Neural network-based bridge deck watering automatic control method and system

Abstract

The invention relates to the field of neural networks, in particular to a bridge deck watering automatic control method and system based on a neural network. According to the method, micrometeorological data, bridge deck thermal state data and historical watering feedback of a bridge site site are collected in real time, the theoretical evaporation capacity of the current atmosphere is calculated based on the thermodynamic principle, and net flux characteristics are constructed. And inputting the processed data into a pre-trained time sequence neural network model to obtain a basic flow prediction value. And then constructing a physical optimization model containing energy balance constraints, and calculating a physical correction gradient to carry out directional correction on a predicted value. Meanwhile, an environment perception hysteresis factor is introduced, the bridge deck moisture dissipation rate is calculated according to the real-time wind speed and humidity, and the dynamic ponding risk probability is solved. And finally, competitive gating weighting is performed on the corrected flow by using the risk probability, a watering instruction is generated, and an execution mechanism is driven, so that the targets of bridge floor cooling and ponding prevention are collaboratively optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of neural networks, in particular to a bridge deck sprinkling automatic control method and system based on neural networks. BACKGROUND

[0002] Under high temperature or extreme weather conditions in summer, the bridge deck of concrete or steel bridge is prone to disease due to thermal expansion effect, and even affects the structural safety. Timed sprinkling evaporation cooling is an economical and effective active cooling means. However, the existing bridge deck sprinkling control method mainly has the following defects: The fixed temperature threshold-based on-off control or simple timing control is generally used, and the real-time changes of the micro-meteorological environment such as wind speed, air humidity, and solar radiation are completely ignored, which has a decisive influence on the evaporation cooling efficiency and water residue time. Under low humidity and high wind conditions, the sprinkling is insufficient, and the cooling potential is wasted; under high humidity and calm wind conditions, the sprinkling is excessive, which causes the bridge deck to be wet and slippery, and forms a safety hazard.

[0003] The simple superposition or fixed priority processing is usually used between effective cooling and water accumulation prevention, which cannot realize the real-time optimal trade-off between the two in the dynamically changing environment.

[0004] Even the pure data-driven predictive control method appeared in recent years can learn complex patterns, but the output instructions may seriously violate the basic physical laws, resulting in invalid or high-energy consumption control instructions, and the engineering application reliability is low.

[0005] Therefore, an automatic control method capable of deeply integrating real-time environment perception, physical law constraint and intelligent learning decision-making is urgently needed, so as to cooperatively optimize the cooling efficiency and operation safety. SUMMARY

[0006] In order to improve the problems of unstable sprinkling instruction prediction model, lack of environment perception, and unreasonable prediction results of the sprinkling instruction prediction model, the present application provides a bridge deck sprinkling automatic control method and system based on neural networks.

[0007] In the first aspect, the present application provides a bridge deck sprinkling automatic control method based on neural networks, which adopts the following technical scheme: Real-time collection of micro-meteorological environment data and bridge deck thermal state data at the bridge site, and acquisition of sprinkling flow feedback in the last control period; Based on the thermodynamic energy conservation and mass transfer principle, the theoretical evaporation capacity of the current atmospheric environment is calculated, and the thermodynamic net flux feature is constructed in combination with the sprinkling flow feedback; The micro-meteorological environment data, bridge deck thermal state data and sprinkling flow feedback after standardization are input into a neural network model with time sequence memory ability which is trained in advance, to generate a basic flow prediction value at the current time. Construct a physical optimization model that includes energy balance constraints, and calculate the partial derivative of the physical loss function with respect to the flow control variable as the physical correction gradient; An environmentally-aware dynamic hysteresis factor is introduced to calculate the bridge surface moisture dissipation rate based on real-time wind speed and humidity characteristics in order to solve the dynamic water accumulation risk probability. Based on the physical correction gradient, the predicted basic flow rate is corrected in a direction that conforms to the law of energy conservation. The corrected flow rate is then subjected to competitive gating weighting using the dynamic water accumulation risk probability to generate the final water spraying command to drive the actuator.

