Intelligent control method and system based on smoke prevention and exhaust system

Abstract

The application relates to an intelligent regulation and control method and system based on a smoke prevention and exhaust system, which comprises collecting real-time state parameters of a fire; inputting the state parameters into a smoke prediction model based on a mixed architecture of a Transformer and a convolutional neural network or a gated recurrent unit, and outputting along-the-way temperature, visibility and smoke layer height distribution; based on the prediction result, a multi-objective optimization function is constructed by comprehensively considering the upstream smoke backflow length, the evacuation area visibility, the smoke layer height and the system energy consumption, the optimal longitudinal wind speed and the exhaust wind speed are solved, the fan rotating speed and the exhaust valve opening degree are adjusted in real time; the real-time monitoring data and the prediction value are compared, and when the error exceeds the threshold value, an online learning mechanism is triggered to finely adjust the parameters of the model. The application realizes the transformation of the smoke prevention and exhaust system from passive response to active prediction, and effectively improves the smoke exhaust efficiency and the personnel evacuation safety in a fire scene.

Description

Technical Field

[0001] This application relates to the field of fire prevention and smoke exhaust technology, and in particular to an intelligent control method and system based on a smoke prevention and exhaust system. Background Technology

[0002] With the rapid development of infrastructure construction in my country, fire safety issues have become increasingly prominent. Tunnel structures and other building structures are characterized by being semi-enclosed and longitudinally narrow, making it difficult to quickly expel the high-temperature and toxic fumes generated during a fire, seriously threatening the safety of personnel evacuation and the stability of the building structure.

[0003] Currently, existing smoke control systems mainly employ a linkage control method based on preset logic. When the fire alarm system detects that the smoke concentration or temperature has reached a preset threshold, it starts the fans and opens the smoke exhaust valves according to a pre-set control scheme. For example, in longitudinal ventilation systems, they typically operate at a constant critical air velocity; in tunnels with cross passages, the angle between the cross passage and the main tunnel is often a fixed 90°, and the smoke exhaust air velocity also uses a preset setting. In recent years, some studies have attempted to combine numerical simulation with empirical formulas, using computational fluid dynamics software to simulate specific fire scenarios, providing a reference for the design of smoke control systems.

[0004] Regarding the aforementioned technologies, existing smoke control and exhaust systems employ static and fixed control strategies, which cannot be dynamically adjusted based on real-time status parameters such as the location of the fire source, the power of the fire source, and the longitudinal wind speed. This results in low smoke exhaust efficiency and makes it difficult to cope with complex and ever-changing real-world scenarios.

[0005] Based on this, this application provides an intelligent control method and system based on a smoke control system. Summary of the Invention

[0006] To address the problem that existing smoke control systems have static and fixed control strategies that cannot be dynamically adjusted based on real-time parameters such as fire source location, fire source power, and longitudinal wind speed, resulting in low smoke extraction efficiency and difficulty in coping with complex and ever-changing real-world scenarios, this application provides an intelligent control method and system based on smoke control systems.

[0007] Firstly, this application provides an intelligent control method based on a smoke control system, which adopts the following technical solution: including: Collect real-time status parameters of the fire-occurring area, including at least the heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed, and time of occurrence. Based on a preset smoke prediction model based on the Transformer architecture, the smoke prediction model adopts a hybrid architecture of Transformer and convolutional neural network or a hybrid architecture of Transformer and gated recurrent unit. Through the smoke prediction model and the real-time state parameters, the smoke parameters along the building space are obtained. The smoke parameters include at least the temperature distribution, visibility distribution and smoke layer height distribution at a set height, where the set height is the critical height for personnel evacuation. Based on the flue gas parameters, a multi-objective optimization function is constructed that comprehensively considers the upstream flue gas counterflow length, visibility in the personnel evacuation area, flue gas layer height, and energy consumption of the smoke exhaust system. The multi-objective optimization function is solved within the preset feasible domain of control parameters to generate the optimal control parameters that minimize the overall cost. The optimal control parameters include at least the optimized longitudinal wind speed and smoke exhaust wind speed. The fan speed and smoke exhaust valve opening are adjusted in real time according to the optimized longitudinal wind speed and smoke exhaust wind speed. The real-time state parameters are compared with the predicted flue gas parameters to calculate the prediction error. When the prediction error exceeds a preset threshold, an online learning mechanism is triggered to fine-tune the parameters of the Transformer prediction model using an incremental learning algorithm, and the model is adaptively updated.

[0008] Preferably, before obtaining the smoke parameters along the building space using the preset Transformer-based smoke prediction model (which employs a hybrid architecture of Transformer and convolutional neural networks or a hybrid architecture of Transformer and gated recurrent units) and the real-time state parameters, the method further includes: Using computational fluid dynamics numerical simulation software, a fire model of a highway tunnel or building space containing cross passages is established, and various fire condition combinations are set. The fire condition combinations include parameter values ​​of different heat release rates of fire sources, different longitudinal wind speeds, different smoke exhaust wind speeds, and different fire occurrence times. The temperature distribution, visibility distribution, and smoke layer height distribution data at a set height along the route under each condition are obtained through numerical simulation to form the original dataset. The original dataset is structured by organizing the heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed and fire occurrence time at each time point into input feature vectors, and organizing the corresponding temperature distribution, visibility distribution and smoke layer height distribution into output labels; all input feature vectors and output labels are normalized and mapped to a uniform numerical range. A sliding time window method is used to extract temporal features and construct training samples with time dependencies. The training samples are divided into training set and test set according to the fire conditions to ensure that the test set includes complete conditions that were not included in the training. A hybrid neural network model based on Transformer is constructed, comprising an input layer, a feature extraction layer, and an output layer. The input layer receives preprocessed temporal feature vectors. The feature extraction layer includes a Transformer encoder module and a local feature extraction module. The Transformer encoder module consists of multiple stacked encoder blocks, each containing a multi-head self-attention layer and a feedforward neural network layer, equipped with residual connections and layer normalization to capture long-range dependencies in the input sequence. The local feature extraction module uses a one-dimensional convolutional neural network or gated recurrent units to extract local spatial features from the features output by the Transformer encoder or to further model temporal dynamic characteristics. The output layer contains multiple fully connected layers that map the extracted features to a dimension equal to the number of measurement points, outputting the temperature distribution, visibility distribution, and smoke layer height distribution at a set height along the path. The mean squared error was used as the loss function, and the Adam optimizer was used to train the hybrid neural network model. The initial learning rate, batch size and maximum number of training epochs were set, and a dynamic learning rate adjustment strategy and an early stopping mechanism were introduced. When the loss on the validation set no longer decreased for several consecutive epochs, the training was terminated early, and the model parameters with the best performance on the validation set were saved. The trained hybrid neural network model was validated using working conditions not included in the training in the test set. The model accuracy was evaluated by comparing the error between the model's predicted values ​​and the simulated data corresponding to the output labels in the training samples, ensuring that the model can accurately predict the spatiotemporal distribution of smoke parameters under different fire conditions.

[0009] Preferably, after constructing the Transformer-based hybrid neural network model, which includes an input layer, a feature extraction layer, and an output layer, the method further includes: The Transformer encoder module is set to consist of two stacked encoder blocks. The number of heads in the multi-head self-attention mechanism in each encoder block is set to 8, and the hidden layer dimension of the feedforward neural network layer is set to 128. When the local feature extraction module uses a one-dimensional convolutional neural network, the number of convolutional layers is set to 2, the number of filters in the first convolutional layer is 64, the number of filters in the second convolutional layer is 128, and the kernel size is set to 2. When the local feature extraction module uses a gated loop unit, the number of gated loop unit layers is set to 2. The first layer is a gated loop unit that returns the sequence and the number of hidden units is set to 128. The second layer is a gated loop unit that only returns the output of the last time step and the number of hidden units is set to 256. The number of fully connected layers in the output layer is set to 2. The hidden layer dimension of the first fully connected layer is set to 512, and the hidden layer dimension of the second fully connected layer is set to 256. A Dropout mechanism is introduced after each fully connected layer to prevent overfitting. Finally, the dimension of the output layer is set to be the same as the number of measurement points, which is used to output the temperature distribution, visibility distribution and flue gas layer height distribution at a set height along the path.

