Tunnel fire exhaust fan adaptive PID intelligent control method based on light-weight PINN fusion

Abstract

The application discloses a tunnel fire smoke exhaust fan adaptive PID intelligent regulation and control method fusing a lightweight PINN, first constructs a tunnel fire FDS numerical model with a dynamic ventilation boundary, and builds a Python-FDS two-way coupling interface; secondly, constructs a lightweight physical information neural network online physical prior model to generate adaptive physical boundary constraints; designs a PINN-LSTM time series prediction module, fuses multi-source sensing data to predict future environmental parameters to generate feedforward estimates; on this basis, constructs a predicted feedforward and adaptive feedback double-layer hybrid PID controller, parameters of which are online set by a deep neural network, and parameter setting and verification are completed by a four-dimensional hybrid objective function; finally, realizes the predicted-feedback closed-loop intelligent regulation and control of the fan air volume through the coupling interface. The application significantly shortens the control regulation time, reduces the overshoot, and improves the control stability and physical rationality under complex fire conditions.

Description

Technical Field

[0001] This invention relates to the field of tunnel fire protection engineering technology, and in particular to an adaptive PID intelligent control method for tunnel fire smoke exhaust fans that integrates lightweight PINN. Background Technology

[0002] The long, narrow, and enclosed spatial structure of tunnels causes hot, toxic smoke to spread rapidly longitudinally after a fire breaks out. When longitudinal ventilation is insufficient, the hot smoke can spread upstream of the fire source, forming a smoke backflow zone. This seriously threatens the evacuation safety of personnel upstream of the fire source and the access route for firefighters, making it the primary cause of casualties in tunnel fires. Precise and rapid control of smoke exhaust fans is the core means to suppress the spread of smoke in tunnel fires and ensure the safety of personnel evacuation.

[0003] Currently, the methods for controlling smoke exhaust fans in tunnel fires are mainly divided into three categories: The first category is the fixed air volume ventilation mode, which presets a fixed fan air volume based on the design conditions. This method cannot adapt to the dynamic increase of fire source power and the spatiotemporal changes in smoke spread during the fire development process. It is prone to insufficient air volume leading to smoke backflow, or excessive air volume leading to turbulent smoke diffusion, resulting in extremely poor fire control effect. The second category is the conventional fixed parameter PID control method, which achieves closed-loop control of fan air volume through single-point temperature feedback. Although this method can achieve dynamic adjustment, tunnel fire systems have strong nonlinearity, large lag, and strong time-varying characteristics. The dynamic characteristics of different stages of fire development vary greatly, and fixed parameter PID cannot achieve optimal control at all times. It is common for the initial stage of a fire to be unsuitable for optimal control. The problems of over-control, large overshoot, delayed adjustment response, and large fluctuations in the steady-state stage make it difficult to meet the rapid and stable control requirements in emergency scenarios. The third category is intelligent control methods, such as fuzzy control and model predictive control. Although these methods can improve control performance, they generally have two core defects: First, they lack specific physical constraints on the flow and ventilation characteristics of smoke in tunnel fires, which can easily lead to control outputs that do not conform to engineering realities, such as reverse ventilation and air volume exceeding the rated operating conditions of the fan. Their physical rationality is insufficient, making them difficult to implement. Second, the coupling between existing control algorithms and CFD numerical simulations is mostly in offline simulation mode, which cannot achieve real-time synchronous iteration between the control algorithm and the dynamic evolution of the fire. The control effect is seriously out of sync with the actual engineering scenario.

[0004] In recent years, deep learning technology has shown great potential in the fields of time series prediction and complex system control, but there are still many challenges in applying it to tunnel fire smoke control: pure data-driven deep learning models lack physical constraints and are prone to producing prediction results and control outputs that do not conform to engineering reality; large models have slow inference speeds and cannot meet the requirements of millisecond-level real-time response in emergency scenarios; the black-box nature of deep learning models makes it difficult to guarantee their reliability, and once the model fails, it will cause serious safety accidents. Summary of the Invention

[0005] This invention aims to at least partially address one of the technical problems in related technologies. To this end, the first objective of this invention is to propose an adaptive PID intelligent control method for tunnel fire smoke exhaust fans that integrates lightweight PINN (Physical Information Network). This method achieves online dynamic self-correction of physical priors by introducing a lightweight physical information neural network; designs a PINN-LSTM multi-source timing prediction module to upgrade pure feedback control to a hybrid predictive feedforward + feedback control, fundamentally solving the large time lag problem; constructs an online PID parameter tuner based on a deep neural network to improve generalization under complex operating conditions; and retains the original patented physical prior PID as a backup control mechanism to ensure the reliability of the system under extreme conditions. This invention significantly improves the response speed, control accuracy, and robustness under complex operating conditions of tunnel fire smoke exhaust control while maintaining millisecond-level real-time response speed, providing more advanced core control technology support for tunnel fire smoke prevention and emergency rescue.

[0006] To achieve the above objectives, a first aspect of the present invention proposes an adaptive PID intelligent control method for tunnel fire smoke exhaust fans that integrates lightweight PINN, comprising the following steps:

[0007] S1. Construct a multi-source sensor fusion FDS numerical model for tunnel fire with dynamic ventilation boundary, and build a Python-FDS bidirectional coupling interface for millisecond-level synchronous iteration of control algorithm and CFD numerical calculation.

[0008] S2. Construct a lightweight PINN online physical prior model specifically for tunnel fires to replace the fixed physical parameters of offline fitting. Solve dynamic physical quantities such as smoke spread length, critical smoke exhaust volume and smoke backflow risk coefficient in real time, and then generate dynamic physical boundary constraints that evolve with the fire situation.

[0009] S3, construct a PINN-LSTM multi-source time series prediction module, integrate temperature, CO concentration, visibility and wind speed data to predict the changes in environmental parameters of the core control area in the future period, and generate feedforward control prediction values ​​accordingly.

[0010] S4. Construct a two-layer hybrid PID controller that includes a predictive feedforward layer and an adaptive feedback layer. The predictive feedforward layer offsets foreseeable flue gas disturbances in advance based on the prediction results, and the adaptive feedback layer eliminates residual errors based on real-time monitoring data. The PID parameters are adaptively tuned online by a deep neural network.

[0011] S5 constructs a multi-source sensor data fusion preprocessing and active anti-interference compensation module to achieve data spatiotemporal alignment, noise reduction, anomaly repair and fault pre-tolerance;

[0012] S6. Construct a four-dimensional hybrid control objective function that integrates control error loss, online physical prior constraint loss, air volume smoothing loss and prediction error loss to complete the initial tuning of controller parameters and verification under multiple extreme conditions.

[0013] S7, based on the tuned hybrid PID controller, performs predictive-feedback closed-loop intelligent control of the volumetric flow rate of the tunnel fire smoke exhaust fan through the Python-FDS bidirectional coupling interface, and outputs the fan air volume timing control command.