[0008] Preferably, the introduction of an environment-aware dynamic hysteresis factor to calculate the bridge deck moisture dissipation rate based on real-time wind speed and humidity characteristics to solve the dynamic water accumulation risk probability specifically includes: constructing an environment-aware dynamic hysteresis factor characterizing the moisture dissipation memory length; establishing a nonlinear mapping relationship between the dynamic hysteresis factor and real-time wind speed and saturated vapor pressure difference, wherein the dynamic hysteresis factor includes a structural gravity drainage baseline term and an environment-sensitive term determined by wind speed and vapor pressure difference; updating the current water accumulation depth state variable based on the dynamic hysteresis factor using a discrete-time analytical solver of a non-homogeneous linear differential equation; and mapping the water accumulation depth state variable to the normalized dynamic water accumulation risk probability through a modified probability activation function.

[0009] Preferably, updating the current water depth state variable using the discrete-time analytical solver of the non-homogeneous linear differential equation specifically refers to: calculating the water depth at the current moment using the discrete-time analytical solver. This calculation process considers the exponential decay term of the water depth at the previous moment and the input-output term contributed by the thermodynamic net flux characteristics. The analytical solver is used to eliminate the accumulated error of the Euler integral method in long-term iterations, ensuring that the evolution of the water depth follows the law of conservation of physical mass.

[0010] Preferably, the construction of a physical optimization model including energy balance constraints, and the calculation of the partial derivative of the physical loss function with respect to the flow control variable as the physical correction gradient, specifically includes: based on the full physical energy balance equation, using the current bridge deck thermal state data and the basic flow prediction value, predicting the theoretical bridge deck surface temperature at the next moment; establishing a physical loss function constrained by the target safe temperature of the road surface, the physical loss function including a temperature tracking error term and a control energy consumption penalty term; applying the chain rule to calculate the partial derivative of the physical loss function with respect to the dimensionless flow rate; introducing a phase transition limit constraint when calculating the partial derivative, forcibly truncating the gradient contribution of the evaporation cooling term to the flow rate when the current total available water flux of the system exceeds the theoretical evaporation capacity.

[0011] Preferably, the competitive gating weighting of the corrected flow rate using the dynamic water accumulation risk probability specifically includes: constructing a weighting coefficient based on a competition mechanism between safety risk and cooling benefit; when the dynamic water accumulation risk probability is lower than the safety benchmark threshold, the weighting coefficient remains in a fully open state; when the dynamic water accumulation risk probability exceeds the safety benchmark threshold, the weighting coefficient exhibits a hyperbolic tangent characteristic and nonlinear decay; multiplying the flow rate value adjusted by the physical correction gradient by the weighting coefficient to obtain the final watering command.

[0012] Preferably, the calculation of the theoretical evaporation carrying capacity of the current atmospheric environment specifically includes: calculation based on the modified Penmanmont formula principle, combined with net radiation intensity, soil heat flux, air density, specific heat at constant pressure and saturated water vapor pressure difference; introducing aerodynamic conductivity based on real-time wind speed calculation, and uniformly converting the calculation results into liquid water flux velocity per unit area as the physical upper limit for determining whether water can be accepted by the atmosphere.

[0013] Preferably, the method further includes a cold start data filling strategy: the system maintains a circular buffer for storing historical time-series data; during the system initialization phase, if the data in the circular buffer is empty, the sensor data at the current moment and the initialized flow feedback are collected, copied and filled into the entire circular buffer to eliminate the distribution bias introduced by zero filling; during the training phase of the neural network model, the sprinkler flow feedback feature is constructed using the label value of the previous moment through autoregression.

[0014] Preferably, the neural network model adopts a long short-term memory network architecture: the neural network model includes two stacked long short-term memory layers; the first layer is configured to return the complete time series to retain micro-meteorological fluctuation characteristics, and the second layer is configured to return the hidden state vector of the final time step to extract thermodynamic trend characteristics; the output layer of the model is mapped to a non-negative space through a non-negative activation function to generate the dimensionless basic flow prediction value.

