A bridge deck runoff real-time monitoring processing system and method based on deep learning

Abstract

The application discloses a bridge deck runoff real-time monitoring processing system and method based on deep learning, relates to the technical field of bridge deck runoff monitoring, and realizes the collaborative control of multiple bridge valves through a federal learning model, dynamically optimizes the resource allocation of an emergency pool, and effectively solves the resource occupation and response lag problems caused by traditional single-bridge independent control; the time sequence modeling of hydrological characteristics is combined with an LSTM network, the valve priority weight is dynamically corrected, and the environmental adaptability limitation of the traditional fixed threshold strategy is broken through; an encryption gradient aggregation mechanism is introduced, data privacy is protected while breaking the information island, and global pollution event rapid response is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bridge deck runoff monitoring, and in particular to a bridge deck runoff real-time monitoring processing system and method based on deep learning. BACKGROUND

[0002] As an important part of highway environmental protection facilities, bridge deck runoff monitoring mainly serves bridge scenes crossing sensitive water bodies, and its core goal is to detect runoff water quality parameters in real time to trigger electric valves and guide contaminated runoff into emergency treatment pools to prevent harmful substances from being directly discharged into the environment.

[0003] The typical system architecture, such as the single-bridge independent control mode adopted by patent CN207484630U, independently deploys a sensor network and a PLC controller at each bridge, and autonomously decides the opening and closing of the valve according to a preset threshold. In the scene where multiple bridges share limited emergency treatment resources, such as a centralized treatment station shared by urban viaduct groups, the valve actions of adjacent bridges lack a coordination mechanism. When the valve of A bridge is opened due to sudden pollution, the normal runoff of B / C bridge may be misdirected into the emergency pool, causing the treatment capacity to be occupied by low-priority flow. The traditional threshold judgment method cannot dynamically perceive the global state, such as the remaining pool capacity and the rainfall intensity, and may cause the treatment system to be overloaded due to simultaneous pollution of multiple bridges. The data between traditional control systems of each bridge are not interconnected, and manual scheduling relies on experience-based decision-making, making it difficult to achieve rapid resource allocation in complex events.

[0004] To alleviate the above problems, some solutions have expanded the emergency pool volume or added distributed treatment units to improve system redundancy, as described in CN201648961U, but this significantly increases land occupation and construction costs. Some solutions use static priority strategies, such as hierarchical control according to the distance of the bridge from the water source, which can reduce the probability of simultaneous pollution, but cannot adapt to the surge in runoff. SUMMARY

[0005] In view of the above existing problems, the present application is proposed.

[0006] The present application provides a bridge deck runoff real-time monitoring processing system and method based on deep learning to solve the problem that the traditional system uses a single-bridge independent control mode, resulting in multiple bridge resource competition, rigid response logic, and information isolation defects, and that hardware expansion and rule optimization solutions cannot fundamentally solve the problem of coordination control.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] In a first aspect, the present application embodiment provides a bridge deck runoff real-time monitoring processing system based on deep learning, which includes,

[0009] Bridge deck monitoring unit, deployed at each bridge runoff collection node, including pH sensor, oil sensor, conductivity sensor, turbidity sensor and video acquisition device;

[0010] Local control module, connected with the bridge deck monitoring unit through RS485 bus, built-in PLC controller and edge computing chip, used for storing sensor data and executing local valve control logic;

[0011] Cooperative decision module, including cloud server and optical fiber communication unit, the cloud server integrates federal learning model, and interacts gradient parameters with each local control module through the optical fiber communication unit;

[0012] Valve execution unit, including electric two-position three-way ball valve installed in each bridge deck drainage pipe, the drive circuit of which is electrically connected with the local control module;

[0013] Emergency pool capacity monitoring device, including ultrasonic liquid level meter and flow switch, which can real-time collect liquid level and inflow flow data of emergency treatment pool.

[0014] As a preferred scheme of the bridge deck runoff real-time monitoring and processing system based on deep learning, wherein: the edge computing chip is built-in sensor data completion module, which is used to perform the following operations when the sensor fails:

[0015] Extracting time series data and historical same working condition records of other sensors on the current bridge deck;

[0016] Calculating the correlation weight of the missing parameter and the context data through the self-attention mechanism;

[0017] Outputting the completed sensor sequence to the threshold judgment logic of the PLC controller.