[0014] Furthermore, the adaptive PID intelligent control method for tunnel fire smoke exhaust fans incorporating lightweight PINN according to the above embodiments of the present invention may also have the following additional technical features:

[0015] According to an embodiment of the present invention, step S1 includes:

[0016] S11. Construct a tunnel fire FDS numerical model that matches the actual engineering situation, and determine the number and corresponding location of fire sources, the configuration range of fire source power, and the configuration parameters of longitudinal ventilation in the FDS numerical model.

[0017] S12, determine the range of the fire source, the power growth of the fire source follows the t² fire model, the total simulation time is the first preset time; the fuel types are gasoline, diesel, and plastic, and the corresponding combustion product generation rate and heat release rate curves are set respectively; the tunnel wall material is defined as concrete, and the wall heat accumulation effect and heat radiation parameters are introduced.

[0018] S13, a dynamic grid partitioning strategy is used to partition the tunnel space;

[0019] S14, the ventilation system includes two sets of jet fans, symmetrically arranged in the tunnel. Each set of jet fans contains 2 fans. The entire cross-section of the fan air supply is defined as a VENT type boundary. A new axial flow smoke exhaust fan and smoke exhaust outlet linkage control interface are added. The smoke exhaust outlet opening degree supports continuous adjustment from 0 to 100%.

[0020] S15, along the longitudinal centerline of the tunnel, multiple composite sensors are deployed at preset intervals. The composite sensors integrate four monitoring functions: temperature, CO concentration, visibility, and wind speed. Multiple core control measuring points are set up upstream of the fire source, and multiple backup monitoring measuring points are set up upstream and downstream of the fire source.

[0021] S16 uses an asynchronous Python script to build a bidirectional coupling interface between the FDS and the control algorithm, completing three synchronization steps in each 0.5s calculation time step: extracting multi-source real-time monitoring data from all sensors in the FDS numerical model; updating the target fan volume flow rate and smoke exhaust port opening output by the hybrid PID controller to the corresponding VENT boundary parameters in real time; synchronizing the fire evolution data to the digital twin buffer; and setting hard constraints on air volume (0~40m³ / s) and smoke exhaust port opening (0~100%).

[0022] According to an embodiment of the present invention, step S2 includes:

[0023] S21. Construct the governing equations for smoke flow in tunnel fires, serving as physical constraints for the lightweight PINN online physical prior model. These equations include the continuity equation, momentum conservation equation, energy conservation equation, and smoke component transport equation, expressed as follows:

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] In the formula, ρ is air density, u is velocity vector, p is pressure, g is gravitational acceleration, τ is viscous stress tensor, h is specific enthalpy, k is thermal conductivity coefficient, T is temperature, and qˊ c Y represents the heat release rate of combustion. i Let D be the mass fraction of the i-th component. i ωˊ is the diffusion coefficient of the component. i The component formation rate;

[0029] S22, a lightweight PINN network architecture is constructed, employing a fully connected neural network with 3 hidden layers and 64 neurons per layer. Channel pruning and quantization techniques are introduced to compress the number of model parameters to less than 10% of the original model, ensuring a single-step inference time of less than 10ms. The network input consists of the current fire source power, ventilation volume, and core measurement point monitoring data, and the output is the smoke spread length L. phy (t), Critical exhaust air volume V cri (t) and flue gas recirculation risk coefficient R back (t);

[0030] S23. Design a hybrid loss function for the lightweight PINN online physics prior model, which is a weighted sum of data fitting loss and physics constraint loss, expressed as follows:

[0031] ;

[0032] In the formula, L data L represents the mean square error loss between the model output and the FDS simulation data. phy To control the residual loss of the equation, This is the weighting coefficient for data loss, with a value ranging from 0.3 to 0.7. This is the weighting coefficient for the total physical constraint loss, with a value ranging from 0.5 to 1.0;

[0033] S24, an online update mechanism for a lightweight PINN online physical prior model is constructed, which fine-tunes the model every 10 seconds using the latest monitoring data to correct model parameters to adapt to the actual fire evolution; when the flue gas backflow risk coefficient R back When (t)>0.7, the model's emergency update mode is triggered, shortening the update interval to 2s;

[0034] S25, based on the dynamic physical quantities output by the lightweight PINN online physical prior model, generates dynamic upper and lower limit constraints for the wind turbine volumetric flow rate, expressed as follows:

[0035] ;

[0036] ;

[0037] In the formula, V rated The rated maximum volumetric flow rate of the fan is taken as 40 m³ / s; Q(t) is the power of the fire source at time t; k1 and k2 are the tunnel ventilation characteristic coefficients, which are adaptively adjusted online by the lightweight PINN online physical prior model.

[0038] According to an embodiment of the present invention, step S3 includes:

[0039] S31, construct a multi-source data attention fusion layer, and use a multi-head self-attention mechanism to automatically learn the weights of four types of data: temperature, CO concentration, visibility, and wind speed, highlighting features that have a greater impact on smoke spread. The fused feature expression is as follows:

[0040] ;

[0041] In the formula, F i (t) represents the normalized feature of the data from the i-th sensor, α i (t) represents the attention weight for the i-th feature;

[0042] S32, construct an LSTM time series prediction network. The input is the fused feature sequence of the past 20 seconds, and the output is the predicted values ​​of temperature, CO concentration and visibility of the core control measurement points at two time steps of 5 seconds and 10 seconds. The LSTM network contains 2 LSTM layers and 1 fully connected output layer, with 128 hidden layer neurons.

[0043] S33, Design a PINN-guided prediction correction mechanism, using the physical priors output by the lightweight PINN online physical prior model as constraints for the LSTM measurement, to correct the physical rationality of the prediction results. The expression for the corrected prediction value is as follows:

[0044] ;

[0045] In the formula, yˊ LSTM (t+τ) represents the original prediction value of the LSTM network, y PINN (t+τ) is the extrapolation value of the lightweight PINN online physics prior model based on physical laws, and β is the correction coefficient, which ranges from 0.6 to 0.9.

[0046] S34: Construct a prediction error assessment and early warning mechanism, calculate the mean square error between the predicted value and the actual monitored value in real time, and when the prediction error exceeds the threshold for three consecutive time steps, automatically reduce the prediction step size to 3s and trigger an emergency update of the lightweight PINN online physical prior model.

[0047] According to an embodiment of the present invention, step S4 includes:

[0048] S41, Construct a control error system, and define the real-time control error e(t) and the predictive control error eˊ(t+τ) as follows:

[0049] ;

[0050] ;

[0051] In the formula, T d (t) represents the real-time temperature of the core control measuring point, Tˊ d (t+τ) represents the predicted temperature after time τ, where T is the temperature. t (t) represents the dynamic target control temperature, which is adaptively adjusted according to the personnel evacuation status and fire level, with a base value of 100℃.

[0052] S42, Design the feedforward control layer, and calculate the feedforward control quantity u based on the predictive control error. ff (t), the fan air volume is adjusted in advance to counteract future flue gas disturbances, and the expression is as follows:

[0053] ;

[0054] In the formula, K ff The feedforward gain coefficient is adaptively adjusted online by the deep neural network.