[0015] Preferably, the step of performing energy-conserving directional correction on the basic flow prediction value based on the physical correction gradient specifically refers to: subtracting the product of the gradient descent step size coefficient and the physical correction gradient from the basic flow prediction value to obtain the corrected flow; the correction is used to constrain the data-driven prediction result to a physical energy-conserving manifold.

[0016] Secondly, the present invention provides an automated bridge deck watering control system based on a neural network, comprising: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned automated bridge deck watering control method is implemented.

[0017] The present invention has the following technical effects: By calculating the theoretical evaporation capacity in real time, a physical upper limit benchmark for sprinkler control is provided, which dynamically changes with wind speed, humidity, and radiation. This improves the ability to perceive the environment, makes the decision on the amount of sprinkler water conform to physical constraints, and avoids blind sprinkler watering.

[0018] By introducing a physical optimization model and calculating the physical correction gradient, the initial prediction of the neural network is directionally guided and corrected, forcing the data-driven prediction results to be constrained within the physical model, thereby improving the reliability of the watering command.

[0019] Based on the dynamic water accumulation risk probability and competitive gating weighting mechanism, the instantaneous water balance is transformed into a cumulative risk measure. It can make an optimal trade-off between maximizing cooling and minimizing water accumulation risk according to the real-time risk level. When the risk is low, it fully guarantees the cooling efficiency, and when the risk is high, it automatically transitions to ensure safety, thereby improving the reliability of water spraying commands.

[0020] By employing a cold-start data filling strategy, the performance fluctuation of the time-series prediction model due to the lack of historical data in the early stages of system startup is effectively solved, thereby improving the reliability of sprinkler commands. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating steps S1 to S6 of an automated control method for bridge deck watering based on neural networks according to the present invention. Detailed Implementation

[0022] This invention discloses an automated control method for bridge deck watering based on neural networks, referring to... Figure 1 ,include: Step S1: Collect real-time micro-meteorological environment data and bridge deck thermal status data at the bridge site, and obtain water spraying flow feedback from the previous control cycle.

[0023] In bridge deck sprinkler systems, especially during cold starts, models that rely on historical data cannot function because the buffer is empty. Simple zero-value filling can introduce severe distribution bias and impair the model's predictive ability.

[0024] To address the issues of data acquisition integrity and cold start reliability, a robust data awareness and buffering mechanism is established to ensure that effective input tensors can be obtained under any initial state.

[0025] A calibrated high-precision sensor array was deployed at key cross-sections of the bridge, specifically including a PT100 platinum resistance temperature sensor installed in a standard Stevenson screen for measuring air temperature. (Unit: degrees Celsius), relative humidity is measured using a capacitive polymer thin-film humidity sensor. (Unit: percentage) An ultrasonic anemometer was installed at a standard meteorological observation height of 2.0 meters above the bridge deck to collect wind speed data. (Unit: meters per second) The installation height of the wind vane can be adjusted by the implementer according to the specific implementation scenario. The total radiation meter, which uses the thermopile principle, is installed horizontally to measure the total solar radiation intensity. (Unit: Watts per square meter) A non-contact infrared thermometer is used to vertically aim at the surface of the asphalt pavement layer on the bridge deck to monitor its surface temperature in real time. (Unit: degrees Celsius), and an electromagnetic flow meter is installed on the branch pipe of the sprinkler pipeline to read the average sprinkler flow rate actually output by the actuator in the previous control cycle. (Unit: cubic meters per second) ).

[0026] Set a fixed control and sampling time step , This value is based on experience and can be adjusted by the implementer according to the specific implementation scenario. A target road surface safety temperature is preset. This value is based on the softening point of typical SBS modified asphalt and the summer bridge rutting prevention specifications. It can be adjusted by the implementer according to the specific implementation scenario, and a flow normalization benchmark can be defined. Dimensionless for use in subsequent calculations.

[0027] To address the cold start problem, construct a length A first-in, first-out circular buffer is used to store historical time-series data. Upon initial startup, if the buffer is detected to be empty, a current value copy-fill strategy is immediately executed: the current time (…) is copied and filled. All sensor data collected and initial traffic feedback copy The portions are sequentially filled into the buffer to construct a shape of... The initial input tensor is 6, where 6 is the number of dimensions of the collected data, which avoids the problem of possible confusion in subsequent sprinkler control commands due to the singular distribution introduced by zero padding.