[0010] Preferably, the smoke prediction model based on a preset Transformer architecture adopts a hybrid architecture of Transformer and convolutional neural network or a hybrid architecture of Transformer and gated recurrent unit. Through the smoke prediction model and the real-time state parameters, smoke parameters along the building space are obtained, including: The real-time state parameters collected at the current moment are concatenated with the historical state parameters collected at previous moments to form an input sequence with time dependence. Each time step in the input sequence contains four feature dimensions: heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed, and fire occurrence time. The input sequence is fed into the Transformer encoder module of the trained flue gas prediction model. The dynamic correlation weights between different time steps in the sequence are calculated through a multi-head self-attention mechanism, so that the flue gas prediction model focuses on the historical moment features that have an important impact on the current prediction. After being processed by stacking multiple encoder blocks, the output is a temporal feature representation containing long-range dependencies. The temporal feature representation is input into the local feature extraction module. When a hybrid architecture of Transformer and convolutional neural network is adopted, the temporal feature is convolved through a one-dimensional convolutional layer to extract local change features along the time dimension. When a hybrid architecture of Transformer and gated recurrent unit is adopted, the temporal feature is recursively processed through a gated recurrent unit layer to further model the dynamic characteristics of flue gas parameters evolving over time. The local variation features or dynamic characteristics output by the local feature extraction module are flattened and then input into the fully connected output layer. Through the nonlinear mapping of the multi-layer fully connected network, the high-dimensional features are converted into an output dimension that matches the number of measurement points along the building space. The fully connected output layer simultaneously outputs three parallel prediction result tensors, corresponding to the temperature distribution, visibility distribution, and smoke layer height distribution at a set height along the building space, respectively; each prediction result tensor contains parameter values ​​at each measurement point along the path.

[0011] Preferably, based on the flue gas parameters, a multi-objective optimization function is constructed that comprehensively considers the upstream flue gas counterflow length, visibility in the personnel evacuation area, flue gas layer height, and energy consumption of the smoke exhaust system. The multi-objective optimization function is solved within the preset feasible region of control parameters to generate optimal control parameters that minimize the overall cost, including: Based on temperature distribution or smoke layer height distribution, identify the position of the smoke front upstream of the fire source and calculate the upstream smoke backflow length between the fire source and the smoke front; extract visibility values ​​in key evacuation areas from the visibility distribution, which include at least the evacuation path upstream of the fire source and the cross passage entrance area; extract the smoke layer height at each location along the path from the smoke layer height distribution and identify dangerous areas where the smoke layer height is lower than human height; calculate the instantaneous energy consumption of the smoke exhaust system based on the current fan speed and smoke exhaust valve opening. A cost function is constructed that comprehensively considers the above-mentioned multiple optimization objectives. The cost function is expressed as the weighted sum of the deviations between each optimization objective and its expected value, and includes at least the following four cost components: The first cost component is proportional to the upstream flue gas backflow length, representing the degree to which the spread of flue gas upstream of the fire source hinders personnel evacuation; the second cost component is proportional to the negative deviation between the visibility in the key personnel evacuation area and the preset safe visibility threshold, representing the impact of insufficient visibility on the personnel evacuation speed; the third cost component is proportional to the negative deviation between the smoke layer height and the preset safe height threshold, representing the degree to which the descent of the smoke layer compresses the effective evacuation space for personnel; and the fourth cost component is proportional to the instantaneous energy consumption of the smoke exhaust system, representing the economic efficiency of the system operation. Based on the physical constraints and safety specifications of the smoke control system, the feasible domain of the control parameters is set. The feasible domain includes at least the upper and lower limits of the adjustable range of the longitudinal wind speed, the upper and lower limits of the adjustable range of the smoke exhaust wind speed, and the adjustment rate limits of the fan speed and the opening of the smoke exhaust valve. Within the feasible region of the control parameters, the optimal control parameters that minimize the cost function are solved by an optimization algorithm. The optimization algorithm adopts a heuristic search algorithm or a gradient descent algorithm, and obtains the longitudinal wind speed and smoke exhaust wind speed that minimize the overall cost through iterative optimization. The optimal longitudinal wind speed and optimal smoke exhaust wind speed obtained from the solution are used as control command parameters and output to the actuator control module for subsequent adjustment of fan speed and smoke exhaust valve opening.

[0012] Preferably, the step of using an optimization algorithm to solve for the optimal control parameters that minimize the cost function within the feasible region of the control parameters includes: At the start of the current control cycle, the optimal control parameters of the previous cycle are used as the initial values ​​for the optimization solution of the current cycle, and the real-time state parameters and the predicted flue gas parameters at the current moment are obtained. Based on the flue gas prediction model, a flue gas parameter evolution model is constructed in the future prediction time domain. The flue gas parameter evolution model takes the current moment as the starting point and the initial value as the input to predict the dynamic changes of temperature distribution, visibility distribution and flue gas layer height distribution at a set height along the path in multiple future time steps. Within the feasible domain of the control parameters, a set of candidate control parameter sequences is generated by random sampling or grid search. Each candidate control parameter sequence contains the value sequences of longitudinal wind speed and smoke exhaust wind speed corresponding to each time step in the future prediction time domain. For each candidate control parameter sequence, the candidate control parameter sequence is input into the flue gas parameter evolution model step by step to obtain the flue gas parameter prediction results for each future time step; according to the multi-objective optimization function, the instantaneous cost corresponding to each future time step is calculated respectively; the instantaneous costs of each future time step are weighted and summed according to the time decay factor to obtain the cumulative comprehensive cost corresponding to the candidate control parameter sequence. The time decay factor is used to balance the weight of the near-term cost and the long-term cost, and the instantaneous cost closer to the current time has a larger weight; Compare the cumulative integrated costs of all candidate control parameter sequences, and select the candidate control parameter sequence that minimizes the cumulative integrated cost as the optimal control parameter for the current period; After the current control cycle is completed, the prediction time domain is rolled forward by one time step, and the optimal control parameters for the next cycle are repeatedly selected to achieve rolling optimization of model predictive control.

[0013] Preferably, the step of comparing the real-time state parameters with the predicted flue gas parameters to calculate the prediction error; when the prediction error exceeds a preset threshold, triggering an online learning mechanism to fine-tune the parameters of the Transformer prediction model using an incremental learning algorithm and adaptively updating the model includes: During system operation, the real-time status parameters and the actual flue gas parameter monitoring values ​​at the corresponding time are continuously collected, and the collected data is stored in a circular buffer; the actual flue gas parameter monitoring values ​​include at least the temperature value, visibility value, and flue gas layer height value measured by sensors at a set height along the path; According to the preset evaluation cycle, historical monitoring data within the preset time window before the current moment is extracted from the circular buffer and compared with the predicted value output by the flue gas prediction model at the corresponding moment. The comprehensive prediction error index is calculated. The comprehensive prediction error index adopts the root mean square error or the mean absolute percentage error, and is calculated independently or by weighted fusion for the three parameters of temperature distribution, visibility distribution and flue gas layer height distribution, respectively. The comprehensive prediction error index is compared with a preset error threshold; when the prediction error of any parameter exceeds the corresponding preset threshold, it is determined that the prediction accuracy of the current model has decreased, and the online learning mechanism is triggered. When the online learning mechanism is triggered, the actual flue gas parameter monitoring values ​​within a preset time window before the trigger time are extracted from the circular buffer to construct an online learning sample set. Each sample in the online learning sample set contains an input feature vector and a corresponding output label. The input feature vector is the real-time state parameter collected at a historical time, and the output label is the actual flue gas parameter monitoring value at the corresponding time. Freeze some network layer parameters in the Transformer prediction model, and fine-tune the parameters of only a few fully connected layers or specific attention layers near the output layer; use mini-batch gradient descent, with the online learning sample set as training data, to perform incremental training for a limited number of rounds on the basis of the original model parameters, and update the network layer parameters other than the frozen layers. The fine-tuned model is quickly validated using the validation samples reserved in the online learning sample set, or using the latest monitoring data collected after the update in the circular buffer, and the prediction error of the updated model is calculated. If the prediction error of the updated model is lower than that before the update and is below a preset threshold, the model update is confirmed to be effective. The updated model parameters that have been verified to be valid will be saved as the current active version for subsequent flue gas parameter prediction. If the performance of the updated model does not improve or deteriorates, it will be automatically rolled back to the previous model version, and the environmental conditions that triggered the rollback will be recorded.

[0014] Secondly, this application discloses an intelligent control device based on a smoke control system, which adopts the following technical solution, including: The data acquisition module is used to collect real-time status parameters of the fire-occurring area. The real-time status parameters include at least the heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed, and the time of occurrence. The smoke prediction module is used to obtain smoke parameters along the building space based on a preset smoke prediction model based on the Transformer architecture. The smoke prediction model adopts a hybrid architecture of Transformer and convolutional neural network or a hybrid architecture of Transformer and gated recurrent unit. Through the smoke prediction model and the real-time state parameters, the smoke parameters include at least the temperature distribution, visibility distribution and smoke layer height distribution at a set height, where the set height is the critical height for personnel evacuation. The parameter control module is used to construct a multi-objective optimization function based on the flue gas parameters, which comprehensively considers the upstream flue gas counterflow length, visibility in the personnel evacuation area, flue gas layer height, and energy consumption of the smoke exhaust system. The module solves the multi-objective optimization function within the preset feasible domain of control parameters to generate the optimal control parameters that minimize the overall cost. The optimal control parameters include at least the optimized longitudinal wind speed and smoke exhaust wind speed. The module adjusts the fan speed and smoke exhaust valve opening in real time based on the optimized longitudinal wind speed and smoke exhaust wind speed. The online update module is used to compare the real-time state parameters with the predicted flue gas parameters and calculate the prediction error. When the prediction error exceeds a preset threshold, the online learning mechanism is triggered to fine-tune the parameters of the Transformer prediction model using an incremental learning algorithm and perform adaptive model updates.