[0055] S43, an adaptive feedback control layer is designed, employing an improved discrete PID algorithm, integrating derivative-first and integral anti-saturation modules, with the feedback control quantity u... fb The expression for (t) is as follows:

[0056] ;

[0057] In the formula, n is the current calculation time step number, and K p (t), K i (t), K d (t) represents the proportional, integral, and differential coefficients of the current step adaptive tuning;

[0058] S44. Construct an online PID parameter tuner based on a fully connected neural network. The inputs are the control error e(t) at the current moment, the error change rate Δe(t) = e(t) - e(t-1), the fire source power Q(t), and the flue gas backflow risk coefficient R. back (t), the output is the adjustment amount ΔK of the PID parameters. p ΔK i ΔK d The parameter tuner contains three hidden layers, each with 32 neurons, and uses the ReLU activation function.

[0059] S45, the total control quantity is the weighted sum of the feedforward control quantity and the feedback control quantity, and is output after correction by the dynamic physical boundary constraints generated in step S42, as shown in the following expression:

[0060] ;

[0061] In the formula, γ is the feedforward-feedback weight coefficient, which is 0.7 when the prediction error is less than the threshold and 0.3 when the prediction error is greater than the threshold, so as to give priority to ensuring the reliability of feedback control.

[0062] According to an embodiment of the present invention, step S5 includes:

[0063] S51 performs spatiotemporal alignment processing on multi-source sensor data, unifying sensor data with different sampling frequencies and spatial locations into a 0.5s time step and the tunnel longitudinal coordinate system.

[0064] S52 employs an improved first-order exponential moving average method to denoise the data. The smoothing coefficient α is adaptively adjusted according to the degree of data fluctuation, taking 0.3 when the data fluctuation exceeds a preset threshold and 0.5 when the data is stable. The denoising formula is as follows:

[0065] ;

[0066] In the formula, x(t) represents the raw sensor data. The data after noise reduction;

[0067] S53 identifies outliers based on the improved 3σ criterion and the isolated forest algorithm. For single-point outliers, the weighted average of the data from the preceding and following 5 frames is used to replace them. For outliers in more than 5 consecutive frames, a sensor fault warning is triggered 10 seconds in advance, and the system automatically switches to backup measurement point data.

[0068] S54 is designed as an active detection and compensation module for sudden changes in operating conditions. Based on the spatiotemporal gradient change rate of multi-source data, it identifies sudden changes in operating conditions such as a sudden increase in fire source power, smoke backflow, and vehicle congestion. When a sudden change in operating conditions is detected, it automatically increases the feedforward control weight and proportional coefficient to shorten the system response time.

[0069] According to an embodiment of the present invention, step S6 includes:

[0070] S61, Construct the main loss term for control error, and use the sum of squared errors index to measure control accuracy, as shown in the following expression:

[0071] ;

[0072] In the formula, N is the total simulation time steps;

[0073] S62, construct an online physics prior constraint loss term to penalize control quantities that exceed the dynamic physics boundaries generated by the lightweight PINN online physics prior model, as shown in the following expression:

[0074] ;

[0075] In the formula, V(t) is the volumetric flow rate control value of the fan at time t, and λ phy This is the weighting coefficient for the physical constraint loss, with a value ranging from 0.8 to 1.2;

[0076] S63, constructs a smoothing loss term for airflow to suppress drastic fluctuations in fan airflow and avoid frequent start-ups and shutdowns and damage to the equipment. The expression is as follows:

[0077] ;

[0078] In the formula, λ smooth The smoothing constraint weight coefficient is set to a value between 0.005 and 0.015.

[0079] S64, Construct the prediction error loss term to measure the accuracy of the time series prediction module, with the following expression:

[0080] ;

[0081] In the formula, λ pred The prediction error weighting coefficient has a value of 0.3 to 0.7.

[0082] S65, the overall control objective function is the weighted sum of all loss terms, expressed as follows:

[0083] ;

[0084] S66, based on the Adam gradient descent algorithm and multi-condition simulation data, takes minimizing the overall control objective function as the optimization objective and completes the initial training of the PID parameter tuner and timing prediction module. Through batch simulations of 7 fire source locations, 10 fire source powers, 15 wind speed conditions and 5 extreme disturbance conditions, the generalization verification and parameter optimization of the controller are completed.

[0085] According to an embodiment of the present invention, step S7 includes:

[0086] S71, load the tuned parameters of the hybrid PID controller, the parameters of the lightweight PINN online physical prior model, the parameters of the LSTM time series prediction model, and the data preprocessing parameters;

[0087] S72 extracts real-time monitoring data from all composite sensors in the tunnel in 0.5s time steps through the Python-FDS bidirectional coupling interface. After being processed by the multi-source data preprocessing module, the data is input into the PINN online physical prior model and the LSTM time series prediction module, respectively.

[0088] S73, a lightweight PINN online physical prior model, calculates dynamic physical quantities and physical boundary constraints in real time; the LSTM module predicts environmental parameter changes in the next 5-10 seconds; the hybrid PID controller combines feedforward control and feedback control to calculate the output target fan volume flow rate and exhaust port opening.

[0089] S74 updates the target control quantity to the corresponding boundary of the FDS numerical model in real time through the coupling interface, completes the CFD calculation and control iteration of the next time step, and realizes the prediction-feedback closed-loop control of the fan air volume.

[0090] The S75 stores multi-source measurement point timing data, fan operation data, PID parameter adjustment data, lightweight PINN online physical prior model output data and prediction error data in real time during the control process, generates control effect analysis reports and visualization curves, and outputs them synchronously to the tunnel fire emergency prevention and control platform.

[0091] Compared with the prior art, the beneficial effects of the present invention are reflected in:

[0092] (1) The control response speed and accuracy are greatly improved: By introducing the PINN-LSTM predictive feedforward mechanism, the adjustment time of tunnel fire smoke exhaust control is shortened by more than 55%, the overshoot is reduced by more than 75%, and the system response lag time is shortened by more than 60%, which fundamentally solves the problem of large lag in traditional feedback control.

[0093] (2) Physical rationality and reliability are significantly enhanced: The lightweight PINN model embeds the flue gas flow control equation as a physical constraint into the neural network to ensure that the control output always conforms to the laws of thermodynamics; at the same time, it retains the offline physical prior backup control mechanism to ensure the system reliability under extreme conditions.

[0094] (3) Strong adaptability to complex fire conditions: The PINN model online update mechanism and deep neural network PID parameter tuner enable the controller to adapt to changes in fire source power, differences in fuel type and ventilation boundary disturbances, significantly improving the generalization ability to multiple fuel types such as gasoline, diesel, and plastic and extreme working conditions such as sudden power increases.

[0095] (4) Millisecond-level real-time response: Through model compression and lightweight design, PINN's single-step inference time is less than 10ms, which meets the stringent real-time requirements in emergency scenarios.

[0096] (5) Comprehensive system robustness: The multi-source data fusion preprocessing and active anti-interference module effectively copes with sensor noise, anomalies and sudden changes in operating conditions, ensuring the quality of control input data and the stability of control actions.

