it is difficult to measure directly in actual fires. Most current studies only focus on the positive prediction of the temperature field and have failed to achieve a high-precision integrated prediction of the fire source location, the maximum ceiling temperature, and the heat release rate.

[0006] Third, existing methods do not fully utilize the spatial distribution characteristics of the temperature field, and the neural network structure is too simple and the parameters lack systematic optimization. Sensor deployment relies heavily on engineering experience and lacks quantitative basis, resulting in an excessive number of sensors or unreasonable placement, making it difficult to achieve high-precision coverage of the entire temperature field with a small number of sensors. Summary of the Invention

[0007] The purpose of this invention is to provide a method for predicting and inverting the characteristics of tunnel fire sources based on a multi-subnet BP neural network. This method aims to overcome the shortcomings of existing single BP neural networks, such as weak generalization ability, easy overfitting, and large prediction errors, and to achieve integrated high-precision prediction and quantitative inversion of tunnel ceiling temperature distribution, fire source location, and fire source heat release rate under conditions with few sensors.

[0008] To achieve the above objectives, this invention provides a method for predicting and inverting the fire source characteristics of tunnel fires based on a multi-subnet BP neural network, comprising the following steps: S1. Deploy temperature sensors and collect temperature monitoring data at various locations through the temperature sensors. The temperature monitoring data includes the spatial coordinates of each temperature sensor and the corresponding temperature value. Preprocess the acquired data. S2. Input the pre-processed spatial coordinates of each temperature sensor and the corresponding temperature value into the pre-trained multi-subnet BP neural network fusion model, and output the temperature distribution curve of the tunnel roof through the temperature monitoring data at a limited number of measuring points. S3. Determine the longitudinal and transverse coordinates of the fire source location based on the peak temperature position on the temperature distribution curve of the tunnel ceiling.

[0009] S4. Substitute the location of the fire source in the tunnel and the corresponding predicted temperature value determined in the temperature distribution curve of the tunnel ceiling into the empirical correlation formula of the ceiling temperature rise, and use the numerical iterative solution method to back-calculate the heat release rate of the fire source.

[0010] Preferably, in step S1, the temperature sensors are installed below the tunnel ceiling, with a series of temperature sensors installed along the longitudinal direction of the tunnel, and a series of temperature sensors installed every 8m along the transverse direction of the tunnel.

[0011] Preferably, the training process of the pre-trained multi-subnet BP neural network fusion model is as follows: Construct a sample set; the sample set includes: the spatial location coordinates of multiple temperature sensors and their corresponding temperature values, as well as their corresponding label values; the labels represent the ceiling temperature; The sample set is preprocessed; the preprocessing includes: smoothing the temperature sequence arranged according to spatial location, and standardizing the spatial location coordinates and the corresponding temperature values ​​respectively; The sample set is filtered to obtain key samples, and the key samples are included in the training set. The remaining data is divided into training set and test set according to a preset ratio. The filtering principle is to identify peak temperature points at different spatial locations from the existing data and use the sample index set corresponding to the peak temperature points as part of the training set. Construct a multi-subnetwork BP neural network fusion model composed of parallel fusion of multiple subnetworks; The training set is input into each sub-network for training. The loss function is used to evaluate the prediction error and update the sub-network parameters. The sample weights are periodically adjusted during the training process. After training, the fusion weights are assigned based on the prediction accuracy of each sub-network. The weighted sum of the prediction values ​​of each sub-network is used as the final temperature prediction value of the multi-sub-network BP neural network fusion model. The final temperature prediction is restored to the actual temperature value through destandardization. The destandardization formula is as follows: ; in, The actual temperature value after reduction, in °C; The standardized temperature prediction value output by the multi-subnet BP neural network fusion model, i.e., the final temperature prediction value. ; The standard deviation of the temperature values; This represents the average temperature value.