[0015] Thirdly, this application also provides a control device, the device comprising: It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described above for the intelligent control method based on the smoke control system.

[0016] Fourthly, this application also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above regarding the intelligent control method based on a smoke control system.

[0017] In summary, this application collects key state parameters of a fire scenario in real time, including the heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed, and the time of fire occurrence. These parameters are then input into a pre-built smoke prediction model based on a Transformer architecture. This model employs a hybrid architecture of Transformer and convolutional neural networks or a hybrid architecture of Transformer and gated recurrent units. It captures long-range dependencies in the time series through a multi-head self-attention mechanism and utilizes a local feature extraction module to model the dynamic characteristics of the time series, quickly outputting the temperature distribution and visibility at a set height along the building space. The system analyzes the distribution of smoke and the height of the smoke layer. Based on predicted smoke parameters, a multi-objective optimization function is constructed, comprehensively considering the upstream smoke counterflow length, visibility in the evacuation area, smoke layer height, and energy consumption of the smoke exhaust system. This function is solved within the feasible region of preset control parameters to generate the optimal longitudinal wind speed and smoke exhaust wind speed that minimize the overall cost. The system adjusts the fan speed and smoke exhaust valve opening in real time according to the optimized control parameters. Simultaneously, the system compares real-time monitoring data with model predictions, calculates the prediction error, and triggers an online learning mechanism when the error exceeds a preset threshold. An incremental learning algorithm is used to fine-tune the parameters of the prediction model, achieving adaptive model updates. This application, through the perception, prediction, execution, and updating processes of the smoke control system, realizes the transformation of the smoke control system from passive response to active prediction, significantly improving smoke exhaust efficiency and personnel evacuation safety in complex fire scenarios, while reducing system energy consumption. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an intelligent control method based on a smoke control system.

[0019] Figure 2 This is a structural block diagram of an intelligent control device based on a smoke control system. Detailed Implementation

[0020] The following combination Figures 1-2 This application will be described in further detail.

[0021] Reference Figure 1 The embodiments of this application include at least steps S10 to S40.

[0022] S10: Collect real-time status parameters of the fire-affected area. The real-time status parameters include at least the heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed, and the time of occurrence.

[0023] S20 is based on a preset smoke prediction model based on the Transformer architecture. The smoke prediction model adopts a hybrid architecture of Transformer and convolutional neural network or a hybrid architecture of Transformer and gated recurrent unit. Through the smoke prediction model and real-time state parameters, the smoke parameters along the building space are obtained. The smoke parameters include at least the temperature distribution, visibility distribution and smoke layer height distribution at a set height. The set height is the critical height for personnel evacuation.

[0024] S30, based on flue gas parameters, constructs a multi-objective optimization function that comprehensively considers the upstream flue gas counterflow length, visibility in the personnel evacuation area, flue gas layer height, and energy consumption of the smoke exhaust system. Solve the multi-objective optimization function within the feasible region of the preset control parameters to generate the optimal control parameters that minimize the overall cost. The optimal control parameters include at least the optimized longitudinal wind speed and smoke exhaust wind speed. Adjust the fan speed and smoke exhaust valve opening in real time according to the optimized longitudinal wind speed and smoke exhaust wind speed.

[0025] S40 compares the real-time state parameters with the predicted flue gas parameters and calculates the prediction error. When the prediction error exceeds a preset threshold, an online learning mechanism is triggered to fine-tune the parameters of the Transformer prediction model using an incremental learning algorithm and perform adaptive model updates.

[0026] Specifically, real-time fire status parameters are collected and input into a Transformer-based hybrid architecture prediction model, which quickly outputs the temperature, visibility, and smoke layer height distribution along the path. Then, based on the prediction results, a multi-objective optimization function is constructed to solve for the optimal control parameters that comprehensively optimize upstream smoke backflow, visibility in evacuation areas, smoke layer height, and system energy consumption, adjusting fan speed and smoke exhaust valve opening in real time. Simultaneously, real-time monitoring data is compared with predicted values; when the error exceeds a threshold, an online learning mechanism is triggered to fine-tune the model parameters. This achieves a shift from passive response to proactive prediction in smoke control systems, effectively improving smoke extraction efficiency and personnel evacuation safety in fire scenarios.

[0027] In some embodiments, considering the construction of the smoke prediction model, the corresponding processing steps are as follows: Using computational fluid dynamics numerical simulation software, a fire model of a highway tunnel or building space containing cross passages is established, setting multiple fire condition combinations. These combinations include parameter values ​​for different fire source heat release rates, different longitudinal wind speeds, different smoke exhaust wind speeds, and different fire occurrence times. Temperature distribution, visibility distribution, and smoke layer height distribution data at a set height along the path under each condition are obtained through numerical simulation to form an original dataset. The original dataset is then structured, with the fire source heat release rate, longitudinal wind speed, smoke exhaust wind speed, and fire occurrence time at each time point organized as input feature vectors, and the corresponding temperature distribution, visibility distribution, and smoke layer height distribution organized as output labels. All input feature vectors and output labels are normalized and mapped to a unified numerical range. A sliding time window method is used to extract temporal features, constructing training samples with time dependencies. The training samples are divided into a training set and a test set according to the fire conditions, ensuring that the test set includes complete conditions not involved in the training. A hybrid neural network model based on Transformer is constructed, comprising an input layer, a feature extraction layer, and an output layer. The input layer receives preprocessed temporal feature vectors. The feature extraction layer includes a Transformer encoder module and a local feature extraction module. The Transformer encoder module consists of multiple stacked encoder blocks, each containing a multi-head self-attention layer and a feedforward neural network layer, equipped with residual connections and layer normalization to capture long-range dependencies in the input sequence. The local feature extraction module uses a one-dimensional convolutional neural network or gated recurrent units to extract local spatial features from the features output by the Transformer encoder or to further model temporal dynamic characteristics. The output layer contains multiple fully connected layers that map the extracted features to a dimension equal to the number of measurement points, outputting the temperature distribution, visibility distribution, and smoke layer height distribution at a set height along the path. The mean squared error (MSE) loss function is used as the training function, and the Adam optimizer is employed to train the hybrid neural network model. Initial learning rate, batch size, and maximum training epochs are set, and a dynamic learning rate adjustment strategy and an early stopping mechanism are introduced. Training is terminated early when the validation set loss no longer decreases for several consecutive epochs, and the model parameters with the best performance on the validation set are saved. The mathematical expression for the MSE loss function is: ; Where MSE is the mean squared error and n is the sample size. Let be the numerical simulation value of the i-th sample. This represents the model's predicted value. This loss function imposes a larger penalty on larger errors, making the model more focused on reducing large errors, which is particularly important for fire smoke prediction.

[0028] The trained hybrid neural network model was validated using untrained fire conditions in the test set. Model accuracy was assessed by comparing the error between the model's predicted values ​​and the simulated data corresponding to the output labels in the training samples, ensuring that the model can accurately predict the spatiotemporal distribution of smoke parameters under different fire conditions. Mean absolute error was also used as an auxiliary evaluation metric. ; MAE is the mean absolute error, which assigns equal weight to each error and provides an intuitive measure of prediction error.

[0029] Specifically, firstly, a fire model of a highway tunnel or building space including a cross passage was established using computational fluid dynamics numerical simulation software. Multiple combinations of operating conditions were set, encompassing different heat release rates from the fire source, longitudinal wind speeds, smoke exhaust wind speeds, and fire occurrence times. Temperature, visibility, and smoke layer height distribution data at a set height along the route were obtained through numerical simulation under each operating condition, forming the raw dataset. Next, the raw data underwent structured processing, with state parameters organized into input feature vectors and smoke parameter distributions organized into output labels. Normalization and sliding time window temporal feature extraction were then performed, dividing the dataset into training and test sets according to the operating conditions. Then, a Transformer-based hybrid neural network model was constructed. Its input layer receives temporal feature vectors, the feature extraction layer consists of a Transformer encoder module and a one-dimensional convolutional neural network or gated recurrent unit local feature extraction module, and the output layer is mapped to the dimension of the number of measurement points through a fully connected layer, outputting the smoke parameter distribution. Finally, the model was trained using the mean squared error as the loss function and the Adam optimizer, with dynamic adjustment of the learning rate and an early stopping mechanism introduced. The model accuracy was verified using the test set. This embodiment constructs a high-precision flue gas prediction model with strong generalization ability through offline training driven by simulation data, providing a reliable decision-making basis for subsequent real-time control.