[0097] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0098] Figure 1 A flowchart of an adaptive PID intelligent control method for tunnel fire smoke exhaust fans incorporating lightweight PINN according to an embodiment of the present invention;

[0099] Figure 2 This is a schematic diagram of a multi-source sensor fusion FDS numerical model for tunnel fires according to an embodiment of the present invention. Detailed Implementation

[0100] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0101] The adaptive PID intelligent control method for tunnel fire smoke exhaust fans, which integrates lightweight PINN, is described below with reference to the accompanying drawings.

[0102] like Figure 1 As shown in the figure, the adaptive PID intelligent control method for tunnel fire smoke exhaust fans integrating lightweight PINN in this embodiment of the invention includes the following steps:

[0103] S1. Construct a multi-source sensor fusion FDS numerical model for tunnel fires with dynamic ventilation boundaries, and build a Python-FDS bidirectional coupling interface for millisecond-level synchronous iteration of control algorithms and CFD numerical calculations.

[0104] According to an embodiment of the present invention, step S1 includes:

[0105] S11. Construct a tunnel fire FDS numerical model that matches the actual engineering situation, and determine the number and corresponding location of fire sources, the configuration range of fire source power, and the configuration parameters of longitudinal ventilation in the FDS numerical model.

[0106] For example, a numerical model of tunnel fire FDS matching the actual engineering situation is constructed. The tunnel dimensions are 300m long, 9m wide, and 6m high. Five fire source locations are set at 50m, 100m, 150m, 200m, and 250m from the tunnel entrance. The fire source power is set in 8 levels with 5MW intervals in the range of 5MW to 40MW. The longitudinal ventilation velocity is set in 12 levels with 1m / s intervals in the range of 1m / s to 12m / s.

[0107] S12, determine the range of the fire source, the power growth of the fire source follows the t² fire model, the total simulation time is the first preset time; the fuel types are gasoline, diesel, and plastic, and the corresponding combustion product generation rate and heat release rate curves are set respectively; the tunnel wall material is defined as concrete, and the wall heat accumulation effect and heat radiation parameters are introduced.

[0108] For example, the fire source is set as a rectangular area with a cross-section of 2m×5m and a height of 0.5m above the ground. Its power growth follows the t² fire model and the exponential decay model. The total simulation time is not less than 600s. The fuel type is expanded to three typical vehicle fuels: gasoline, diesel, and plastic. Corresponding combustion product generation rate and heat release rate curves are set respectively. The tunnel wall material is defined as concrete, and the wall heat accumulation effect and heat radiation parameters are introduced.

[0109] S13 uses a dynamic grid partitioning strategy to divide the tunnel space.

[0110] For example, an adaptive dynamic grid partitioning strategy is adopted. A high-precision grid of 0.25m×0.25m×0.25m is used in the 50m radius area of ​​the fire source center and the 5m area around the fan. The flue gas front movement area is automatically densified to a 0.3m×0.3m×0.3m grid. The remaining area of ​​the tunnel uses a standard grid of 0.5m×0.5m×0.5m. The boundary conditions at both ends of the tunnel are set to "OPEN", and the boundary conditions of the connecting passage are set to "OPEN / CLOSED" which can be dynamically switched.

[0111] S14, the ventilation system includes two sets of jet fans, symmetrically arranged in the tunnel. Each set of jet fans contains 2 fans. The entire cross-section of the fan air supply is defined as a VENT type boundary. A new axial flow smoke exhaust fan and smoke exhaust outlet linkage control interface are added. The smoke exhaust outlet opening degree supports continuous adjustment from 0 to 100%.

[0112] For example, the ventilation system consists of two sets of jet fans, symmetrically arranged at distances of 75m and 225m from the tunnel entrance, with each set containing two fans; the rated volumetric flow rate of a single fan is 40m³ / s, and the entire cross-section of the fan's air supply is defined as a VENT type boundary, supporting continuous airflow adjustment from 0 to 40m³ / s; a new axial flow smoke exhaust fan and smoke exhaust outlet linkage control interface are added, and the smoke exhaust outlet opening supports continuous adjustment from 0 to 100%.

[0113] S15, along the longitudinal centerline of the tunnel, is equipped with multiple composite sensors at preset intervals. The composite sensors integrate four monitoring functions: temperature, CO concentration, visibility, and wind speed. Multiple core control monitoring points are set up upstream of the fire source, and multiple backup monitoring points are set up upstream and downstream of the fire source.

[0114] For example, composite sensors are deployed at 5m intervals along the longitudinal centerline of the tunnel. Each sensor integrates four monitoring functions: temperature, CO concentration, visibility, and wind speed. A total of 61 sensors are used, with a sampling time interval of 0.5s. Three core control monitoring points are set at the ceiling positions 10m, 20m, and 30m upstream of the fire source, and five backup monitoring points are set at positions both upstream and downstream of the fire source. The multi-source sensor fusion tunnel fire FDS numerical model is as follows: Figure 2 As shown.

[0115] S16 uses an asynchronous Python script to build a bidirectional coupling interface between the FDS and the control algorithm, completing three synchronization steps in each 0.5s calculation time step: extracting multi-source real-time monitoring data from all sensors in the FDS numerical model; updating the target fan volume flow rate and smoke exhaust port opening output by the hybrid PID controller to the corresponding VENT boundary parameters in real time; synchronizing the fire evolution data to the digital twin buffer; and setting hard constraints on air volume (0~40m³ / s) and smoke exhaust port opening (0~100%).

[0116] S2 constructs a lightweight PINN online physical prior model specifically for tunnel fires to replace the fixed physical parameters of offline fitting. It solves the dynamic physical quantities of smoke spread length, critical smoke exhaust volume and smoke backflow risk coefficient in real time, thereby generating dynamic physical boundary constraints that evolve with the fire situation.

[0117] According to an embodiment of the present invention, step S2 includes:

[0118] S21. Construct the governing equations for smoke flow in tunnel fires, serving as physical constraints for the lightweight PINN online physical prior model. These equations include the continuity equation, momentum conservation equation, energy conservation equation, and smoke component transport equation, expressed as follows:

[0119] ;

[0120] ;

[0121] ;

[0122] ;

[0123] In the formula, ρ is air density, u is velocity vector, p is pressure, g is gravitational acceleration, τ is viscous stress tensor, h is specific enthalpy, k is thermal conductivity coefficient, T is temperature, and qˊ c Y represents the heat release rate of combustion. i Let D be the mass fraction of the i-th component. i ωˊ is the diffusion coefficient of the component. i The component formation rate;

[0124] S22, a lightweight PINN network architecture is constructed, employing a fully connected neural network with 3 hidden layers and 64 neurons per layer. Channel pruning and quantization techniques are introduced to compress the number of model parameters to less than 10% of the original model, ensuring a single-step inference time of less than 10ms. The network input consists of the current fire source power, ventilation volume, and core measurement point monitoring data, and the output is the smoke spread length L. phy (t), Critical exhaust air volume V cri (t) and flue gas recirculation risk coefficient R back (t);

[0125] S23. Design a hybrid loss function for the lightweight PINN online physics prior model, which is a weighted sum of data fitting loss and physics constraint loss, expressed as follows:

[0126] ;

[0127] In the formula, L data L represents the mean square error loss between the model output and the FDS simulation data. phyTo control the residual loss of the equation, This is the weighting coefficient for data loss, with a value ranging from 0.3 to 0.7. This is the weighting coefficient for the total physical constraint loss, with a value ranging from 0.5 to 1.0; It should be noted that, and Two coefficients are used in the hybrid loss function of the lightweight PINN online physics prior model. Their role is to balance the constraints of fitting simulation data and satisfying the residuals of the governing equations during model training. and The values ​​of 0.3~0.7 and 0.5~1.0 are empirical ranges obtained through simulation trial and error under various typical tunnel fire conditions, ensuring that the magnitudes of the two losses are comparable and avoiding one of them dominating the training process.