[0012] Preferably, the preprocessing includes: S21. The temperature sequence is smoothed using a Gaussian filter, with the following formula: ; in, For the original temperature sequence at position The value at that location, For smoothed position The value at that location, These are the Gaussian filter smoothing coefficients; S22. Spatial position coordinates and the corresponding temperature value Standardize them separately. The standardization formula is: ; ; in, This represents the mean value corresponding to the spatial location coordinates. , The number of samples. For sample number, For the first The location coordinates of each sample; The standard deviation of the spatial location coordinates. ; For standardized spatial coordinates, This is the standardized temperature value.

[0013] Preferably, the multi-subnetwork BP neural network fusion model consists of 7 subnetworks fused in parallel. Each subnetwork contains one input layer neuron, a first hidden layer of 16 neurons, a second hidden layer of 32 neurons, a third hidden layer of 16 neurons, and one output layer neuron. All hidden layers use the ReLU activation function, and the output layer uses a linear activation function. The forward propagation formula for a single subnetwork is: ; ; ; ; in, This is the weight matrix from the input layer to the first hidden layer. This is the weight matrix from the first hidden layer to the second hidden layer. This is the weight matrix from the second hidden layer to the third hidden layer. This is the weight matrix from the third hidden layer to the output layer. For each layer's bias vector, For activation function, These are the output values ​​of each hidden layer after activation by the ReLU activation function. These are the output values ​​of each hidden layer after linear transformation. For standardized spatial coordinates, It is the first The standardized temperature prediction value output by each sub-network. Number the sub-network. .

[0014] Preferably, the loss function of each sub-network during the training process is: ; in, For the first The training loss of each subnetwork For the number of training sets, For the first The standardized temperature value of each sample. For the first Subnetworks for the first Standardized temperature prediction values ​​for each sample. For sample number, Number the subnetwork.

[0015] Preferably, during the training process, the multi-subnetwork BP neural network fusion model uses the Adam optimizer to update the subnetwork parameters. The update formula for the Adam optimizer is: ; ; ; ; ; in, , They are respectively The first and second momentum at time t. =0.9、 =0.999 is the momentum decay coefficient. loss function For parameters gradient, , This is the momentum value after deviation correction. These are the weight parameters of the subnetwork. For the first The parameter values ​​after the next iteration. For learning rate, =10 -8 To prevent the minimum value where the denominator is 0.

[0016] Preferably, during the training process, the multi-subnet BP neural network fusion model updates the sample weights after every 1000 training rounds, assigning higher weights to samples with larger absolute prediction errors; the sample weight update formula is: ; in, For the first Temperature values ​​of each sample For the first In the subnetwork, the first The weights of each sample satisfy... , For the first Subnetworks for the first The predicted temperature value for each sample. For the number of training sets, 、 All are sample numbers.

[0017] Preferably, the formula for allocating the fusion weights of each subnetwork in the multi-subnetwork BP neural network fusion model is as follows: ; ; in, For the first The root mean square error of each subnetwork on the training set For the first Temperature values ​​of each sample For the first Subnetworks for the first The predicted temperature value for each sample. For the first The fusion weights of each subnetwork satisfy , The total number of sub-networks, , For sample number, This represents the size of the training set.

[0018] Preferably, the formula for the final temperature prediction value of the multi-subnet BP neural network fusion model is: ; in, This represents the final temperature prediction value from the multi-subnet BP neural network fusion model. For the first The fusion weights of individual subnetworks For the first Temperature predictions after denormalization of the subnetwork The total number of sub-networks, , Number the subnetwork.

[0019] Therefore, the tunnel fire source feature prediction and inversion method based on multi-subnet BP neural network described above has the following significant advantages and beneficial effects: (1) Multi-subnet fusion suppresses overfitting, significantly improves prediction accuracy, and the temperature error does not exceed 5%.

[0020] (2) The number of sensors is small, the layout is clear, and the economy is strong: a series of sensors in the longitudinal direction and a series of sensors in the transverse direction of 8 m, making the project easy to implement.

[0021] (3) Combined with the empirical correlation of roof temperature rise, the physical meaning is clear and the accuracy is controllable. For the 3 MW working condition, the prediction error of the heat release rate of the fire source does not exceed 15%.

[0022] (4) It can be directly used for early warning of tunnel fires, situation assessment and emergency decision-making.