[0018] As a preferred scheme of the bridge deck runoff real-time monitoring and processing system based on deep learning, wherein: in the local control module, the way of calculating the correlation weight of the missing parameter and the context data is:

[0019] Position encoding fusion is performed on the input sensor data, so that X' = X + PE, wherein X' represents the input data sequence after position encoding fusion, X represents the original sensor data sequence, and PE represents the position encoding matrix;

[0020] The position encoding is defined by using sine and cosine functions, and the formula is:

[0021]

[0022] Where i represents the position index in the data sequence, j represents the baseline index of the feature dimension, d represents the total feature dimension of the input data sequence, the constant 10000 is used to construct the periodic scale, and 2j and 2j+1 correspond to the index positions calculated by the sine and cosine functions, respectively.

[0023] A linear mapping is performed on the fused location-encoded data X′ to generate query, key, and value vectors, as shown in the formula:

[0024] Q = X′W Q K = X′W K V=X′W V ,

[0025] Where Q represents the generated query vector sequence, K represents the generated key vector sequence, V represents the generated value vector sequence, X′ is the data sequence after positional encoding fusion, and W Q W K and W V These are the corresponding linear mapping weight matrices;

[0026] A multi-head attention mechanism is employed to enhance model capabilities. Let H be the number of multi-head attention mechanisms and d be the feature dimension of each head. h :

[0027] d h =d / H, for the i-th head, the attention calculation formula is:

[0028] Among them, head i Q represents the output of the i-th attention head. i K i and V i Let i be the query, key, and value after linear projection of the i-th head, respectively. K represents i transpose, The scaling factor is used to balance the magnitude of the vector inner product, and softmax(·) represents the normalization operation;

[0029] Generate the vector projections for each head using the formula: Q i =X′W i Q ,K i =X′W i K V i =X′W i V Among them, W i Q W i K and W i VLet Q represent the linear mapping weights of the i-th head, respectively. i K i and V i This is the result of the corresponding linear transformation;

[0030] After concatenating the outputs of all heads, the multi-head attention result is obtained through output transformation:

[0031] MHA(X′)=Concat(head1,…,head H W O Where MHA(X′) represents the final output of the multi-head self-attention mechanism, and Concat(·) represents concatenating the outputs of each attention head according to its dimension. i For the i-th head output, W O The output is a fused linear transformation matrix, where H represents the total number of attention heads;

[0032] The number of attention heads is determined based on the input feature dimension and the minimum requirement per head, expressed as:

[0033]

[0034] Where H represents the final number of attention heads selected, and d is the total feature dimension of the input data sequence. min This represents the minimum feature dimension required for each attention head. This indicates a floor operation. In addition, considering the actual parallel computing capabilities of the hardware, the local attention strategy can constrain the computation range in certain regions, thereby reducing computational costs.

[0035] As a preferred embodiment of the deep learning-based bridge deck runoff real-time monitoring and processing system described in this invention, the federated learning model generates multi-bridge cooperative control strategies in the following ways:

[0036] Each local control module trains a local sub-model based on a deep reinforcement learning framework, with inputs including the current bridge deck water quality parameter sequence, the historical valve status of adjacent bridges, and the remaining capacity of the emergency pool.

[0037] The gradient parameters of the local sub-model are encrypted and uploaded to the cloud server for aggregation to generate global model parameters;

[0038] The global model parameters are sent to each local control module to dynamically adjust the opening combination of the electric two-position three-way ball valve.

[0039] As a preferred embodiment of the deep learning-based bridge surface runoff real-time monitoring and processing system described in this invention, the training process of the local sub-model includes:

[0040] A state space is constructed based on historical pollution events, which includes a four-dimensional vector of the current bridge deck water quality parameters, the valve action sequence of adjacent bridges in the past 5 minutes, and the emergency pool capacity change rate.

[0041] The PPO algorithm is used to optimize the motion space, which is the opening adjustment amount of the electric two-position three-way ball valve on each bridge deck;

[0042] The total inflow is constrained to not exceed the real-time capacity of the emergency pool by using the Lagrange relaxation method.