[0030] Furthermore, the Transformer encoder module is configured to consist of two stacked encoder blocks, with each encoder block containing 8 heads for the multi-head self-attention mechanism and a hidden layer dimension of 128 for the feedforward neural network. When the local feature extraction module uses a one-dimensional convolutional neural network, it is configured with two convolutional layers, with 64 filters in the first layer and 128 filters in the second layer, and both kernel sizes are set to 2. When the local feature extraction module uses a gated recurrent unit, it is configured with two gated recurrent unit layers, with the first layer being the return layer. The sequence is gated recurrent unit with 128 hidden units. The second layer is a gated recurrent unit that only returns the output of the last time step with 256 hidden units. The number of fully connected layers in the output layer is set to 2. The hidden layer dimension of the first fully connected layer is set to 512, and the hidden layer dimension of the second fully connected layer is set to 256. A Dropout mechanism is introduced after each fully connected layer to prevent overfitting. Finally, the dimension of the output layer is set to be the same as the number of measurement points, which is used to output the temperature distribution, visibility distribution and flue gas layer height distribution at a set height along the path.

[0031] In some embodiments, step S20 specifically includes the following steps: concatenating the real-time state parameters collected at the current moment with the historical state parameters collected at historical moments to form an input sequence with time dependencies, wherein each time step in the input sequence contains four feature dimensions: heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed, and fire occurrence time; inputting the input sequence into the Transformer encoder module in the trained smoke prediction model, calculating the dynamic correlation weights between different time steps in the sequence through a multi-head self-attention mechanism, so that the smoke prediction model focuses on the historical moment features that have an important impact on the current prediction; and outputting a temporal feature representation containing long-range dependencies after being processed by stacking multiple encoder blocks. The temporal feature representation is input into the local feature extraction module. When a hybrid architecture of Transformer and convolutional neural network is adopted, the temporal feature is convolved through a one-dimensional convolutional layer to extract local variation features along the time dimension. When a hybrid architecture of Transformer and gated recurrent unit is adopted, the temporal feature is recursively processed through the gated recurrent unit layer to further model the dynamic characteristics of flue gas parameters evolving over time. The local variation features or dynamic characteristics output by the local feature extraction module are flattened and then input into the fully connected output layer. Through the nonlinear mapping of the multi-layer fully connected network, the high-dimensional features are converted into an output dimension that matches the number of measuring points along the building space. The fully connected output layer simultaneously outputs three parallel prediction result tensors, corresponding to the temperature distribution, visibility distribution, and flue gas layer height distribution at a set height along the building space, respectively. Each prediction result tensor contains the parameter values ​​at each measuring point along the path.

[0032] The core of the flue gas prediction model is the multi-head self-attention mechanism in the Transformer architecture, and its calculation process can be expressed as follows: ; The calculation for each attention head is as follows: ; Where Q is the query matrix, K is the key matrix, and V is the value matrix. , , These are the weight matrices for the query, key, and value corresponding to the i-th attention head, respectively. The output weight matrix is ​​given by h, where h is the number of attention heads. The dimension of the key vector, divided by This is to prevent the softmax function from entering the saturation region due to excessively large dot product values. This mechanism enables the model to capture the dynamic correlation between different time steps in the input sequence, laying the foundation for accurately predicting the evolution of fire smoke.

[0033] Specifically, the real-time state parameters of the current moment and historical moments are concatenated into an input sequence with time dependencies. Each time step includes four feature dimensions: heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed, and fire occurrence time. The input sequence is fed into the Transformer encoder module, which calculates the dynamic correlation weights between different time steps through a multi-head self-attention mechanism, enabling the model to focus on historical features that have a significant impact on the current prediction. After being processed by stacking multiple encoder blocks, the output is a temporal feature representation containing long-range dependencies. Subsequently, the temporal feature representation is fed into the local feature extraction module. When using a one-dimensional convolutional neural network, local change features along the time dimension are extracted through convolution operations. When using a gated recurrent unit, the dynamic evolution characteristics of the smoke parameters are further modeled through recursive processing. The features output by the local feature extraction module are flattened and fed into the fully connected output layer. Through the nonlinear mapping of multiple fully connected networks, the output dimension is transformed to match the number of measurement points along the path. Finally, three prediction result tensors—temperature distribution, visibility distribution, and smoke layer height distribution—are output in parallel. It achieves an efficient mapping from input state parameters to the spatiotemporal distribution of multidimensional flue gas parameters, providing an accurate predictive data foundation for subsequent optimization decisions.

[0034] In some embodiments, step S30 specifically includes the following steps: based on temperature distribution or smoke layer height distribution, identify the position of the upstream smoke front of the fire source, and calculate the upstream smoke backflow length between the fire source and the smoke front; extract visibility values ​​in key personnel evacuation areas from the visibility distribution, the key personnel evacuation areas including at least the personnel evacuation path upstream of the fire source and the cross passage entrance area; extract the smoke layer height at each location along the path from the smoke layer height distribution, and identify the dangerous area range where the smoke layer height is lower than the height of human bodies; calculate the instantaneous energy consumption of the smoke exhaust system based on the current fan speed and smoke exhaust valve opening; The dimensionless model for the flue gas counter-current length is as follows: ; in, Let L be the dimensionless flue gas counter-current length. Let g be the tunnel height, g be the acceleration due to gravity, and Q be the rate of heat release from the fire source. air density, For flue gas temperature, Where is the specific heat capacity of air at constant pressure, and V is the longitudinal ventilation velocity. This represents the cross-sectional area of ​​the tunnel.

[0035] The upstream flue gas counterflow length can be quantitatively characterized using a dimensionless model, as shown in the following formula: ; in, To correct the Richardson number, The rate of heat release is dimensionless. The longitudinal wind speed is dimensionless. This model reflects the nonlinear relationship between the flue gas counterflow length, the fire source power, and the longitudinal wind speed.

[0036] The Li model can be used to describe the highest temperature rise below the ceiling: in, Let Q be the maximum temperature rise below the ceiling, Q be the heat release rate of the fire source, V be the longitudinal wind speed, and b be the radius of the fire source. The height of the tunnel. The longitudinal wind speed is dimensionless. The above model is used to assess the thermal impact of fire on the tunnel structure.

[0037] A cost function is constructed that comprehensively considers the above-mentioned multiple optimization objectives. The cost function is expressed as the weighted sum of the deviations between each optimization objective and its expected value, and includes at least the following four cost components: The first cost component is proportional to the upstream flue gas backflow length, representing the degree to which the spread of flue gas upstream of the fire source hinders personnel evacuation; the second cost component is proportional to the negative deviation between the visibility in the key personnel evacuation area and the preset safe visibility threshold, representing the impact of insufficient visibility on the personnel evacuation speed; the third cost component is proportional to the negative deviation between the smoke layer height and the preset safe height threshold, representing the degree to which the descent of the smoke layer compresses the effective evacuation space for personnel; and the fourth cost component is proportional to the instantaneous energy consumption of the smoke exhaust system, representing the economic efficiency of the system operation. Based on the physical constraints and safety specifications of the smoke control system, a feasible region for the control parameters is defined. This feasible region includes at least the upper and lower limits of the adjustable range of the longitudinal wind speed and the smoke exhaust wind speed, as well as the adjustment rate limits for the fan speed and the smoke exhaust valve opening. Within this feasible region, an optimization algorithm is used to find the optimal control parameters that minimize the cost function. The optimization algorithm employs a heuristic search algorithm or a gradient descent algorithm, iteratively optimizing to obtain the longitudinal wind speed and smoke exhaust wind speed that minimize the overall cost. The optimal longitudinal wind speed and optimal smoke exhaust wind speed obtained are then used as control command parameters and output to the actuator control module for subsequent adjustment of the fan speed and smoke exhaust valve opening.

[0038] Specifically, the system identifies the upstream smoke front position based on temperature distribution or smoke layer height distribution, and calculates the upstream smoke backflow length. Visibility values ​​for key evacuation areas (including upstream evacuation paths and cross passage entrances) are extracted from the visibility distribution. Dangerous areas below human height are identified from the smoke layer height distribution. Instantaneous energy consumption of the smoke extraction system is calculated based on the current fan speed and smoke exhaust valve opening. A multi-objective optimization function with four cost components is then constructed: the first component is proportional to the upstream smoke backflow length; the second component is proportional to the negative deviation of visibility in the evacuation area below the safety threshold; the third component is proportional to the negative deviation of smoke layer height below the safety threshold; and the fourth component is proportional to system energy consumption. After setting the adjustment range and rate limit for longitudinal wind speed and smoke exhaust wind speed, a heuristic search or gradient descent algorithm is used to solve for the optimal longitudinal wind speed and smoke exhaust wind speed that minimize the overall cost, and the results are output to the actuator for control. This achieves a quantitative decision mapping from prediction results to optimal control parameters.