[0023] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the method for predicting and inverting the fire source characteristics of tunnel fires based on a multi-subnet BP neural network according to the present invention. Figure 2 This is a schematic diagram of the temperature sensor arrangement according to an embodiment of the present invention; Figure 3 This is a flowchart of the algorithm prediction process of the multi-subnet BP neural network fusion model according to an embodiment of the present invention; Figure 4 This is a heatmap of the root mean square error (RMSE) of 100 sets of parameters (learning rate and number of sub-networks) in Embodiment 1 of the present invention. Figure 5 The following are diagrams showing the fire source location and temperature prediction results according to an embodiment of the present invention: (a) a comparison between the predicted temperature and the measured value at the longitudinal centerline of the tunnel; (b) a comparison between the predicted temperature and the measured value at 4 m downstream of the tunnel. Figure 6 This is a diagram of the scaled-down experimental apparatus according to an embodiment of the present invention; Figure 7 This is a graph showing the predicted temperature results of the size reduction experiment in an embodiment of the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example like Figure 1 As shown, this embodiment provides a method for predicting and inverting the fire source characteristics of tunnel fires based on a multi-subnet BP neural network, including the following steps: S1. Deploy temperature sensors and collect temperature monitoring data at various locations. The temperature monitoring data includes the spatial coordinates of each temperature sensor and the corresponding temperature value. Preprocess the acquired data. like Figure 2 As shown, a series of temperature sensors are installed longitudinally along the tunnel, 0.1 m below the tunnel ceiling, and another series of temperature sensors are installed transversely every 8 m. This installation scheme balances the spatial coverage of the temperature field with the economic efficiency of the temperature sensor deployment. After the temperature sensors are installed, temperature monitoring data is collected at each sensor's location. The temperature monitoring data includes the spatial coordinates of each temperature sensor and its corresponding temperature value, forming a one-to-one data pair.

[0028] S2. Input the pre-processed spatial coordinates of each temperature sensor and the corresponding temperature value into the pre-trained multi-subnet BP neural network fusion model, and output the temperature distribution curve of the tunnel roof through the temperature monitoring data at a limited number of measuring points. The training process of the pre-trained multi-subnet BP neural network fusion model is as follows: First, a sample set is constructed. The sample set includes the spatial coordinates of multiple temperature sensors and their corresponding temperature values, as well as their corresponding label values, with the labels representing the ceiling temperature.

[0029] Then, the sample set is preprocessed. Preprocessing includes the following steps: The first step is to perform Gaussian filtering on the temperature series arranged according to spatial location to smooth the data and reduce random noise interference in the temperature monitoring data. The Gaussian filtering formula is: ; in, For the original temperature sequence at position The value at that location, For smoothed position The value at that location, This is the Gaussian filter smoothing coefficient, with a value of 2.0. This smoothing process can effectively eliminate minor noise introduced during sensor acquisition and subsequent calculations, while almost without losing key indicators such as the highest temperature value and the location of the fire source.

[0030] The second step is to determine the spatial location coordinates. and the corresponding temperature value Standardization is performed separately to eliminate the dimensional differences between spatial coordinates and temperature values, thus avoiding gradient imbalance caused by inconsistent numerical scales during model training. The standardization formula is: ; ; in, This represents the mean value corresponding to the spatial location coordinates. , The number of samples. For sample number, For the first The location coordinates of each sample; The standard deviation of the spatial location coordinates. ; For standardized spatial coordinates, This is the standardized temperature value.

[0031] The third step is to select key samples and divide the dataset. To ensure the model can learn the features of the temperature field peak region, when dividing the training set, peak temperature points at various spatial locations are identified from the existing samples. Samples corresponding to these peak temperature points are prioritized for inclusion in the training set, ensuring that the training set contains samples corresponding to peak temperature points. The remaining data is divided into training and test sets in a 7:3 ratio. The purpose of this is to allow the model to access key feature samples of the temperature field during the training phase, thereby improving the model's prediction accuracy for temperature peaks.