[0128] S24, an online update mechanism for a lightweight PINN online physical prior model is constructed, which fine-tunes the model every 10 seconds using the latest monitoring data to correct model parameters to adapt to the actual fire evolution; when the flue gas backflow risk coefficient R back When (t)>0.7, the model's emergency update mode is triggered, shortening the update interval to 2s;

[0129] S25, based on the dynamic physical quantities output by the lightweight PINN online physical prior model, generates dynamic upper and lower limit constraints for the wind turbine volumetric flow rate, expressed as follows:

[0130] ;

[0131] ;

[0132] In the formula, V rated The rated maximum volumetric flow rate of the fan is taken as 40 m³ / s; Q(t) is the power of the fire source at time t; k1 and k2 are the tunnel ventilation characteristic coefficients, which are adaptively adjusted online by the lightweight PINN online physical prior model.

[0133] S3 constructs a PINN-LSTM multi-source time-series prediction module, which integrates temperature, CO concentration, visibility and wind speed data to predict changes in environmental parameters in the core control area in the future period, and generates feedforward control prediction values ​​accordingly.

[0134] According to an embodiment of the present invention, step S3 includes:

[0135] S31, construct a multi-source data attention fusion layer, and use a multi-head self-attention mechanism to automatically learn the weights of four types of data: temperature, CO concentration, visibility, and wind speed, highlighting features that have a greater impact on smoke spread. The fused feature expression is as follows:

[0136] ;

[0137] In the formula, F i (t) represents the normalized feature of the data from the i-th sensor, α i (t) represents the attention weight for the i-th feature;

[0138] S32, construct an LSTM time series prediction network. The input is the fused feature sequence of the past 20 seconds, and the output is the predicted values ​​of temperature, CO concentration and visibility of the core control measurement points at two time steps of 5 seconds and 10 seconds. The LSTM network contains 2 LSTM layers and 1 fully connected output layer, with 128 hidden layer neurons.

[0139] S33, Design a PINN-guided prediction correction mechanism, using the physical priors output by the lightweight PINN online physical prior model as constraints for the LSTM measurement, to correct the physical rationality of the prediction results. The expression for the corrected prediction value is as follows:

[0140] ;

[0141] In the formula, yˊ LSTM (t+τ) represents the original prediction value of the LSTM network, y PINN (t+τ) is the extrapolation value of the lightweight PINN online physics prior model based on physical laws, and β is the correction coefficient, which ranges from 0.6 to 0.9.

[0142] S34: Construct a prediction error assessment and early warning mechanism, calculate the mean square error between the predicted value and the actual monitored value in real time, and when the prediction error exceeds the threshold for three consecutive time steps, automatically reduce the prediction step size to 3s and trigger an emergency update of the lightweight PINN online physical prior model.

[0143] S4. A two-layer hybrid PID controller is constructed, which includes a predictive feedforward layer and an adaptive feedback layer. The predictive feedforward layer offsets foreseeable flue gas disturbances in advance based on the prediction results, and the adaptive feedback layer eliminates residual errors based on real-time monitoring data. The PID parameters are adaptively tuned online by a deep neural network.

[0144] According to an embodiment of the present invention, step S4 includes:

[0145] S41, Construct a control error system, and define the real-time control error e(t) and the predictive control error eˊ(t+τ) as follows:

[0146] ;

[0147] ;

[0148] In the formula, T d(t) represents the real-time temperature of the core control measuring point, Tˊ d (t+τ) represents the predicted temperature after time τ, where T is the temperature. t (t) represents the dynamic target control temperature, which is adaptively adjusted according to the personnel evacuation status and fire level, with a base value of 100℃.

[0149] S42, Design the feedforward control layer, and calculate the feedforward control quantity u based on the predictive control error. ff (t), the fan air volume is adjusted in advance to counteract future flue gas disturbances, and the expression is as follows:

[0150] ;

[0151] In the formula, K ff The feedforward gain coefficient is adaptively adjusted online by the deep neural network.

[0152] S43, an adaptive feedback control layer is designed, employing an improved discrete PID algorithm, integrating derivative-first and integral anti-saturation modules, with the feedback control quantity u... fb The expression for (t) is as follows:

[0153] ;

[0154] In the formula, n is the current calculation time step number, and K p (t), K i (t), K d (t) represents the proportional, integral, and differential coefficients of the current step adaptive tuning;

[0155] S44. Construct an online PID parameter tuner based on a fully connected neural network. The inputs are the control error e(t) at the current moment, the error change rate Δe(t) = e(t) - e(t-1), the fire source power Q(t), and the flue gas backflow risk coefficient R. back (t), the output is the adjustment amount ΔK of the PID parameters. p ΔK i ΔK d The parameter tuner contains three hidden layers, each with 32 neurons, and uses the ReLU activation function.

[0156] S45, the total control quantity is the weighted sum of the feedforward control quantity and the feedback control quantity, and is output after correction by the dynamic physical boundary constraints generated in step S42, as shown in the following expression:

[0157] ;

[0158] In the formula, γ is the feedforward-feedback weight coefficient, which is 0.7 when the prediction error is less than the threshold and 0.3 when the prediction error is greater than the threshold, so as to give priority to ensuring the reliability of feedback control.

[0159] S5 constructs a multi-source sensor data fusion preprocessing and active anti-interference compensation module to achieve data spatiotemporal alignment, noise reduction, anomaly repair, and fault pre-tolerance.

[0160] According to an embodiment of the present invention, step S5 includes:

[0161] S51 performs spatiotemporal alignment processing on multi-source sensor data, unifying sensor data with different sampling frequencies and spatial locations into a 0.5s time step and the tunnel longitudinal coordinate system.

[0162] S52 employs an improved first-order exponential moving average method to denoise the data. The smoothing coefficient α is adaptively adjusted according to the degree of data fluctuation. It is set to 0.3 when the data fluctuation exceeds a preset threshold (the preset threshold can be calibrated according to the actual situation), and to 0.5 when the data is stable. The denoising formula is as follows:

[0163] ;

[0164] In the formula, x(t) represents the raw sensor data. The data after noise reduction;

[0165] S53 identifies outliers based on the improved 3σ criterion and the isolated forest algorithm. For single-point outliers, the weighted average of the data from the preceding and following 5 frames is used to replace them. For outliers in more than 5 consecutive frames, a sensor fault warning is triggered 10 seconds in advance, and the system automatically switches to backup measurement point data.