[0039] In some embodiments, considering the specific process of minimizing the cost function, the corresponding processing steps are as follows: At the beginning of the current control cycle, the optimal control parameters of the previous cycle are used as the initial values ​​for the optimization solution of the current cycle, and the real-time state parameters and predicted flue gas parameters at the current moment are obtained; Based on the flue gas prediction model, a flue gas parameter evolution model in the future prediction time domain is constructed. The flue gas parameter evolution model takes the current moment as the starting point and the initial values ​​as input to predict the dynamic changes of temperature distribution, visibility distribution and flue gas layer height distribution at a set height along the path in multiple future time steps; Within the feasible domain of the control parameters, a set of candidate control parameter sequences is generated by random sampling or grid search. Each candidate control parameter sequence contains the value sequences of longitudinal wind speed and exhaust wind speed corresponding to each time step in the future prediction time domain; For each candidate control parameter sequence, the sequence is input into the flue gas parameter evolution model step by step to obtain the predicted flue gas parameters for each future time step. Based on the multi-objective optimization function, the instantaneous cost for each future time step is calculated. The instantaneous costs for each future time step are weighted and summed according to a time decay factor to obtain the cumulative comprehensive cost corresponding to the candidate control parameter sequence. The time decay factor is used to balance the weights of near-term and long-term costs, with the weight of instantaneous costs closer to the current time being larger. The cumulative comprehensive costs of all candidate control parameter sequences are compared, and the candidate control parameter sequence that minimizes the cumulative comprehensive cost is selected as the optimal control parameter for the current period. After the current control period is completed, the prediction time domain is rolled forward by one time step, and the optimal control parameter for the next period is repeatedly selected, achieving rolling optimization of the model predictive control.

[0040] Specifically, the optimal control parameters from the previous cycle are used as the initial values ​​for the current cycle. An evolutionary model is constructed based on the flue gas prediction model to predict the dynamic changes of flue gas parameters over multiple future time steps. Subsequently, candidate control parameter sequences are generated within the feasible domain of the control parameters. Each sequence contains the values ​​of longitudinal wind speed and exhaust wind speed for each future time step. For each candidate sequence, it is input into the evolutionary model step by step to obtain the predicted results of future flue gas parameters. The instantaneous cost corresponding to each time step is calculated, and then weighted and summed according to the time decay factor to obtain the cumulative comprehensive cost, where the cost closer to the current moment has a larger weight. The cumulative comprehensive costs of all candidate sequences are compared, and the sequence that minimizes the cost is selected as the optimal control parameter for the current cycle. After the current cycle is completed, the prediction time domain is rolled forward by one time step, and the above process is repeated to achieve rolling optimization of model predictive control. Thus, through forward-looking multi-step optimization and rolling updates, dynamic adaptive adjustment of the control strategy is achieved.

[0041] In some embodiments, step S40 specifically includes the following steps: During system operation, continuously collect real-time status parameters and actual flue gas parameter monitoring values ​​at corresponding times, and store the collected data in a circular buffer; the actual flue gas parameter monitoring values ​​include at least the temperature value, visibility value, and flue gas layer height value measured by sensors at a set height along the path; according to a preset evaluation cycle, extract historical monitoring data within a preset time window before the current time from the circular buffer, compare it with the predicted value output by the flue gas prediction model at the corresponding time, calculate the comprehensive prediction error index, the comprehensive prediction error index adopts root mean square error or mean absolute percentage error, and is calculated independently or by weighted fusion for the three parameters of temperature distribution, visibility distribution, and flue gas layer height distribution respectively; compare the comprehensive prediction error index with a preset error threshold; when the prediction error of any parameter exceeds the corresponding preset threshold, it is determined that the prediction accuracy of the current model has decreased, and the online learning mechanism is triggered; When the online learning mechanism is triggered, the actual flue gas parameter monitoring values ​​within a preset time window prior to the trigger time are extracted from the circular buffer to construct an online learning sample set. Each sample in the online learning sample set contains an input feature vector and a corresponding output label. The input feature vector consists of real-time state parameters collected at historical moments, and the output label is the actual flue gas parameter monitoring value at the corresponding moment. Some network layer parameters in the Transformer prediction model are frozen, and parameter fine-tuning is performed only on a few fully connected layers or specific attention layers near the output layer. A mini-batch gradient descent method is used, with the online learning sample set as training data, to refine the parameters of the original model. Based on the existing data, incremental training is performed in a limited number of rounds to update the parameters of the network layers other than the frozen layers. The fine-tuned model is quickly validated using the validation samples reserved in the online learning sample set or the latest monitoring data collected after the update in the circular buffer, and the prediction error of the updated model is calculated. If the prediction error of the updated model is lower than that before the update and is below a preset threshold, the model update is confirmed to be effective. The validated updated model parameters are saved as the current active version for subsequent flue gas parameter prediction. If the performance of the updated model does not improve or deteriorates, it is automatically rolled back to the model version before the update, and the environmental conditions that triggered the rollback are recorded.

[0042] Specifically, the system continuously collects real-time state parameters and flue gas parameter values ​​measured by sensors and stores them in a circular buffer. Historical monitoring data is extracted and compared with model predictions at preset intervals, and a comprehensive prediction error index is calculated using root mean square error or mean absolute percentage error. When the prediction error of any parameter exceeds a preset threshold, the model accuracy is deemed to have decreased, triggering online learning. After triggering, measured data prior to the trigger time is extracted from the buffer to construct an online learning sample set. Each sample contains historical state parameters as input features and the corresponding measured flue gas parameter as the output label. During training, some network layers in the Transformer model are frozen, and parameters are fine-tuned only for fully connected layers or specific attention layers near the output layer. Mini-batch gradient descent is used for a limited number of incremental training rounds. After the update, reserved validation samples or the latest monitoring data are used to verify model performance. If the error is lower than before the update and below the threshold, the update is confirmed as effective and saved as a new version; if the performance does not improve or deteriorates, the system automatically rolls back to the original version and records the environmental conditions that triggered the rollback to ensure system stability and reliability.

[0043] The implementation principle of an intelligent control method based on a smoke control system in this application is as follows: Key state parameters of a fire scenario are collected in real time, including the heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed, and the time of fire occurrence. Subsequently, these parameters are input into a pre-built smoke prediction model based on a Transformer architecture. This model employs a hybrid architecture of Transformer and convolutional neural networks or a hybrid architecture of Transformer and gated recurrent units. It captures long-range dependencies in the time series through a multi-head self-attention mechanism and utilizes a local feature extraction module to model the dynamic characteristics of the time series, quickly outputting the set height along the building space. The system analyzes the temperature distribution, visibility distribution, and smoke layer height distribution at a given location. Based on the predicted smoke parameters, a multi-objective optimization function is constructed, comprehensively considering the upstream smoke backflow length, visibility in the evacuation area, smoke layer height, and energy consumption of the smoke exhaust system. This function is solved within the feasible region of preset control parameters to generate the optimal longitudinal wind speed and smoke exhaust wind speed that minimize the overall cost. The system adjusts the fan speed and smoke exhaust valve opening in real time according to the optimized control parameters. Simultaneously, the system compares real-time monitoring data with model predictions, calculates the prediction error, and triggers an online learning mechanism when the error exceeds a preset threshold. An incremental learning algorithm is used to fine-tune the prediction model parameters, achieving adaptive model updates. This application, through the perception, prediction, execution, and updating processes of the smoke control system, realizes the transformation of the smoke control system from passive response to active prediction, significantly improving smoke exhaust efficiency and personnel evacuation safety in complex fire scenarios, while reducing system energy consumption.

[0044] Figure 1 This is a flowchart illustrating an intelligent control method based on a smoke control system in one embodiment. It should be understood that, although... Figure 1The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows; unless explicitly stated otherwise, there is no strict order requirement for the execution of these steps, and they can be executed in other orders; and Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0045] Based on the same technical concept, referring to Figure 2 This application also provides an intelligent control device based on a smoke control system, which adopts the following technical solution: the device includes: The data acquisition module is used to collect real-time status parameters of the fire-affected area. These real-time status parameters include at least the heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed, and the time of occurrence. The smoke prediction module is used to obtain smoke parameters along the building space based on a preset smoke prediction model based on the Transformer architecture. The smoke prediction model adopts a hybrid architecture of Transformer and convolutional neural network or a hybrid architecture of Transformer and gated recurrent unit. Through the smoke prediction model and real-time state parameters, the smoke parameters include at least the temperature distribution, visibility distribution and smoke layer height distribution at a set height. The set height is the critical height for personnel evacuation. The parameter control module is used to construct a multi-objective optimization function based on flue gas parameters, taking into account factors such as upstream flue gas counterflow length, visibility in the personnel evacuation area, flue gas layer height, and energy consumption of the smoke exhaust system. The multi-objective optimization function is solved within the preset feasible domain of control parameters to generate the optimal control parameters that minimize the overall cost. The optimal control parameters include at least the optimized longitudinal wind speed and smoke exhaust wind speed. The fan speed and smoke exhaust valve opening are adjusted in real time based on the optimized longitudinal wind speed and smoke exhaust wind speed. The online update module is used to compare real-time state parameters with predicted flue gas parameters and calculate prediction error. When the prediction error exceeds a preset threshold, the online learning mechanism is triggered to fine-tune the parameters of the Transformer prediction model using an incremental learning algorithm and perform adaptive model updates.