[0028] Step S2: Based on the principles of thermodynamic energy conservation and mass transport, calculate the theoretical evaporation carrying capacity of the current atmospheric environment, and construct the thermodynamic net flux characteristics by combining the water spray flow feedback.

[0029] Cooling by sprinkling water involves absorbing heat through the evaporation phase change of water, but the evaporation rate is strictly controlled by the local weather conditions.

[0030] If the amount of water sprayed exceeds the upper limit of atmospheric evaporation, the excess water will not only fail to contribute to the cooling effect, but will also accumulate to form a water film, reducing the anti-skid performance of the road surface and posing a safety hazard.

[0031] Quantifying the theoretical evaporation carrying capacity of the atmospheric environment is crucial for avoiding ineffective watering and controlling the risk of water accumulation. The Penman-Montes formula from micrometeorology can be applied. This formula, based on the theory of energy conservation and turbulent diffusion, decomposes the evaporation process into radiation-driven terms and aerodynamic terms, thereby calculating the maximum evaporation rate that the underlying surface (bridge surface) may reach under given meteorological conditions.

[0032] First, the key meteorological parameters are derived: based on the current air temperature. The Tetens formula was used to calculate the saturated vapor pressure. (Unit: kPa) ), and then combined with relative humidity Calculate actual water vapor pressure and saturated water vapor pressure difference Simultaneously calculate the slope of the saturated water vapor pressure-temperature curve. (unit: ),in The unit is kPa, which is the fitting coefficient of the saturated water vapor pressure formula near the triple point of water (0.01°C). The units are dimensionless and °C, respectively. These are empirical values ​​in the Tetens formula and can be adjusted by the implementer according to the specific implementation scenario; 100 is a unit conversion factor for humidity values; The unit is kPa·℃, and it is a derived constant of the slope of the saturated water vapor pressure curve.

[0033] To standardize the unit system, an air density is preset. Specific heat of air at constant pressure And introduce the unit conversion constant. , used to Convert to

[0034] The specific implementation for calculating the theoretical evaporation carrying capacity of the current atmospheric environment is as follows: First, calculate the net radiation received by the bridge deck. Considering the balance between shortwave and longwave radiation, the formula is: ,in The albedo of the asphalt surface. It is the surface emissivity of asphalt (also its long-wave absorption rate). It is the Stefan-Boltzmann constant. and Absolute temperatures (Kelvin) of air and surface, respectively, and effective emissivity of the sky. From empirical formula Calculation, here The unit remains the same. .

[0035] Secondly, based on measured wind speed Calculate the convective heat transfer coefficient (unit: Then, based on the Lewis analogy, the aerodynamic conductivity is obtained. (unit: ).

[0036] For the surface of a free water film, its surface conductivity is set. For example, a maximum value, such as This is approximated by assuming an infinite supply of surface moisture. Finally, the latent heat of vaporization of water is considered. and water density The theoretical evaporation capacity was calculated using the modified Penman-Monteith formula. Its physical meaning is the maximum water flux velocity (unit: ) that a unit area of ​​bridge deck can theoretically achieve through evaporation under current meteorological and radiation conditions. The calculation formula is: In the formula, This is the constant of the wet / dry meter.

[0037] In obtaining Then, the net thermodynamic flux characteristics can be constructed, specifically by defining the bridge deck area corresponding to a single control unit. ,For example It can be adjusted by the implementer according to the specific implementation scenario, and a dynamic evaporation efficiency coefficient can be introduced. The effect of wind speed on the actual evaporation process is characterized by 1, where 1 is the theoretical upper limit of evaporation efficiency, indicating that the actual evaporation rate cannot exceed the theoretical maximum evaporation capacity determined by energy and aerodynamic conditions. 5 represents the basic evaporation efficiency coefficient under windless conditions. This value is based on experience and can be adjusted by the implementer according to the specific implementation scenario. The coefficient for enhancing evaporation efficiency is denoted as the factor that increases evaporation efficiency by 1 m / s. The increase, in units of This value is based on experience and can be adjusted by the implementer according to the specific implementation scenario.