[0032] After completing the sample set preprocessing, a multi-subnet BP neural network fusion model is constructed. For example... Figure 3 As shown, the multi-subnetwork BP neural network fusion model consists of 7 subnetworks fused in parallel. Each subnetwork contains one input layer neuron, 16 neurons in the first hidden layer, 32 neurons in the second hidden layer, 16 neurons in the third hidden layer, and one output layer neuron. All hidden layers use the ReLU activation function, and the output layer uses a linear activation function.

[0033] The forward propagation process of a single subnetwork is as follows: The standardized spatial coordinates are... The input network undergoes linear transformation and ReLU activation in each layer, and finally outputs the standardized temperature prediction value from the output layer.

[0034] The forward propagation formula is: ; ; ; ; in, This is the weight matrix from the input layer to the first hidden layer. This is the weight matrix from the first hidden layer to the second hidden layer. This is the weight matrix from the second hidden layer to the third hidden layer. This is the weight matrix from the third hidden layer to the output layer. For each layer's bias vector, For activation function, These are the output values ​​of each hidden layer after activation by the ReLU activation function. These are the output values ​​of each hidden layer after linear transformation. For standardized spatial coordinates, It is the first The standardized temperature prediction value output by each sub-network. Number the sub-network. .

[0035] The training set is input into each sub-network for training. The root mean square error is used as the loss function for training each sub-network to measure the deviation between the predicted and the true values.

[0036] No. The loss function for each subnetwork is: ; in, For the first The training loss of each subnetwork For the number of training sets, For the first The standardized temperature value of each sample. For the first Subnetworks for the first Standardized temperature prediction values ​​for each sample. For sample number, Number the subnetwork.

[0037] The Adam optimizer is used to update the sub-network parameters. The Adam optimizer adaptively adjusts the learning rate by combining the first-order momentum (mean) and second-order momentum (variance) of the gradient. The parameter update formula is: ; ; ; ; ; in, , They are respectively The first and second momentum at time t. =0.9、 =0.999 is the momentum decay coefficient. loss function For parameters gradient, , This is the momentum value after deviation correction. These are the weight parameters of the subnetwork. For the first The parameter values ​​after the next iteration. For learning rate, =10 -8 To prevent the minimum value where the denominator is 0.

[0038] Sample weights are periodically adjusted during training. To enhance the model's learning of key temperature field samples (such as the highest temperature point and samples with large prediction errors), the sample weights of each sub-network are updated every 1000 training rounds, with samples having larger absolute prediction errors assigned higher weights. The sample weight update formula is: ; in, For the first Temperature values ​​of each sample For the first In the subnetwork, the first The weights of each sample satisfy... , For the first Subnetworks for the first The predicted temperature value for each sample. For the number of training sets, 、 All are sample numbers.

[0039] After training, fusion weights are assigned based on the training error of each sub-network. Sub-networks with smaller training errors are assigned higher fusion weights, and the sum of the fusion weights of all sub-networks is 1. The fusion weight allocation formula is: ; ; in, For the first The root mean square error of each subnetwork on the training set For the first Temperature values ​​of each sample For the first Subnetworks for the first The predicted temperature value for each sample. For the first The fusion weights of each subnetwork satisfy , The total number of sub-networks, , For sample number, This represents the size of the training set.

[0040] The weighted sum of the predicted values ​​from each subnetwork is used as the final temperature prediction value of the multi-subnetwork BP neural network fusion model, and the formula is: ; in, This represents the final temperature prediction value from the multi-subnet BP neural network fusion model. For the first The fusion weights of individual subnetworks For the first Temperature predictions after denormalization of the subnetwork The total number of sub-networks, , Number the subnetwork.

[0041] The final temperature prediction is restored to the actual temperature value through destandardization. The destandardization formula is as follows: ; in, The actual temperature value after reduction, in °C; The standardized temperature prediction value output by the multi-subnet BP neural network fusion model, i.e., the final temperature prediction value. ; The standard deviation of the temperature values; This represents the average temperature value.