[0043] As a preferred embodiment of the deep learning-based bridge surface runoff real-time monitoring and processing system described in this invention, the collaborative decision-making module encrypts the gradient parameters of the local sub-model and uploads them to the cloud server for aggregation in the following way:

[0044] Homomorphic encryption is used to encrypt the gradient parameters of the local sub-models, and a secure aggregation rule is used to complete the global update. Let g be the gradient parameter calculated by each local sub-model. i ,i=1,2,…,N, where g i This represents the gradient parameter vector of the i-th local sub-model, and N represents the total number of sub-models participating in the co-training.

[0045] Encryption is achieved by using the homomorphic encryption function E(·) on each gradient parameter vector, resulting in... Where E(·) represents the homomorphic encryption operation function, and the encrypted gradient parameter is denoted as...

[0046] After receiving all the encryption gradients, the cloud server directly aggregates them using the homomorphic property, and the calculation formula is as follows: in, This represents the addition operation under homomorphic encryption, and the aggregation result is the global gradient under the encryption state;

[0047] After aggregation, the result is restored to the global gradient using a threshold decryption scheme:

[0048] Where D(·) represents the decryption function, which restores the encrypted result to the true gradient parameters. In order to balance the model convergence speed and data privacy protection, the amplitude of the encryption noise and the key parameters can be adjusted during the encryption process to ensure both aggregation accuracy and privacy requirements.

[0049] As a preferred embodiment of the deep learning-based bridge deck runoff real-time monitoring and processing system of the present invention, wherein: during the training of the local sub-model, the process of constraining by the Lagrange relaxation method includes:

[0050] A constrained optimization problem is constructed using the Lagrange relaxation method, and the adjustment amount of each valve is denoted as Δv.i , Where, Δv i This represents the opening adjustment amount of the i-th valve, m is the number of valves, and C represents the real-time reception capacity of the emergency pool.

[0051] By introducing the Lagrange multiplier λ, the original constraint is transformed into an unconstrained problem. The Lagrange function is then constructed for the objective function L(Δv):

[0052]

[0053] Where L(Δv) represents the loss function defined during training, and λ is the Lagrange multiplier;

[0054] While solving this optimization problem, λ is dynamically adjusted, and the updated formula is:

[0055]

[0056] Where, λ (t) Let represent the multiplier at the t-th iteration, and η represent the multiplier update step size. + The non-negative part is taken, λ≥0. This update strategy realizes the gradual adjustment of the multiplier according to the default amount, so that the total inflow strictly satisfies the constraint conditions.

[0057] As a preferred embodiment of the deep learning-based bridge deck runoff real-time monitoring and processing system described in this invention, the control strategy of the valve actuator includes a hierarchical response mechanism:

[0058] When the emergency pool level reaches the first threshold, the valve priority is dynamically allocated according to the distance between the bridge and the sensitive water body.

[0059] When the liquid level reaches the second threshold, the electric two-position three-way ball valve of the low-priority bridge is forcibly closed.

[0060] The priority weights are dynamically adjusted based on the regional hydrological characteristics using an LSTM network;

[0061] In the valve actuator unit, the method for dynamically correcting the regional hydrological characteristics is as follows:

[0062] Define a regional hydrological feature vector and use a fully connected layer to linearly encode the regional hydrological features;

[0063] The encoded feature sequence is input into the LSTM network;

[0064] LSTM networks capture the temporal dynamics of hydrological characteristics through gating mechanisms, enabling the system to adjust valve execution strategies in real time, thereby adapting to the ever-changing regional water conditions and ensuring that valve control responds to the actual conditions of the regional environment.

[0065] Secondly, this invention provides a method for real-time monitoring and processing of bridge surface runoff based on deep learning, including:

[0066] Step S1: Collect water quality parameters and video data through each bridge deck monitoring unit, and generate standardized feature vectors through local control module preprocessing.

[0067] Step S2: Input the feature vector into the federated learning model, and combine it with the real-time capacity of the emergency pool and meteorological forecast data to generate a multi-bridge collaborative valve control instruction set;

[0068] Step S3: Dynamically adjust the opening combination of each bridge deck electric two-position three-way ball valve according to the instruction set, so that bridge deck runoff with a pollution probability higher than the set threshold is preferentially introduced into the emergency pool.