[0166] S54 is designed as an active detection and compensation module for sudden changes in operating conditions. Based on the spatiotemporal gradient change rate of multi-source data, it identifies sudden changes in operating conditions such as a sudden increase in fire source power, smoke backflow, and vehicle congestion. When a sudden change in operating conditions is detected, it automatically increases the feedforward control weight and proportional coefficient to shorten the system response time.

[0167] S6 constructs a four-dimensional hybrid control objective function that integrates control error loss, online physical prior constraint loss, airflow smoothing loss, and prediction error loss to complete the initial tuning of controller parameters and verification under multiple extreme conditions.

[0168] According to an embodiment of the present invention, step S6 includes:

[0169] S61, Construct the main loss term for control error, and use the sum of squared errors index to measure control accuracy, as shown in the following expression:

[0170] ;

[0171] In the formula, N is the total simulation time steps;

[0172] S62, construct an online physics prior constraint loss term to penalize control quantities that exceed the dynamic physics boundaries generated by the lightweight PINN online physics prior model, as shown in the following expression:

[0173] ;

[0174] In the formula, V(t) is the volumetric flow rate control value of the fan at time t, and λ phy This is the weighting coefficient for the physical constraint loss, with a value ranging from 0.8 to 1.2; it should be noted that... λ phy The online physical prior constraint loss term is used in the four-dimensional hybrid objective function of the controller. Its function is to penalize the behavior of the fan airflow exceeding the dynamic boundary generated by PINN. The value is 0.8~1.2, which comes from the sensitivity analysis in the controller parameter tuning and optimization process.

[0175] S63, constructs a smoothing loss term for airflow to suppress drastic fluctuations in fan airflow and avoid frequent start-ups and shutdowns and damage to the equipment. The expression is as follows:

[0176] ;

[0177] In the formula, λ smooth The smoothing constraint weight coefficient is set to a value between 0.005 and 0.015.

[0178] S64, Construct the prediction error loss term to measure the accuracy of the time series prediction module, with the following expression:

[0179] ;

[0180] In the formula, λ pred The prediction error weighting coefficient has a value of 0.3 to 0.7.

[0181] S65, the overall control objective function is the weighted sum of all loss terms, expressed as follows:

[0182] ;

[0183] S66, based on the Adam gradient descent algorithm and multi-condition simulation data, takes minimizing the overall control objective function as the optimization objective and completes the initial training of the PID parameter tuner and timing prediction module. Through batch simulations of 7 fire source locations, 10 fire source powers, 15 wind speed conditions and 5 extreme disturbance conditions, the generalization verification and parameter optimization of the controller are completed.

[0184] S7, based on the tuned hybrid PID controller, performs predictive-feedback closed-loop intelligent control of the volumetric flow rate of the tunnel fire smoke exhaust fan through the Python-FDS bidirectional coupling interface, and outputs the fan air volume timing control command.

[0185] According to an embodiment of the present invention, step S7 includes:

[0186] S71, load the tuned parameters of the hybrid PID controller, the parameters of the lightweight PINN online physical prior model, the parameters of the LSTM time series prediction model, and the data preprocessing parameters;

[0187] S72 extracts real-time monitoring data from all composite sensors in the tunnel in 0.5s time steps through the Python-FDS bidirectional coupling interface. After being processed by the multi-source data preprocessing module, the data is input into the PINN online physical prior model and the LSTM time series prediction module, respectively.

[0188] S73, a lightweight PINN online physical prior model, calculates dynamic physical quantities and physical boundary constraints in real time; the LSTM module predicts environmental parameter changes in the next 5-10 seconds; the hybrid PID controller combines feedforward control and feedback control to calculate the output target fan volume flow rate and exhaust port opening.

[0189] S74 updates the target control quantity to the corresponding boundary of the FDS numerical model in real time through the coupling interface, completes the CFD calculation and control iteration of the next time step, and realizes the prediction-feedback closed-loop control of the fan air volume.

[0190] The S75 stores multi-source measurement point timing data, fan operation data, PID parameter adjustment data, lightweight PINN online physical prior model output data and prediction error data in real time during the control process, generates control effect analysis reports and visualization curves, and outputs them synchronously to the tunnel fire emergency prevention and control platform.

[0191] In summary, the adaptive PID intelligent control method for tunnel fire smoke exhaust fans, incorporating lightweight PINN, firstly, establishes a tunnel fire FDS numerical model that matches the actual engineering conditions, integrates multi-source composite sensors and dynamic ventilation boundaries (such as continuously adjustable smoke exhaust port opening), and constructs a Python-FDS bidirectional coupling interface to achieve millisecond-level data interaction and synchronous iteration between CFD numerical calculation and the control algorithm at each time step. Based on this, a lightweight physical information neural network online physical prior model is constructed, embedding the control equations for smoke flow as physical constraints into the network to solve dynamic physical quantities such as smoke spread length and critical smoke exhaust volume in real time. This generates dynamic physical boundary constraints such as upper and lower limits of airflow that adaptively adjust with the fire situation, replacing traditional fixed parameters. Simultaneously, a PINN-LSTM multi-source time-series prediction module is constructed, utilizing a multi-head self-attention mechanism to fuse multi-source sensor data such as temperature and CO concentration to predict environmental changes in the core area in the next few seconds. The physical extrapolation value of PINN is then introduced to correct the prediction results, generating a physically reasonable feedforward control estimate. Furthermore, a dual-layer hybrid PID controller was designed. Its predictive feedforward layer acts in advance based on the predicted value to offset foreseeable smoke disturbances, while the adaptive feedback layer compensates for real-time monitoring errors. The PID parameters are tuned online by an independent deep neural network based on the fire situation. Throughout the process, multi-source data undergoes preprocessing such as spatiotemporal alignment, adaptive denoising, and anomaly repair before being fed into the control loop. The system parameters are initially tuned offline and verified under multiple operating conditions using a four-dimensional hybrid objective function that integrates control error, physical boundary violation penalties, airflow smoothness, and prediction accuracy, and a backup control mechanism is included. Finally, the tuned controller calculates the optimal fan volumetric flow rate command based on real-time sensing and prediction data through a bidirectional coupling interface and executes it in a closed loop, achieving precise, rapid, and physically consistent predictive-feedback intelligent control of smoke extraction in tunnel fires. This invention can reduce the adjustment time of smoke exhaust control in tunnel fires by more than 55%, reduce the overshoot by more than 75%, and shorten the system response lag time by more than 60%. While maintaining millisecond-level real-time response, it significantly improves the control stability and physical rationality under complex fire conditions, providing more advanced control technology support for smoke control and emergency rescue in tunnel fires.