[0046] In some embodiments, the smoke prediction module is also used to establish a fire model of a highway tunnel or building space containing a cross passage using computational fluid dynamics numerical simulation software, set up a variety of fire condition combinations, and the fire condition combinations include at least the parameter values ​​of different fire source heat release rates, different longitudinal wind speeds, different smoke exhaust wind speeds, and different fire occurrence times; and obtain the temperature distribution, visibility distribution, and smoke layer height distribution data at a set height along the path under each condition through numerical simulation to form an original dataset; The original dataset is structured by organizing the heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed and fire occurrence time at each time point into input feature vectors, and organizing the corresponding temperature distribution, visibility distribution and smoke layer height distribution into output labels; all input feature vectors and output labels are normalized and mapped to a uniform numerical range. A sliding time window method is used to extract temporal features and construct training samples with time dependencies. The training samples are divided into training set and test set according to the fire conditions to ensure that the test set includes complete conditions that were not included in the training. A hybrid neural network model based on Transformer is constructed, comprising an input layer, a feature extraction layer, and an output layer. The input layer receives preprocessed temporal feature vectors. The feature extraction layer includes a Transformer encoder module and a local feature extraction module. The Transformer encoder module consists of multiple stacked encoder blocks, each containing a multi-head self-attention layer and a feedforward neural network layer, equipped with residual connections and layer normalization to capture long-range dependencies in the input sequence. The local feature extraction module uses a one-dimensional convolutional neural network or gated recurrent units to extract local spatial features from the features output by the Transformer encoder or to further model temporal dynamic characteristics. The output layer contains multiple fully connected layers that map the extracted features to a dimension equal to the number of measurement points, outputting the temperature distribution, visibility distribution, and smoke layer height distribution at a set height along the path. The mean squared error was used as the loss function, and the Adam optimizer was used to train the hybrid neural network model. The initial learning rate, batch size and maximum number of training epochs were set, and a dynamic learning rate adjustment strategy and an early stopping mechanism were introduced. When the loss on the validation set no longer decreased for several consecutive epochs, the training was terminated early, and the model parameters with the best performance on the validation set were saved. The trained hybrid neural network model was validated using working conditions not included in the training in the test set. The model accuracy was evaluated by comparing the error between the model's predicted values ​​and the simulated data corresponding to the output labels in the training samples, ensuring that the model can accurately predict the spatiotemporal distribution of smoke parameters under different fire conditions.

[0047] In some embodiments, the flue gas prediction module is further configured to set the Transformer encoder module to consist of two stacked encoder blocks, with the number of heads of the multi-head self-attention mechanism in each encoder block set to 8, and the hidden layer dimension of the feedforward neural network layer set to 128. When the local feature extraction module uses a one-dimensional convolutional neural network, the number of convolutional layers is set to 2, the number of filters in the first convolutional layer is 64, the number of filters in the second convolutional layer is 128, and the kernel size is set to 2. When the local feature extraction module uses a gated loop unit, the number of gated loop unit layers is set to 2. The first layer is a gated loop unit that returns the sequence and the number of hidden units is set to 128. The second layer is a gated loop unit that only returns the output of the last time step and the number of hidden units is set to 256. The number of fully connected layers in the output layer is set to 2. The hidden layer dimension of the first fully connected layer is set to 512, and the hidden layer dimension of the second fully connected layer is set to 256. A Dropout mechanism is introduced after each fully connected layer to prevent overfitting. Finally, the dimension of the output layer is set to be the same as the number of measurement points, which is used to output the temperature distribution, visibility distribution and flue gas layer height distribution at a set height along the path.

[0048] In some embodiments, the flue gas prediction module is specifically used to concatenate the real-time state parameters collected at the current moment with the historical state parameters collected at historical moments to form an input sequence with time dependence. Each time step in the input sequence includes four feature dimensions: heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed, and fire occurrence time. The input sequence is fed into the Transformer encoder module of the trained flue gas prediction model. The dynamic correlation weights between different time steps in the sequence are calculated through a multi-head self-attention mechanism, so that the flue gas prediction model focuses on the historical moment features that have an important impact on the current prediction. After being processed by stacking multiple encoder blocks, the output is a temporal feature representation containing long-range dependencies. The temporal feature representation is input into the local feature extraction module. When a hybrid architecture of Transformer and convolutional neural network is adopted, the temporal feature is convolved through a one-dimensional convolutional layer to extract local change features along the time dimension. When a hybrid architecture of Transformer and gated recurrent unit is adopted, the temporal feature is recursively processed through a gated recurrent unit layer to further model the dynamic characteristics of flue gas parameters evolving over time. The local variation features or dynamic characteristics output by the local feature extraction module are flattened and then input into the fully connected output layer. Through the nonlinear mapping of the multi-layer fully connected network, the high-dimensional features are converted into an output dimension that matches the number of measurement points along the building space. The fully connected output layer simultaneously outputs three parallel prediction result tensors, corresponding to the temperature distribution, visibility distribution, and smoke layer height distribution at a set height along the building space, respectively; each prediction result tensor contains parameter values ​​at each measurement point along the path.

[0049] In some embodiments, the parameter control module is specifically used to identify the position of the upstream smoke front of the fire source based on temperature distribution or smoke layer height distribution, calculate the upstream smoke backflow length between the fire source and the smoke front; extract visibility values ​​in key personnel evacuation areas from the visibility distribution, the key personnel evacuation areas including at least the personnel evacuation path upstream of the fire source and the cross passage entrance area; extract the smoke layer height at each position along the path from the smoke layer height distribution, identify the dangerous area range where the smoke layer height is lower than the height of human bodies; and calculate the instantaneous energy consumption of the smoke exhaust system based on the current fan speed and smoke exhaust valve opening. A cost function is constructed that comprehensively considers the above-mentioned multiple optimization objectives. The cost function is expressed as the weighted sum of the deviations between each optimization objective and its expected value, and includes at least the following four cost components: The first cost component is proportional to the upstream flue gas backflow length, representing the degree to which the spread of flue gas upstream of the fire source hinders personnel evacuation; the second cost component is proportional to the negative deviation between the visibility in the key personnel evacuation area and the preset safe visibility threshold, representing the impact of insufficient visibility on the personnel evacuation speed; the third cost component is proportional to the negative deviation between the smoke layer height and the preset safe height threshold, representing the degree to which the descent of the smoke layer compresses the effective evacuation space for personnel; and the fourth cost component is proportional to the instantaneous energy consumption of the smoke exhaust system, representing the economic efficiency of the system operation. Based on the physical constraints and safety specifications of the smoke control system, the feasible domain of the control parameters is set. The feasible domain includes at least the upper and lower limits of the adjustable range of the longitudinal wind speed, the upper and lower limits of the adjustable range of the smoke exhaust wind speed, and the adjustment rate limits of the fan speed and the opening of the smoke exhaust valve. Within the feasible region of the control parameters, the optimal control parameters that minimize the cost function are solved by an optimization algorithm. The optimization algorithm adopts a heuristic search algorithm or a gradient descent algorithm, and obtains the longitudinal wind speed and smoke exhaust wind speed that minimize the overall cost through iterative optimization. The optimal longitudinal wind speed and optimal smoke exhaust wind speed obtained from the solution are used as control command parameters and output to the actuator control module for subsequent adjustment of fan speed and smoke exhaust valve opening.

[0050] In some embodiments, the parameter control module is specifically used to take the optimal control parameters of the previous cycle as the initial values ​​for the optimization solution of the current cycle at the beginning of the current control cycle, and at the same time obtain the real-time state parameters and the predicted flue gas parameters at the current moment. Based on the flue gas prediction model, a flue gas parameter evolution model is constructed in the future prediction time domain. The flue gas parameter evolution model takes the current moment as the starting point and the initial value as the input to predict the dynamic changes of temperature distribution, visibility distribution and flue gas layer height distribution at a set height along the path in multiple future time steps. Within the feasible domain of the control parameters, a set of candidate control parameter sequences is generated by random sampling or grid search. Each candidate control parameter sequence contains the value sequences of longitudinal wind speed and smoke exhaust wind speed corresponding to each time step in the future prediction time domain. For each candidate control parameter sequence, the candidate control parameter sequence is input into the flue gas parameter evolution model step by step to obtain the flue gas parameter prediction results for each future time step; according to the multi-objective optimization function, the instantaneous cost corresponding to each future time step is calculated respectively; the instantaneous costs of each future time step are weighted and summed according to the time decay factor to obtain the cumulative comprehensive cost corresponding to the candidate control parameter sequence. The time decay factor is used to balance the weight of the near-term cost and the long-term cost, and the instantaneous cost closer to the current time has a larger weight; Compare the cumulative comprehensive costs of all candidate control parameter sequences, and select the candidate control parameter sequence that minimizes the cumulative comprehensive cost as the optimal control parameter for the current period; After the current control cycle is completed, the prediction time domain is rolled forward by one time step, and the optimal control parameters for the next cycle are repeatedly selected to achieve rolling optimization of model predictive control.