[0069] Step S4: When a communication interruption is detected, switch to the redundant control strategy generated by the local sub-model to maintain the single-bridge foundation drainage function.

[0070] As a preferred embodiment of the deep learning-based real-time monitoring and processing method for bridge surface runoff described in this invention, the generation of the multi-bridge collaborative valve control instruction set includes:

[0071] The urgency of pollution is calculated based on the abrupt change characteristics of water quality parameters on each bridge deck.

[0072] Dynamic programming algorithm is used to optimize valve opening combinations, with the emission weight of high-urgency bridge decks being greater than that of low-urgency bridge decks;

[0073] In heavy rain, the anti-interference mode is activated, and the leakage identification results of the video acquisition device are used first to correct the sensor data.

[0074] Step S2, the steps for calculating the urgency of pollution, include:

[0075] A quantitative method for mutation indices is introduced, and the formula is defined as follows:

[0076]

[0077] in, This represents the current observed value of the l-th water quality parameter at time t. This represents the average value of the ιth water quality parameter over a past period of time. Let represent the standard deviation of the l-th water quality parameter within the same historical window, ∈ , and ∈ , where ...

[0078] Based on the abrupt change indicators of each sensor, a weighted summation is used to calculate the comprehensive pollution urgency, using the following formula:

[0079] Among them, E t represents the comprehensive pollution emergency degree at time t, and w ι is the weight coefficient corresponding to the ι-th type of water quality parameter and satisfies

[0080] The classification logic of the pollution emergency degree is classified according to the preset threshold:

[0081] When E t < T1, it is classified as a low emergency degree,

[0082] When T1 ≤ E t < T2, it is classified as a medium emergency degree,

[0083] When E t ≥ T2, it is classified as a high emergency degree,

[0084] Among them, T1 and T2 are emergency degree thresholds.

[0085] The beneficial effects of the present invention are as follows: The present invention realizes the collaborative control of multiple-bridge valves through a federated learning model, dynamically optimizes the resource allocation of the emergency pool, and effectively solves the problems of resource occupation and response lag caused by traditional single-bridge independent control; combined with the LSTM network for time-series modeling of hydrological characteristics, dynamically corrects the valve priority weight, and breaks through the environmental adaptability limitation of the traditional fixed threshold strategy; introduces an encrypted gradient aggregation mechanism to break the information island while protecting data privacy and realizes a rapid response to global pollution events; through the sensor data completion module and the hierarchical control strategy, significantly improves the robustness of the system in complex scenarios such as hardware failures and rainstorm interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0087] Figure 1 It is a schematic framework diagram of the real-time monitoring and processing system for bridge deck runoff based on deep learning in Embodiment 1.

[0088] Figure 2 It is a schematic flow diagram of the real-time monitoring and processing method for bridge deck runoff based on deep learning in Embodiment 1. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0089] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following will make a detailed description of the specific embodiments of the present invention in conjunction with the drawings of the specification.

[0090] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0091] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0092] Example 1, referring to Figure 1 and Figure 2 This embodiment provides a deep learning-based real-time monitoring and processing system for bridge deck runoff, including:

[0093] The bridge deck monitoring unit is deployed at each bridge runoff collection point and includes a pH sensor, an oil sensor, a conductivity sensor, a turbidity sensor, and a video acquisition device.

[0094] The local control module is connected to the bridge deck monitoring unit via an RS485 bus. It has a built-in PLC controller and edge computing chip, which are used to store sensor data and execute local valve control logic.

[0095] Edge computing chips have a built-in sensor data completion module to perform the following operations when sensors fail:

[0096] Extract time-series data and historical records of the same working conditions from other sensors on the current bridge deck;

[0097] The association weights between missing parameters and contextual data are calculated using a self-attention mechanism;

[0098] Output the completed sensor sequence to the threshold judgment logic of the PLC controller;

[0099] In the local control module, the method for calculating the association weight between missing parameters and context data is as follows:

[0100] The input sensor data is fused with position codes. Let X′=X+PE, where X′ represents the input data sequence after fusion with position codes, X represents the original sensor data sequence, and PE represents the position coding matrix.