[0038] Thermodynamic net flux characteristics Calculated by the following formula: The calculation expresses the rate of water flux per unit area supplied in the previous cycle. The actual water flux rate per unit area that can evaporate in the current environment The instantaneous difference between them.

[0039] This means there is a water surplus and a tendency for water accumulation. This indicates that there is still a surplus in evaporation capacity.

[0040] Step S3: Input the standardized micro-meteorological environment data, bridge deck thermal state data, and sprinkler flow feedback into a pre-trained neural network model with time-series memory capabilities to generate the basic flow prediction value for the current moment.

[0041] Bridge deck temperature changes are strongly influenced by transient microclimates (such as gusts of wind and cloud cover). Simple real-time input-output mapping models cannot capture this complex spatiotemporal dynamic relationship, resulting in delayed or drastic fluctuations in prediction instructions.

[0042] Choose neural networks with temporal memory capabilities. These networks can learn and remember key patterns and trends in long-term sequences through their internal states, thereby making more accurate predictions.

[0043] First, the raw time-series data acquired in step S1 undergoes rigorous standardization preprocessing. For data of length... Each feature dimension within the historical window (6 dimensions in total): Independent Z-score standardization, i.e. ,in and The mean and standard deviation of this feature are calculated from a large-scale training dataset during the model training phase and then fixed for use in network inference.

[0044] For traffic characteristics Before standardization, dimensionless processing is performed: Then on Z-score standardization is performed, a process that is completely consistent with the model training phase.

[0045] The neural network model adopted is a Long Short-Term Memory network architecture, specifically: a two-layer stacked LSTM network is constructed; the first LSTM layer is configured with 64 hidden units and obtains the hidden state of each time step in the complete time series, preserving the high-frequency details of micro-meteorological fluctuations; the second LSTM layer is configured with 32 hidden units and only returns the hidden state vector of the final time step, which is used to extract and condense the thermodynamic trend features within the entire time window.

[0046] A Dropout layer (dropout rate 0.1) is used between the two LSTM layers to prevent overfitting. The Dropout value is set to 0.1, which can be adjusted by the implementer according to the specific implementation scenario.

[0047] Finally, a fully connected output layer with only one neuron is connected, using ReLU (RectifiedLinearUnit) as the activation function to ensure that the model's output is non-negative. This output value is the dimensionless basic flow prediction. .

[0048] The training data for this model can be generated through high-fidelity offline physics simulation or collected and labeled by relevant personnel. During offline physics simulation, the controller follows a PID algorithm with physical limit constraints to ensure the physical validity of the training labels. In the deployment phase, the standardized historical time-series tensor is input into this pre-trained model, and after forward propagation, the model is obtained. .

[0049] Step S4: Construct a physical optimization model that includes energy balance constraints, and calculate the partial derivative of the physical loss function with respect to the flow control variables as the physical correction gradient.

[0050] Prediction results of neural network models It may violate basic physical laws in certain edge cases, such as predicting a large amount of water spray at night when radiation is low and humidity is high, which obviously cannot produce an evaporative cooling effect.

[0051] Therefore, it is necessary to perform physical-based corrections on the neural network output, constructing a differentiable physical forward model, and using its gradient with respect to the control input to guide the operation. Fine-tune in the direction of reducing physical loss.

[0052] The specific implementation includes: First, constructing a physical optimization model; second, establishing the thermal balance equation for the bridge deck surface based on the law of conservation of energy; and third, defining the equivalent area heat capacity of the pavement layer. ,in , , These are the density, specific heat, and effective thickness of the asphalt layer, respectively.

[0053] The net heat flux received per unit area of ​​the bridge deck at the current moment is ,in Net radiation, For convective heat transfer, For evaporative cooling heat flux, it is a dimensionless control quantity. The function.

[0054] Evaporative cooling flux The calculation is strictly constrained by physical limits: the total water flux available for evaporation at the current moment includes the water flux corresponding to the current command. and the current water depth of the bridge The equivalent water flux represented .