[0042] Determine the optimal parameter combination. The prediction result of a BP neural network is related to its learning rate, and the number of subnetworks also affects the prediction result. For example... Figure 4 As shown, the root mean square error (RMSE) is used as the evaluation metric—the smaller the RMSE value, the smaller the deviation between the predicted and actual values, and the better the model's predictive performance. Through comparative experiments with 100 different combinations of learning rate and number of subnetworks, the analysis revealed that when the learning rate is 0.02 and the number of subnetworks is 7, the model's RMSE value reaches the minimum of 0.09766%. Therefore, the optimal parameter combination is determined as follows: learning rate... =0.002, Number of subnetworks =7.

[0043] After model training is complete, the prediction phase begins. The preprocessed spatial coordinates and corresponding temperature values ​​of each temperature sensor are input into the trained multi-subnet BP neural network fusion model. Because the temperature sensors are sparsely distributed vertically and horizontally at 8-meter intervals below the tunnel ceiling, the number of measuring points is limited, making it impossible to directly obtain continuous temperature distributions in all directions. The multi-subnet BP neural network fusion model learns the distribution pattern of the temperature field with spatial location from the training set, enabling it to predict the temperature at any location in the longitudinal and transverse directions of the entire tunnel based on a finite number of discrete measuring points, generating complete temperature distribution curves for the tunnel ceiling in all directions.

[0044] S3. Determine the longitudinal and transverse coordinates of the fire source location based on the peak temperature position on the temperature distribution curve of the tunnel ceiling.

[0045] Specifically, the temperature distribution curves of the tunnel ceiling predicted by two strings of sensors in the longitudinal and transverse directions are Gaussian smoothed to remove false extreme values ​​caused by random noise. The global maximum temperature value is then searched on the smoothed continuous curve, and the coordinates corresponding to this maximum value are the coordinates of the fire source.

[0046] like Figure 5 As shown, the vertical and horizontal coordinates of the fire source can be determined by the peak temperature position on the tunnel ceiling temperature distribution curve output by the model. Figure 5 In (a), you can see "Predicted location of fire source (0.11m)". Figure 5 In (b), we can see "Predicted location of fire source (0.51m)", which means the coordinates of the fire source are (0.11, 0.51). Figure 5 (a) shows a comparison between the predicted and measured values ​​of the temperature along the longitudinal centerline of the tunnel. Figure 5 Figure (b) shows a comparison between the predicted and measured temperatures 4 m downstream of the tunnel. The figure shows that the temperature prediction error is less than 5%, which accurately pinpoints the location of the fire source.

[0047] S4. Substitute the location of the fire source in the tunnel and the corresponding predicted temperature value determined in the temperature distribution curve of the tunnel ceiling into the empirical correlation of the temperature rise of the ceiling, and use the numerical iterative solution method to back-calculate the heat release rate of the fire source.

[0048] Substitute the predicted temperature values ​​into the empirical correlations for roof temperature rise in tunnel fires listed in Table 1. The empirical correlations in Table 1 cover a variety of operating conditions.

[0049] Table 1 Empirical Correlation of Temperature Rise in Tunnel Fire Ceiling

[0050] In the formulas in the table, The ceiling temperature rise (K). The heat release rate of the heat source (kW). The tunnel height is in meters (m). It is the ambient temperature. For the upstream air supply speed, It is gravitational acceleration. It is the dimensionless upstream tunnel length. The power of the dimensionless fire source; It is the maximum smoke temperature rise (K) at a certain lateral position. It is the distance (m) from the fire source. It is the distance (m) from the fire source to the side wall. It is the tunnel width (m); The temperature rise from the highest point of the ceiling Gas temperature (°C) at (m); The temperature rise of the ceiling at any longitudinal location in the tunnel. The difference in length between the upstream and downstream sides of the tunnel (m). The total length of the tunnel (m); These are the equation coefficients; The longitudinal distance (m) from the maximum temperature rise; , For upstream coefficients, , This is the downstream coefficient.

[0051] Based on the above correlation, a nonlinear mapping relationship is established between the roof temperature rise and the heat release rate of the fire source, the tunnel cross-sectional geometry, the initial ambient temperature, the distance between longitudinal measuring points, the location of the fire source, and the flow state of the plume. A numerical iterative solution method is used to back-calculate the heat release rate of the fire source based on this nonlinear mapping relationship.