[0192] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0193] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0194] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0195] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for adaptive PID intelligent control of tunnel fire smoke exhaust fans integrating lightweight PINN, characterized in that, Includes the following steps: S1. Construct a multi-source sensor fusion FDS numerical model for tunnel fire with dynamic ventilation boundary, and build a Python-FDS bidirectional coupling interface for millisecond-level synchronous iteration of control algorithm and CFD numerical calculation. S2. Construct a lightweight PINN online physical prior model specifically for tunnel fires to replace the fixed physical parameters of offline fitting. Solve dynamic physical quantities such as smoke spread length, critical smoke exhaust volume and smoke backflow risk coefficient in real time, and then generate dynamic physical boundary constraints that evolve with the fire situation. S3, construct a PINN-LSTM multi-source time series prediction module, integrate temperature, CO concentration, visibility and wind speed data to predict the changes in environmental parameters of the core control area in the future period, and generate feedforward control prediction values ​​accordingly. S4. Construct a dual-layer hybrid PID controller comprising a predictive feedforward layer and an adaptive feedback layer. The predictive feedforward layer preemptively offsets foreseeable flue gas disturbances based on the prediction results, and the adaptive feedback layer eliminates residual errors based on real-time monitoring data. The PID parameters are adaptively tuned online by a deep neural network. S5 constructs a multi-source sensor data fusion preprocessing and active anti-interference compensation module to achieve data spatiotemporal alignment, noise reduction, anomaly repair and fault pre-tolerance; S6, construct a four-dimensional hybrid control objective function that integrates control error loss, online physical prior constraint loss, airflow smoothing loss, and prediction error loss to complete the initial tuning of controller parameters and verification under multiple extreme conditions. Step S6 includes: S61, Construct the main loss term for control error, and use the sum of squared errors index to measure control accuracy, as shown in the following expression: ； In the formula, N is the total simulation time steps; S62, construct an online physics prior constraint loss term to penalize control quantities that exceed the dynamic physics boundaries generated by the lightweight PINN online physics prior model, as shown in the following expression: ； In the formula, V(t) is the volumetric flow rate control value of the fan at time t, and λ phy This is the weighting coefficient for the physical constraint loss, with a value ranging from 0.8 to 1.2; S63, constructs a smoothing loss term for airflow to suppress drastic fluctuations in fan airflow and avoid frequent start-ups and shutdowns and damage to the equipment. The expression is as follows: ； In the formula, λ smooth The smoothing constraint weight coefficient is set to a value between 0.005 and 0.

015. S64, Construct the prediction error loss term to measure the accuracy of the time series prediction module, with the following expression: ； In the formula, λ pred The prediction error weighting coefficient has a value of 0.3 to 0.

7. S65, the overall control objective function is the weighted sum of all loss terms, expressed as follows: ； S66, based on the Adam gradient descent algorithm and multi-condition simulation data, takes minimizing the overall control objective function as the optimization objective and completes the initial training of the PID parameter tuner and timing prediction module; through batch simulation of 7 fire source locations, 10 fire source powers, 15 wind speed conditions and 5 extreme disturbance conditions, the generalization verification and parameter optimization of the controller are completed. S7, based on the tuned hybrid PID controller, executes the prediction-feedback closed-loop intelligent control of the volumetric flow rate of the tunnel fire smoke exhaust fan through the Python-FDS bidirectional coupling interface, and outputs the fan air volume timing control command.

2. The adaptive PID intelligent control method for tunnel fire smoke exhaust fans integrating lightweight PINN as described in claim 1, characterized in that, Step S1 includes: S11. Construct a tunnel fire FDS numerical model that matches the actual engineering situation, and determine the number and corresponding location of fire sources, the configuration range of fire source power, and the configuration parameters of longitudinal ventilation in the FDS numerical model. S12, determine the range of the fire source, the power growth of the fire source follows the t² fire model, the total simulation time is the first preset time; the fuel types are gasoline, diesel, and plastic, and the corresponding combustion product generation rate and heat release rate curves are set respectively; the tunnel wall material is defined as concrete, and the wall heat accumulation effect and heat radiation parameters are introduced. S13, a dynamic grid partitioning strategy is used to partition the tunnel space; S14, the ventilation system includes two sets of jet fans, symmetrically arranged in the tunnel. Each set of jet fans contains 2 fans. The entire cross-section of the fan air supply is defined as a VENT type boundary. A new axial flow smoke exhaust fan and smoke exhaust outlet linkage control interface are added. The smoke exhaust outlet opening degree supports continuous adjustment from 0 to 100%. S15, along the longitudinal centerline of the tunnel, multiple composite sensors are deployed at preset intervals. The composite sensors integrate four monitoring functions: temperature, CO concentration, visibility, and wind speed. Multiple core control measuring points are set up upstream of the fire source, and multiple backup monitoring measuring points are set up upstream and downstream of the fire source. S16 uses an asynchronous Python script to build a bidirectional coupling interface between the FDS and the control algorithm, completing three synchronization steps in each 0.5s calculation time step: extracting multi-source real-time monitoring data from all sensors in the FDS numerical model; updating the target fan volume flow rate and smoke exhaust port opening output by the hybrid PID controller to the corresponding VENT boundary parameters in real time; synchronizing the fire evolution data to the digital twin buffer; and setting hard constraints on air volume (0~40m³ / s) and smoke exhaust port opening (0~100%).

3. The adaptive PID intelligent control method for tunnel fire smoke exhaust fans integrating lightweight PINN as described in claim 1, characterized in that, Step S2 includes: S21. Construct the governing equations for smoke flow in tunnel fires, serving as physical constraints for the lightweight PINN online physical prior model. These equations include the continuity equation, momentum conservation equation, energy conservation equation, and smoke component transport equation, expressed as follows: ； ； ； ； In the formula, ρ is air density, u is velocity vector, p is pressure, g is gravitational acceleration, τ is viscous stress tensor, h is specific enthalpy, k is thermal conductivity coefficient, T is temperature, and qˊ c Y represents the heat release rate of combustion. i Let D be the mass fraction of the i-th component. i Let ω be the diffusion coefficient of the component. i The component formation rate; S22, a lightweight PINN network architecture is constructed, employing a fully connected neural network with 3 hidden layers and 64 neurons per layer. Channel pruning and quantization techniques are introduced to compress the number of model parameters to less than 10% of the original model, ensuring a single-step inference time of less than 10ms. The network input consists of the current fire source power, ventilation volume, and core measurement point monitoring data, and the output is the smoke spread length L. phy (t), Critical exhaust air volume V cri (t) and flue gas recirculation risk coefficient R back (t); S23. Design a hybrid loss function for the lightweight PINN online physics prior model, which is a weighted sum of data fitting loss and physics constraint loss, expressed as follows: ； In the formula, L data To account for the mean square error loss between the model output and the FDS simulation data, L phy To control the residual loss of the equation, This is the weighting coefficient for data loss, with a value ranging from 0.3 to 0.7; This is the weighting coefficient for the total physical constraint loss, with a value ranging from 0.5 to 1.0; S24, an online update mechanism for a lightweight PINN online physical prior model is constructed, which fine-tunes the model every 10 seconds using the latest monitoring data to correct model parameters to adapt to the actual fire evolution; when the flue gas backflow risk coefficient R back When (t)>0.7, the model's emergency update mode is triggered, shortening the update interval to 2s; S25, based on the dynamic physical quantities output by the lightweight PINN online physical prior model, generates dynamic upper and lower limit constraints for the wind turbine volumetric flow rate, expressed as follows: ； ； In the formula, V rated The rated maximum volumetric flow rate of the fan is taken as 40 m³ / s; Q(t) is the power of the fire source at time t; k1 and k2 are the tunnel ventilation characteristic coefficients, which are adaptively adjusted online by the lightweight PINN online physical prior model.