[0055] The total water flux rate and the actual atmospheric evaporation capacity Comparing the two, the actual evaporation rate is the smaller of the two values, i.e. ,but .

[0056] Based on this, the predicted surface temperature of the bridge deck at the next moment is: Therefore, a physical loss function can be established. The function has two objectives: first, to make the predicted temperature close to the target temperature, and second, to control energy consumption (i.e., the amount of water sprayed).

[0057] The specific form is as follows: in, This is the temperature scaling factor. and This is a weighting coefficient used to balance the cooling effect and energy-saving requirements; it is a hyperparameter that can be adjusted by the implementer according to the specific implementation scenario.

[0058] This allows for the calculation of the physical correction gradient, guiding the neural network model to fine-tune in the direction of reducing physical loss, and the application of the chain rule to calculate the loss function. Regarding control variables partial derivatives : in, .

[0059] The calculation requires the introduction of phase transition limit constraints: when When, it indicates that watering should be increased. It can effectively increase evaporation, at this time ; Conversely, when the total water flux has reached or exceeded the evaporation limit, increasing... It will not increase evaporation at this time. .

[0060] This gradient That is, to use neural network prediction The direction and force of correction required to guide the physical feasible region.

[0061] Step S5: Introduce an environmentally sensitive dynamic hysteresis factor to calculate the bridge surface moisture dissipation rate based on real-time wind speed and humidity characteristics in order to solve the dynamic water accumulation risk probability.

[0062] The risk of water accumulation on bridges is not determined solely by the instantaneous amount of water sprayed, but rather by the dynamic balance between historical water accumulation and current dissipation capacity.

[0063] The judgment method based on instantaneous flow threshold cannot take into account the combined effects of residual moisture from previous watering and the effects of wind speed and humidity on moisture evaporation, which can easily lead to misjudgment of risk.

[0064] Real-time estimation of the bridge surface's "moisture inventory" to construct a probability characteristic of water accumulation risk.

[0065] The specific implementation is as follows: First, construct an environment-aware dynamic hysteresis factor. This factor characterizes the overall dissipation rate of water per unit area on the bridge deck (unit: It consists of structural drainage items and environmental enhancement items: in, The structural gravity drainage reference term is determined by the cross slope of the bridge deck. and longitudinal slope The decision is made, and the calculation formula is as follows: empirical coefficient Determined based on the depth of the bridge deck pavement structure; This is an empirical coefficient. This item reflects wind speed, with reference to saturated water vapor pressure difference. and drying power It promotes the evaporation of water.

[0066] Secondly, the discrete-time analytical solution operator of the non-homogeneous linear differential equation is used to update the water accumulation depth state variable, and the bridge water accumulation depth is updated. The dynamic evolution (unit: meter) is described by a first-order linear differential equation: To avoid the cumulative error generated by the Euler integral method in long-term iterations, its accurate discrete-time analytical solution is used for updating: It ensures that regardless of the time step Regardless of the value chosen, the update process strictly satisfies the physical constraints and is unconditionally stable.

[0067] when When the expression approaches 0, the second term of the formula degenerates into... Maintaining physical continuity A modified probabilistic activation function is used to map the water depth state variable to a normalized dynamic water accumulation risk probability, while a translated and scaled sigmoid function is used to achieve this mapping: in, The slope coefficient controls the sensitivity of the probability as the water depth increases. This is the offset coefficient. and Hyperparameters can be adjusted by the implementer according to the specific implementation scenario.

[0068] For example, when the bridge deck is completely dry ( )hour, This indicates extremely low underlying risk.

[0069] For example, when the water depth reaches (i.e., 1.5 cm) , This indicates a medium risk level.

[0070] so, Provided in Risk indicators that are continuous, smooth, and have clear physical meaning within a given range.

[0071] Step S6: Based on the physical correction gradient, the predicted basic flow rate is corrected in a direction that conforms to the law of energy conservation, and the corrected flow rate is competitively gated and weighted using the dynamic water accumulation risk probability to generate the final water spraying command to drive the actuator.

[0072] After the aforementioned steps, data-driven predicted values ​​were obtained. Physically guided correction gradient and security risk assessment .