[0052] This method can control the inversion calculation error within 15% under the condition of 3 MW fire source, and the inversion accuracy meets the engineering requirements for tunnel fire safety assessment and quantitative fire situation determination.

[0053] To further verify the accuracy and reliability of the prediction results of the multi-subnet BP neural network fusion model of this invention, a scaled-down experiment was used as an example for verification.

[0054] like Figure 6 As shown, the tunnel model used has a scale of 1:10 and dimensions of 6.0 m in length, 0.8 m in height, and 0.5 m in width. Within the range of 0–3.0 m, the thermocouples are spaced 0.1 m apart, totaling 31 thermocouples; within the range of 3.0–5.4 m, the thermocouples are spaced 0.3 m apart, totaling 8 thermocouples. All 39 thermocouples are placed 0.01 m below the tunnel ceiling. The study was conducted under the condition that the fire source is located at the closed end of the tunnel, 0.0 m away, with a fire source heat release rate of 4.3 kW and an ambient temperature of approximately 10℃. Under this condition, the corresponding full-scale tunnel fire source heat release rate is 2.72 MW.

[0055] like Figure 7 As shown, the temperature prediction results of the full-size tunnel corresponding to the scaled-down model are highly consistent with the temperature change trend recorded in the experiments, and the overall prediction error is controlled within a small range. Meanwhile, the inverted heat release rate of the fire source is 2.43 MW, which is extremely close to the actual heat release rate of 2.72 MW, with a relative error controlled within 11%. The above experimental results verify the accuracy and reliability of the multi-subnet BP neural network fusion model prediction method and the fire source heat release rate inversion method of this invention.

[0056] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting and inverting the fire source characteristics of tunnel fires based on a multi-subnet BP neural network, characterized in that, Includes the following steps: S1. Deploy temperature sensors and collect temperature monitoring data at various locations through the temperature sensors. The temperature monitoring data includes the spatial coordinates of each temperature sensor and the corresponding temperature value. Preprocess the acquired data. S2. Input the pre-processed spatial coordinates of each temperature sensor and the corresponding temperature value into the pre-trained multi-subnet BP neural network fusion model, and output the temperature distribution curve of the tunnel roof through the temperature monitoring data at a limited number of measuring points. S3. Determine the longitudinal and transverse coordinates of the fire source location based on the peak temperature position on the tunnel ceiling temperature distribution curve; S4. Substitute the location of the fire source in the tunnel and the corresponding predicted temperature value determined in the temperature distribution curve of the tunnel ceiling into the empirical correlation formula of the ceiling temperature rise, and use the numerical iterative solution method to back-calculate the heat release rate of the fire source.

2. The method according to claim 1, characterized in that, In S1, the temperature sensors are installed below the tunnel ceiling, with a series of temperature sensors installed along the longitudinal direction of the tunnel, and a series of temperature sensors installed every 8m along the transverse direction of the tunnel.

3. The method according to claim 1, characterized in that, The training process of the pre-trained multi-subnet BP neural network fusion model is as follows: Construct a sample set; The sample set includes: the spatial coordinates of multiple temperature sensors and their corresponding temperature values, as well as their corresponding tag values; the tags represent the ceiling temperature. The sample set is preprocessed; the preprocessing includes: smoothing the temperature sequence arranged according to spatial location, and standardizing the spatial location coordinates and the corresponding temperature values ​​respectively; The sample set is filtered to obtain key samples, and the key samples are included in the training set. The remaining data is divided into training set and test set according to a preset ratio. The filtering principle is to identify peak temperature points at different spatial locations from the existing data and use the sample index set corresponding to the peak temperature points as part of the training set. Construct a multi-subnetwork BP neural network fusion model composed of parallel fusion of multiple subnetworks; The training set is input into each sub-network for training. The loss function is used to evaluate the prediction error and update the sub-network parameters. The sample weights are periodically adjusted during the training process. After training, the fusion weights are assigned based on the prediction accuracy of each sub-network. The weighted sum of the prediction values ​​of each sub-network is used as the final temperature prediction value of the multi-sub-network BP neural network fusion model. The final temperature prediction is restored to the actual temperature value through destandardization. The destandardization formula is as follows: ； in, The actual temperature value after reduction, in °C; The standardized temperature prediction value output by the multi-subnet BP neural network fusion model, i.e., the final temperature prediction value. ; The standard deviation of the temperature values; This represents the average temperature value.