[0051] In some embodiments, the online update module is specifically used to continuously collect real-time status parameters and actual flue gas parameter monitoring values ​​at corresponding times during system operation, and store the collected data in a circular buffer; the actual flue gas parameter monitoring values ​​include at least the temperature value, visibility value, and flue gas layer height value measured by sensors at a set height along the path; According to the preset evaluation cycle, historical monitoring data within the preset time window before the current moment is extracted from the circular buffer and compared with the predicted value output by the flue gas prediction model at the corresponding moment. The comprehensive prediction error index is calculated. The comprehensive prediction error index adopts the root mean square error or the mean absolute percentage error, and is calculated independently or by weighted fusion for the three parameters of temperature distribution, visibility distribution and flue gas layer height distribution, respectively. The comprehensive prediction error index is compared with a preset error threshold; when the prediction error of any parameter exceeds the corresponding preset threshold, it is determined that the prediction accuracy of the current model has decreased, and the online learning mechanism is triggered. When the online learning mechanism is triggered, the actual flue gas parameter monitoring values ​​within a preset time window before the trigger time are extracted from the circular buffer to construct an online learning sample set. Each sample in the online learning sample set contains an input feature vector and a corresponding output label. The input feature vector is the real-time status parameter collected at a historical moment, and the output label is the actual flue gas parameter monitoring value at the corresponding moment. The parameters of some network layers in the Transformer prediction model are frozen, and only the parameters of a few fully connected layers or specific attention layers near the output layer are fine-tuned. The mini-batch gradient descent method is used to perform incremental training in a limited number of rounds on the basis of the original model parameters using online learning sample sets as training data, and the parameters of network layers other than the frozen layers are updated. The fine-tuned model is quickly validated using the validation samples reserved in the online learning sample set or the latest monitoring data collected after the update in the circular buffer, and the prediction error of the updated model is calculated. If the prediction error of the updated model is lower than that before the update and is below the preset threshold, the model update is confirmed to be effective. The updated model parameters that have been verified to be valid will be saved as the current active version for subsequent flue gas parameter prediction. If the performance of the updated model does not improve or deteriorates, it will be automatically rolled back to the previous model version, and the environmental conditions that triggered the rollback will be recorded.

[0052] This application also discloses a control device.

[0053] Specifically, the control device includes a memory and a processor. The memory stores a computer program that can be loaded by the processor and executed to implement the aforementioned intelligent control method based on the smoke control system.

[0054] This application also discloses a computer-readable storage medium.

[0055] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed as described above for the intelligent control method based on the smoke control system. The computer-readable storage medium includes, for example, 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.

[0056] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An intelligent control method based on a smoke control system, characterized in that, include: Collect real-time status parameters of the fire-occurring area, including at least the heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed, and time of occurrence. Based on a preset smoke prediction model based on the Transformer architecture, the smoke prediction model adopts a hybrid architecture of Transformer and convolutional neural network or a hybrid architecture of Transformer and gated recurrent unit. Through the smoke prediction model and the real-time state parameters, the smoke parameters along the building space are obtained. The smoke parameters include at least the temperature distribution, visibility distribution and smoke layer height distribution at a set height, where the set height is the critical height for personnel evacuation. Based on the flue gas parameters, a multi-objective optimization function is constructed that comprehensively considers the upstream flue gas counterflow length, visibility in the personnel evacuation area, flue gas layer height, and energy consumption of the smoke exhaust system. The multi-objective optimization function is solved within the preset feasible domain of control parameters to generate the optimal control parameters that minimize the overall cost. The optimal control parameters include at least the optimized longitudinal wind speed and smoke exhaust wind speed. The fan speed and smoke exhaust valve opening are adjusted in real time according to the optimized longitudinal wind speed and smoke exhaust wind speed. The real-time state parameters are compared with the predicted flue gas parameters to calculate the prediction error; When the prediction error exceeds a preset threshold, an online learning mechanism is triggered, which uses an incremental learning algorithm to fine-tune the parameters of the Transformer prediction model and perform adaptive model updates.

2. The intelligent control method based on a smoke control system according to claim 1, characterized in that, Before obtaining the smoke parameters along the building space using the preset Transformer-based smoke prediction model (which employs a hybrid architecture of Transformer and convolutional neural networks or a hybrid architecture of Transformer and gated recurrent units) and the real-time state parameters, the following steps are also included: Using computational fluid dynamics numerical simulation software, a fire model of a highway tunnel or building space containing cross passages is established, and various fire condition combinations are set. The fire condition combinations include parameter values ​​of different heat release rates of fire sources, different longitudinal wind speeds, different smoke exhaust wind speeds, and different fire occurrence times. The temperature distribution, visibility distribution, and smoke layer height distribution data at a set height along the route under each condition are obtained through numerical simulation to form the original dataset. The original dataset is structured by organizing the heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed and fire occurrence time at each time point into input feature vectors, and organizing the corresponding temperature distribution, visibility distribution and smoke layer height distribution into output labels; all input feature vectors and output labels are normalized and mapped to a uniform numerical range. A sliding time window method is used to extract temporal features and construct training samples with time dependencies. The training samples are divided into training set and test set according to the fire conditions to ensure that the test set includes complete conditions that were not included in the training. A hybrid neural network model based on Transformer is constructed, comprising an input layer, a feature extraction layer, and an output layer. The input layer receives preprocessed temporal feature vectors. The feature extraction layer includes a Transformer encoder module and a local feature extraction module. The Transformer encoder module consists of multiple stacked encoder blocks, each containing a multi-head self-attention layer and a feedforward neural network layer, equipped with residual connections and layer normalization to capture long-range dependencies in the input sequence. The local feature extraction module uses a one-dimensional convolutional neural network or gated recurrent units to extract local spatial features from the features output by the Transformer encoder or to further model temporal dynamic characteristics. The output layer contains multiple fully connected layers that map the extracted features to a dimension equal to the number of measurement points, outputting the temperature distribution, visibility distribution, and smoke layer height distribution at a set height along the path. The mean squared error was used as the loss function, and the Adam optimizer was used to train the hybrid neural network model. The initial learning rate, batch size and maximum number of training epochs were set, and a dynamic learning rate adjustment strategy and an early stopping mechanism were introduced. When the loss on the validation set no longer decreased for several consecutive epochs, the training was terminated early, and the model parameters with the best performance on the validation set were saved. The trained hybrid neural network model was validated using working conditions not included in the training in the test set. The model accuracy was evaluated by comparing the error between the model's predicted values ​​and the simulated data corresponding to the output labels in the training samples, ensuring that the model can accurately predict the spatiotemporal distribution of smoke parameters under different fire conditions.

3. The intelligent control method based on a smoke control system according to claim 2, characterized in that, After constructing the Transformer-based hybrid neural network model, which includes an input layer, a feature extraction layer, and an output layer, the model further includes: The Transformer encoder module is set to consist of two stacked encoder blocks. The number of heads in the multi-head self-attention mechanism in each encoder block is set to 8, and the hidden layer dimension of the feedforward neural network layer is set to 128. When the local feature extraction module uses a one-dimensional convolutional neural network, the number of convolutional layers is set to 2, the number of filters in the first convolutional layer is 64, the number of filters in the second convolutional layer is 128, and the kernel size is set to 2. When the local feature extraction module uses a gated loop unit, the number of gated loop unit layers is set to 2. The first layer is a gated loop unit that returns the sequence and the number of hidden units is set to 128. The second layer is a gated loop unit that only returns the output of the last time step and the number of hidden units is set to 256. The number of fully connected layers in the output layer is set to 2. The hidden layer dimension of the first fully connected layer is set to 512, and the hidden layer dimension of the second fully connected layer is set to 256. A Dropout mechanism is introduced after each fully connected layer to prevent overfitting. Finally, the dimension of the output layer is set to be the same as the number of measurement points, which is used to output the temperature distribution, visibility distribution and flue gas layer height distribution at a set height along the path.

4. The intelligent control method based on a smoke control system according to claim 3, characterized in that, The pre-defined flue gas prediction model based on the Transformer architecture, employing a hybrid architecture of Transformer and convolutional neural network or a hybrid architecture of Transformer and gated recurrent units, obtains flue gas parameters along the building space using the flue gas prediction model and the real-time state parameters, including: The real-time state parameters collected at the current moment are concatenated with the historical state parameters collected at previous moments to form an input sequence with time dependence. Each time step in the input sequence contains four feature dimensions: heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed, and fire occurrence time. The input sequence is fed into the Transformer encoder module of the trained flue gas prediction model. The dynamic correlation weights between different time steps in the sequence are calculated through a multi-head self-attention mechanism, so that the flue gas prediction model focuses on the historical moment features that have an important impact on the current prediction. After being processed by stacking multiple encoder blocks, the output is a temporal feature representation containing long-range dependencies. The temporal feature representation is input into the local feature extraction module. When a hybrid architecture of Transformer and convolutional neural network is adopted, the temporal feature is convolved through a one-dimensional convolutional layer to extract local change features along the time dimension. When a hybrid architecture of Transformer and gated recurrent unit is adopted, the temporal feature is recursively processed through a gated recurrent unit layer to further model the dynamic characteristics of flue gas parameters evolving over time. The local variation features or dynamic characteristics output by the local feature extraction module are flattened and then input into the fully connected output layer. Through the nonlinear mapping of the multi-layer fully connected network, the high-dimensional features are converted into an output dimension that matches the number of measurement points along the building space. The fully connected output layer simultaneously outputs three parallel prediction result tensors, corresponding to the temperature distribution, visibility distribution, and smoke layer height distribution at a set height along the building space, respectively; each prediction result tensor contains parameter values ​​at each measurement point along the path.