[0073] However, in high-risk situations, water spraying should be significantly reduced to ensure safety. Even if physical corrections suggest increasing water spraying, this should be rejected. Therefore, the intelligent prediction should first be fine-tuned using physical gradients to make it conform to energy laws. Then, the fine-tuned instructions should be reviewed for safety using risk probability, and strong attenuation should be applied in high-risk situations.

[0074] First, directional correction is performed, using the basic idea of ​​gradient descent, to adjust the base flow forecast value. By moving one step along the direction of decrease in the physical loss function, the physically corrected flow rate value is obtained. : in, It is the gradient descent step size coefficient, which is an adjustable hyperparameter that can be adjusted by the implementer according to the specific implementation scenario.

[0075] This will enable purely data-driven prediction Projecting onto a physical mapping that conforms to thermodynamic energy conservation, we obtain... .

[0076] Therefore, competitive gating weighting is performed to construct a dynamic weight coefficient based on the hyperbolic tangent function. This coefficient represents the probability of flooding. Functions: In the formula, As a safety baseline threshold, Steepness coefficient, and All of these are adjustable hyperparameters, which can be adjusted by the implementer according to the specific implementation scenario.

[0077] when hour, The item is 0. This means that the weighting coefficients remain in a fully passable state, and the security constraints are not effective.

[0078] when hour, The value exhibits a hyperbolic tangent characteristic and decays nonlinearly, smoothly approaching 0 from 1. The higher the risk, the stronger the decay.

[0079] The corrected flow With gate weights Multiply the results, ensuring the result is non-negative, to obtain the final dimensionless watering command: At the same time, the dimensionless command is converted into an actual physical flow command. The signal is transmitted to the frequency converter or electric regulating valve of the sprinkler system via an industrial fieldbus (such as Modbus RTU or Profinet), which drives the actuator to complete a complete control cycle.

[0080] This invention also discloses an automated bridge deck watering control system based on a neural network, comprising: a system including a processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement the automated bridge deck watering control method according to the above embodiments of the invention.

[0081] The present invention also discloses a system, which further includes other components well known to those skilled in the art, such as a communication bus and a communication interface. The settings and functions of these components are known in the art and will not be described in detail here.

[0082] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An automated control method for bridge deck watering based on neural networks, characterized in that, include: Real-time collection of micro-meteorological environment data and bridge deck thermal status data at the bridge site, and obtaining feedback on water spraying flow rate from the previous control cycle; Based on the principles of thermodynamic energy conservation and mass transport, the theoretical evaporation carrying capacity of the current atmospheric environment is calculated, and the thermodynamic net flux characteristics are constructed by combining the water spray flow feedback. The standardized micrometeorological environment data, bridge deck thermal state data, and sprinkler flow feedback are input into a pre-trained neural network model with time-series memory capabilities to generate the basic flow prediction value for the current moment. Construct a physical optimization model that includes energy balance constraints, and calculate the partial derivative of the physical loss function with respect to the flow control variable as the physical correction gradient; An environmentally-aware dynamic hysteresis factor is introduced to calculate the bridge surface moisture dissipation rate based on real-time wind speed and humidity characteristics in order to solve the dynamic water accumulation risk probability. Based on the physical correction gradient, the predicted basic flow rate is corrected in a direction that conforms to the law of energy conservation. The corrected flow rate is then subjected to competitive gating weighting using the dynamic water accumulation risk probability to generate the final water spraying command to drive the actuator.

2. The method according to claim 1, characterized in that, The introduction of an environmentally sensitive dynamic hysteresis factor, which calculates the bridge deck moisture dissipation rate based on real-time wind speed and humidity characteristics to determine the probability of dynamic water accumulation risk, specifically includes: An environmentally sensitive dynamic hysteresis factor is constructed to characterize the length of moisture dissipation memory. A nonlinear mapping relationship is established between the dynamic hysteresis factor and real-time wind speed and saturated water vapor pressure difference. The dynamic hysteresis factor includes a structural gravity drainage reference term and an environmentally sensitive term determined by wind speed and water vapor pressure difference. Based on the dynamic hysteresis factor, the current water depth state variable is updated using the discrete-time analytical solution operator of the non-homogeneous linear differential equation. The water depth state variable is mapped to the normalized dynamic water accumulation risk probability by a modified probabilistic activation function.