4. The method according to claim 3, characterized in that, The preprocessing includes: S21. The temperature sequence is smoothed using a Gaussian filter, with the following formula: ； in, For the original temperature sequence at position The value at that location, For smoothed position The value at that location, These are the Gaussian filter smoothing coefficients; S22. Spatial position coordinates and the corresponding temperature value Standardize them separately. The standardization formula is: ； ； in, This represents the mean value corresponding to the spatial location coordinates. , The number of samples. For sample number, For the first The location coordinates of each sample; The standard deviation of the spatial location coordinates. ; For standardized spatial coordinates, This is the standardized temperature value.

5. The method according to claim 3, characterized in that, The multi-subnetwork BP neural network fusion model consists of 7 subnetworks fused in parallel. Each subnetwork contains one input layer neuron, 16 neurons in the first hidden layer, 32 neurons in the second hidden layer, 16 neurons in the third hidden layer, and one output layer neuron. All hidden layers use the ReLU activation function, and the output layer uses a linear activation function. The forward propagation formula for a single subnetwork is: ； ； ； ； in, This is the weight matrix from the input layer to the first hidden layer. This is the weight matrix from the first hidden layer to the second hidden layer. This is the weight matrix from the second hidden layer to the third hidden layer. This is the weight matrix from the third hidden layer to the output layer. For each layer's bias vector, For activation function, These are the output values ​​of each hidden layer after activation by the ReLU activation function. These are the output values ​​of each hidden layer after linear transformation. For standardized spatial coordinates, It is the first The standardized temperature prediction value output by each sub-network. Number the sub-network. .

6. The method according to claim 5, characterized in that, The loss function for each sub-network during training is: ； in, For the first The training loss of each subnetwork For the number of training sets, For the first The standardized temperature value of each sample. For the first Subnetworks for the first Standardized temperature prediction values ​​for each sample. For sample number, Number the subnetwork.

7. The method according to claim 3, characterized in that, During training, the multi-subnetwork BP neural network fusion model uses the Adam optimizer to update the subnetwork parameters. The update formula for the Adam optimizer is as follows: ； ； ； ； ； in, , They are respectively The first and second momentum at time t. =0.9、 =0.999 is the momentum decay coefficient. loss function For parameters gradient, , This is the momentum value after deviation correction. These are the weight parameters of the subnetwork. For the first The parameter values ​​after the next iteration. For learning rate, =10 -8 To prevent the minimum value where the denominator is 0.

8. The method according to claim 7, characterized in that, During the training process, the sample weights of the multi-subnet BP neural network fusion model are updated after every 1000 training rounds, and samples with larger absolute prediction errors are assigned higher weights. The formula for updating sample weights is: ； in, For the first Temperature values ​​of each sample For the first In the subnetwork, the first The weights of each sample satisfy... , For the first Subnetworks for the first The predicted temperature value for each sample. For the number of training sets, , All are sample numbers.

9. The method according to claim 5, characterized in that, The formula for allocating the fusion weights of each subnetwork in the multi-subnetwork BP neural network fusion model is as follows: ； ； in, For the first The root mean square error of each subnetwork on the training set For the first Temperature values ​​of each sample For the first Subnetworks for the first The predicted temperature value for each sample. For the first The fusion weights of each subnetwork satisfy , The total number of sub-networks, , For sample number, This represents the size of the training set.

10. The method according to claim 3, characterized in that, The formula for the final temperature prediction value of the multi-subnet BP neural network fusion model is as follows: ； in, This represents the final temperature prediction value from the multi-subnet BP neural network fusion model. For the first The fusion weights of individual subnetworks For the first Temperature predictions after denormalization of the subnetwork The total number of sub-networks, , Number the subnetwork.

Images