5. The intelligent control method based on a smoke control system according to claim 1, characterized in that, Based on the flue gas parameters, a multi-objective optimization function is constructed that comprehensively considers the upstream flue gas counterflow length, visibility in the personnel evacuation area, flue gas layer height, and energy consumption of the smoke exhaust system. The multi-objective optimization function is solved within the preset feasible region of control parameters to generate optimal control parameters that minimize the overall cost, including: Based on temperature distribution or smoke layer height distribution, identify the position of the smoke front upstream of the fire source and calculate the upstream smoke backflow length between the fire source and the smoke front; extract visibility values ​​in key evacuation areas from the visibility distribution, which include at least the evacuation path upstream of the fire source and the cross passage entrance area; extract the smoke layer height at each location along the path from the smoke layer height distribution and identify dangerous areas where the smoke layer height is lower than human height; calculate the instantaneous energy consumption of the smoke exhaust system based on the current fan speed and smoke exhaust valve opening. A cost function is constructed that comprehensively considers the above-mentioned multiple optimization objectives. The cost function is expressed as the weighted sum of the deviations between each optimization objective and its expected value, and includes at least the following four cost components: The first cost component is proportional to the upstream flue gas backflow length, representing the degree to which the spread of flue gas upstream of the fire source hinders personnel evacuation; the second cost component is proportional to the negative deviation between the visibility in the key personnel evacuation area and the preset safe visibility threshold, representing the impact of insufficient visibility on the personnel evacuation speed; the third cost component is proportional to the negative deviation between the smoke layer height and the preset safe height threshold, representing the degree to which the descent of the smoke layer compresses the effective evacuation space for personnel; and the fourth cost component is proportional to the instantaneous energy consumption of the smoke exhaust system, representing the economic efficiency of the system operation. Based on the physical constraints and safety specifications of the smoke control system, the feasible domain of the control parameters is set. The feasible domain includes at least the upper and lower limits of the adjustable range of the longitudinal wind speed, the upper and lower limits of the adjustable range of the smoke exhaust wind speed, and the adjustment rate limits of the fan speed and the opening of the smoke exhaust valve. Within the feasible region of the control parameters, the optimal control parameters that minimize the cost function are solved by an optimization algorithm. The optimization algorithm adopts a heuristic search algorithm or a gradient descent algorithm, and obtains the longitudinal wind speed and smoke exhaust wind speed that minimize the overall cost through iterative optimization. The optimal longitudinal wind speed and optimal smoke exhaust wind speed obtained from the solution are used as control command parameters and output to the actuator control module for subsequent adjustment of fan speed and smoke exhaust valve opening.

6. The intelligent control method based on a smoke control system according to claim 5, characterized in that, The step of finding the optimal control parameters that minimize the cost function within the feasible region of the control parameters using an optimization algorithm includes: At the start of the current control cycle, the optimal control parameters of the previous cycle are used as the initial values ​​for the optimization solution of the current cycle, and the real-time state parameters and the predicted flue gas parameters at the current moment are obtained. Based on the flue gas prediction model, a flue gas parameter evolution model is constructed in the future prediction time domain. The flue gas parameter evolution model takes the current moment as the starting point and the initial value as the input to predict the dynamic changes of temperature distribution, visibility distribution and flue gas layer height distribution at a set height along the path in multiple future time steps. Within the feasible domain of the control parameters, a set of candidate control parameter sequences is generated by random sampling or grid search. Each candidate control parameter sequence contains the value sequences of longitudinal wind speed and smoke exhaust wind speed corresponding to each time step in the future prediction time domain. For each candidate control parameter sequence, the candidate control parameter sequence is input into the flue gas parameter evolution model step by step to obtain the flue gas parameter prediction results for each future time step; according to the multi-objective optimization function, the instantaneous cost corresponding to each future time step is calculated respectively; the instantaneous costs of each future time step are weighted and summed according to the time decay factor to obtain the cumulative comprehensive cost corresponding to the candidate control parameter sequence. The time decay factor is used to balance the weight of the near-term cost and the long-term cost, and the instantaneous cost closer to the current time has a larger weight; Compare the cumulative integrated costs of all candidate control parameter sequences, and select the candidate control parameter sequence that minimizes the cumulative integrated cost as the optimal control parameter for the current period; After the current control cycle is completed, the prediction time domain is rolled forward by one time step, and the optimal control parameters for the next cycle are repeatedly selected to achieve rolling optimization of model predictive control.

7. The intelligent control method based on a smoke control system according to claim 1, characterized in that, The process involves comparing the real-time state parameters with the predicted flue gas parameters to calculate the prediction error; when the prediction error exceeds a preset threshold, an online learning mechanism is triggered to fine-tune the parameters of the Transformer prediction model using an incremental learning algorithm, and to perform adaptive model updates, including: During system operation, the real-time status parameters and the actual flue gas parameter monitoring values ​​at the corresponding time are continuously collected, and the collected data is stored in a circular buffer; the actual flue gas parameter monitoring values ​​include at least the temperature value, visibility value, and flue gas layer height value measured by sensors at a set height along the path; According to the preset evaluation cycle, historical monitoring data within the preset time window before the current moment is extracted from the circular buffer and compared with the predicted value output by the flue gas prediction model at the corresponding moment. The comprehensive prediction error index is calculated. The comprehensive prediction error index adopts the root mean square error or the mean absolute percentage error, and is calculated independently or by weighted fusion for the three parameters of temperature distribution, visibility distribution and flue gas layer height distribution, respectively. The comprehensive prediction error index is compared with a preset error threshold; when the prediction error of any parameter exceeds the corresponding preset threshold, it is determined that the prediction accuracy of the current model has decreased, and the online learning mechanism is triggered. When the online learning mechanism is triggered, the actual flue gas parameter monitoring values ​​within a preset time window before the trigger time are extracted from the circular buffer to construct an online learning sample set. Each sample in the online learning sample set contains an input feature vector and a corresponding output label. The input feature vector is the real-time state parameter collected at a historical time, and the output label is the actual flue gas parameter monitoring value at the corresponding time. Freeze some network layer parameters in the Transformer prediction model, and fine-tune the parameters of only a few fully connected layers or specific attention layers near the output layer; use mini-batch gradient descent, with the online learning sample set as training data, to perform incremental training for a limited number of rounds on the basis of the original model parameters, and update the network layer parameters other than the frozen layers. The fine-tuned model is quickly validated using the validation samples reserved in the online learning sample set, or using the latest monitoring data collected after the update in the circular buffer, and the prediction error of the updated model is calculated. If the prediction error of the updated model is lower than that before the update and is below a preset threshold, the model update is confirmed to be effective. The updated model parameters that have been verified to be valid will be saved as the current active version for subsequent flue gas parameter prediction. If the performance of the updated model does not improve or deteriorates, it will be automatically rolled back to the previous model version, and the environmental conditions that triggered the rollback will be recorded.

8. An intelligent control device based on a smoke control system, characterized in that, The device includes: The data acquisition module is used to collect real-time status parameters of the fire-occurring area. The real-time status parameters include at least the heat release rate of the fire source, longitudinal wind speed, smoke exhaust wind speed, and the time of occurrence. The smoke prediction module is used to obtain smoke parameters along the building space based on a preset smoke prediction model based on the Transformer architecture. The smoke prediction model adopts a hybrid architecture of Transformer and convolutional neural network or a hybrid architecture of Transformer and gated recurrent unit. Through the smoke prediction model and the real-time state parameters, the smoke parameters include at least the temperature distribution, visibility distribution and smoke layer height distribution at a set height, where the set height is the critical height for personnel evacuation. The parameter control module is used to construct a multi-objective optimization function based on the flue gas parameters, which comprehensively considers the upstream flue gas counterflow length, visibility in the personnel evacuation area, flue gas layer height, and energy consumption of the smoke exhaust system. The module solves the multi-objective optimization function within the preset feasible domain of control parameters to generate the optimal control parameters that minimize the overall cost. The optimal control parameters include at least the optimized longitudinal wind speed and smoke exhaust wind speed. The module adjusts the fan speed and smoke exhaust valve opening in real time based on the optimized longitudinal wind speed and smoke exhaust wind speed. The online update module is used to compare the real-time state parameters with the predicted flue gas parameters and calculate the prediction error. When the prediction error exceeds a preset threshold, the online learning mechanism is triggered to fine-tune the parameters of the Transformer prediction model using an incremental learning algorithm and perform adaptive model updates.

9. A control device, characterized in that, The device includes: A memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 7.

Images