3. The method according to claim 2, characterized in that, The method of updating the current water depth state variable using a discrete-time analytical solution operator for a non-homogeneous linear differential equation specifically refers to: The current water depth is calculated using the discrete-time analytical solver. This calculation process considers the exponential decay term of the water depth at the previous time step and the input-output term contributed by the thermodynamic net flux characteristics. The analytical solver is used to eliminate the cumulative error of the Euler integral method in long-term iterations and ensure that the evolution of the water depth follows the law of conservation of physical mass.

4. The method according to claim 1, characterized in that, The construction of a physical optimization model with energy balance constraints, and the calculation of the partial derivative of the physical loss function with respect to the flow control variable as the physical correction gradient, specifically includes: Based on the full physical energy balance equation, using the current bridge deck thermal state data and the basic flow prediction value, the theoretical bridge deck surface temperature at the next moment is predicted. A physical loss function constrained by the target safe temperature of the road surface is established, which includes a temperature tracking error term and a control energy consumption penalty term; the chain rule is applied to calculate the partial derivative of the physical loss function with respect to the dimensionless flow rate. A phase transition limit constraint is introduced when calculating the partial derivatives. When the current total available water flux of the system exceeds the theoretical evaporation capacity, the gradient contribution of the evaporative cooling term to the flow rate is forcibly cut off.

5. The method according to claim 1, characterized in that, The competitive gating weighting of the corrected flow rate using the dynamic water accumulation risk probability specifically includes: A weighting coefficient based on a competition mechanism between safety risk and cooling benefits is constructed; when the probability of dynamic water accumulation risk is lower than the safety benchmark threshold, the weighting coefficient remains in a fully unobstructed state. When the probability of dynamic water accumulation exceeds the safety benchmark threshold, the weighting coefficient exhibits a hyperbolic tangent characteristic and decays nonlinearly. The final watering command is obtained by multiplying the flow rate value after physical correction gradient adjustment by the weighting coefficient.

6. The method according to claim 1, characterized in that, The calculation of the theoretical evaporation carrying capacity of the current atmospheric environment specifically includes: The calculation is based on the modified Penmanmont formula principle, combined with net radiation intensity, soil heat flux, air density, specific heat at constant pressure and saturated water vapor pressure difference; We introduce aerodynamic conductivity based on real-time wind speed calculation and convert the calculation results into liquid water flux velocity per unit area as the physical upper limit for determining whether water can be accepted by the atmosphere.

7. The method according to claim 1, characterized in that, The method also includes a cold start data population strategy: The system maintains a circular buffer for storing historical time-series data. During the system initialization phase, if the circular buffer is empty, the sensor data at the current moment and the initial flow feedback are collected, copied, and filled into the entire circular buffer to eliminate the distribution bias introduced by zero filling. During the training phase of the neural network model, the water flow feedback feature is constructed using the label value from the previous time step through autoregression.

8. The method according to claim 1, characterized in that, The neural network model adopts a long short-term memory network architecture: The neural network model includes two stacked long short-term memory layers; The first layer is configured to return the complete time series to preserve micro-meteorological fluctuation characteristics, and the second layer is configured to return the hidden state vector of the final time step to extract thermodynamic trend characteristics. The model's output layer is mapped to a non-negative space through a non-negative activation function to generate the dimensionless basic flow prediction value.

9. The method according to claim 1, characterized in that, The directional correction of the basic flow prediction value based on the physical correction gradient in accordance with energy conservation specifically refers to: subtracting the product of the gradient descent step size coefficient and the physical correction gradient from the basic flow prediction value to obtain the corrected flow. The correction is used to constrain data-driven predictions to a physical energy-conserving manifold.

10. An automated control system for bridge deck watering based on neural networks, characterized in that, It includes: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, it implements the bridge deck watering automation control method based on a neural network as described in claims 1-9.

Images