4. The adaptive PID intelligent control method for tunnel fire smoke exhaust fans integrating lightweight PINN as described in claim 1, characterized in that, Step S3 includes: S31, construct a multi-source data attention fusion layer, and use a multi-head self-attention mechanism to automatically learn the weights of four types of data: temperature, CO concentration, visibility, and wind speed, highlighting features that have a greater impact on smoke spread. The fused feature expression is as follows: ； In the formula, F i (t) represents the normalized feature of the data from the i-th sensor, α i (t) represents the attention weight for the i-th feature; S32, construct an LSTM time series prediction network. The input is the fused feature sequence of the past 20 seconds, and the output is the predicted values ​​of temperature, CO concentration and visibility of the core control measurement points at two time steps of 5 seconds and 10 seconds. The LSTM network contains 2 LSTM layers and 1 fully connected output layer, with 128 hidden layer neurons. S33, Design a PINN-guided prediction correction mechanism, using the physical priors output by the lightweight PINN online physical prior model as constraints for the LSTM measurement, to correct the physical rationality of the prediction results. The expression for the corrected prediction value is as follows: ； In the formula, yˊ LSTM (t+τ) represents the original prediction value of the LSTM network, y PINN (t+τ) is the extrapolation value of the lightweight PINN online physics prior model based on physical laws, and β is the correction coefficient, which ranges from 0.6 to 0.9; S34: Construct a prediction error assessment and early warning mechanism, calculate the mean square error between the predicted value and the actual monitored value in real time, and when the prediction error exceeds the threshold for three consecutive time steps, automatically reduce the prediction step size to 3s and trigger an emergency update of the lightweight PINN online physical prior model.

5. The adaptive PID intelligent control method for tunnel fire smoke exhaust fans integrating lightweight PINN as described in claim 1, characterized in that, Step S4 includes: S41, Construct a control error system, and define the real-time control error e(t) and the predictive control error eˊ(t+τ) as follows: ； ； In the formula, T d (t) represents the real-time temperature of the core control measuring point, Tˊ d (t+τ) represents the predicted temperature after time τ, where T is the temperature. t (t) represents the dynamic target control temperature, which is adaptively adjusted according to the personnel evacuation status and fire level, with a base value of 100℃. S42, Design the feedforward control layer, and calculate the feedforward control quantity u based on the predictive control error. ff (t), the fan air volume is adjusted in advance to counteract future flue gas disturbances, and the expression is as follows: ； In the formula, K ff The feedforward gain coefficient is adaptively adjusted online by the deep neural network. S43, an adaptive feedback control layer is designed, employing an improved discrete PID algorithm, integrating derivative-first and integral anti-saturation modules, with the feedback control quantity u... fb The expression for (t) is as follows: ； In the formula, n is the current calculation time step number, and K p (t), K i (t), K d (t) represents the proportional, integral, and differential coefficients of the current step adaptive tuning; S44. Construct an online PID parameter tuner based on a fully connected neural network. The inputs are the control error e(t) at the current moment, the error change rate Δe(t) = e(t) − e(t−1), the fire source power Q(t), and the flue gas backflow risk coefficient R. back (t), the output is the adjustment amount ΔK of the PID parameters. p ΔK i ΔK d The parameter tuner contains three hidden layers, each with 32 neurons, and uses the ReLU activation function. S45, the total control quantity is the weighted sum of the feedforward control quantity and the feedback control quantity, and is output after correction by the dynamic physical boundary constraints generated in step S42, as shown in the following expression: ； In the formula, γ is the feedforward-feedback weight coefficient, which is 0.7 when the prediction error is less than the threshold and 0.3 when the prediction error is greater than the threshold, so as to give priority to ensuring the reliability of feedback control.

6. The adaptive PID intelligent control method for tunnel fire smoke exhaust fans integrating lightweight PINN as described in claim 1, characterized in that, Step S5 includes: S51 performs spatiotemporal alignment processing on multi-source sensor data, unifying sensor data with different sampling frequencies and spatial locations into a 0.5s time step and the tunnel longitudinal coordinate system. S52 employs an improved first-order exponential moving average method to denoise the data. The smoothing coefficient α is adaptively adjusted according to the degree of data fluctuation, taking 0.3 when the data fluctuation exceeds a preset threshold and 0.5 when the data is stable. The denoising formula is as follows: ； In the formula, x(t) is the original sensor data, and x′(t) is the data after noise reduction; S53 identifies outliers based on the improved 3σ criterion and the isolated forest algorithm. For single-point outliers, the weighted average of the data from the preceding and following 5 frames is used to replace them. For outliers in more than 5 consecutive frames, a sensor fault warning is triggered 10 seconds in advance, and the system automatically switches to backup measurement point data. S54 is designed as an active detection and compensation module for sudden changes in operating conditions. Based on the spatiotemporal gradient change rate of multi-source data, it identifies sudden changes in operating conditions such as a sudden increase in fire source power, smoke backflow, and vehicle congestion. When a sudden change in operating conditions is detected, it automatically increases the feedforward control weight and proportional coefficient to shorten the system response time.

7. The adaptive PID intelligent control method for tunnel fire smoke exhaust fans integrating lightweight PINN as described in claim 1, characterized in that, Step S7 includes: S71, load the tuned parameters of the hybrid PID controller, the parameters of the lightweight PINN online physical prior model, the parameters of the LSTM time series prediction model, and the data preprocessing parameters; S72 extracts real-time monitoring data from all composite sensors in the tunnel in 0.5s time steps through the Python-FDS bidirectional coupling interface. After being processed by the multi-source data preprocessing module, the data is input into the PINN online physical prior model and the LSTM time series prediction module, respectively. S73, a lightweight PINN online physical prior model, calculates dynamic physical quantities and physical boundary constraints in real time; the LSTM module predicts environmental parameter changes in the next 5-10 seconds; the hybrid PID controller combines feedforward control and feedback control to calculate the output target fan volume flow rate and exhaust port opening. S74 updates the target control quantity to the corresponding boundary of the FDS numerical model in real time through the coupling interface, completes the CFD calculation and control iteration of the next time step, and realizes the prediction-feedback closed-loop control of the fan air volume. The S75 stores multi-source measurement point timing data, fan operation data, PID parameter adjustment data, lightweight PINN online physical prior model output data and prediction error data in real time during the control process, generates control effect analysis reports and visualization curves, and outputs them synchronously to the tunnel fire emergency prevention and control platform.

Images