[0101] Position coding is defined using sine and cosine functions, with the following formula:

[0102]

[0103] Where i represents the position index in the data sequence, j represents the baseline index of the feature dimension, d represents the total feature dimension of the input data sequence, the constant 10000 is used to construct the periodic scale, and 2j and 2j+1 correspond to the index positions calculated by the sine and cosine functions, respectively.

[0104] A linear mapping is performed on the fused location-encoded data X′ to generate query, key, and value vectors, as shown in the formula:

[0105] Q = X′W Q K = X′W K V=X′W V ,

[0106] Where Q represents the generated query vector sequence, K represents the generated key vector sequence, V represents the generated value vector sequence, X′ is the data sequence after positional encoding fusion, and W Q W K and W V These are the corresponding linear mapping weight matrices;

[0107] A multi-head attention mechanism is employed to enhance model capabilities. Let H be the number of multi-head attention mechanisms and d be the feature dimension of each head. h :

[0108] d h =d / H, for the i-th head, the attention calculation formula is:

[0109] Among them, head i Q represents the output of the i-th attention head. i K i and V i Let i be the query, key, and value after linear projection of the i-th head, respectively. K represents i transpose, The scaling factor is used to balance the magnitude of the vector inner product, and softmax(·) represents the normalization operation;

[0110] Generate the vector projections for each head using the formula: Q i =X′W i Q ,K i =X′W i K V i =X′W i V Among them, W i Q W i K and W i VLet Q represent the linear mapping weights of the i-th head, respectively. i K i and V i This is the result of the corresponding linear transformation;

[0111] After concatenating the outputs of all heads, the multi-head attention result is obtained through output transformation:

[0112] MHA(X′)=Concat(head1,…,head H W O Where MHA(X′) represents the final output of the multi-head self-attention mechanism, and Concat(·) represents concatenating the outputs of each attention head according to its dimension. i For the i-th head output, W O The output is a fused linear transformation matrix, where H represents the total number of attention heads;

[0113] The number of attention heads is determined based on the input feature dimension and the minimum requirement per head, expressed as:

[0114]

[0115] Where H represents the final number of attention heads selected, and d is the total feature dimension of the input data sequence. min This represents the minimum feature dimension required for each attention head. This indicates a floor operation. In addition, considering the actual parallel computing capabilities of the hardware, the local attention strategy can constrain the computation range in certain regions, thereby reducing computational costs.

[0116] Specifically, by controlling the number of attention heads, a balance is achieved between the model's expressive power and hardware resources; to ensure the processing quality of sub-modules, a minimum feature dimension constraint is introduced, so that the model can still run efficiently under resource constraints. The local attention strategy further reduces unnecessary computational burden, which helps to shorten processing latency and maintain a stable response when hardware computing power is limited.

[0117] The collaborative decision-making module includes a cloud server and a fiber optic communication unit. The cloud server integrates a federated learning model and interacts with each local control module for gradient parameters through the fiber optic communication unit.

[0118] Federated learning models generate multi-bridge cooperative control policies in the following ways:

[0119] Each local control module trains a local sub-model based on a deep reinforcement learning framework, with inputs including the current bridge deck water quality parameter sequence, the historical valve status of adjacent bridges, and the remaining capacity of the emergency pool.

[0120] The gradient parameters of the local sub-model are encrypted and uploaded to the cloud server for aggregation to generate global model parameters;

[0121] Global model parameters are sent to each local control module to dynamically adjust the opening combination of the electric two-position three-way ball valve;

[0122] The training process for the local sub-model includes:

[0123] A state space is constructed based on historical pollution events, which includes a four-dimensional vector of the current bridge deck water quality parameters, the valve action sequence of adjacent bridges in the past 5 minutes, and the emergency pool capacity change rate.

[0124] The PPO algorithm is used to optimize the motion space, which is the opening adjustment amount of the electric two-position three-way ball valve on each bridge deck;

[0125] The total inflow rate is constrained to not exceed the real-time reception capacity of the emergency pool by using the Lagrange relaxation method.

[0126] In the collaborative decision-making module, the method for encrypting the gradient parameters of the local sub-model and uploading them to the cloud server for aggregation is as follows:

[0127] Homomorphic encryption is used to encrypt the gradient parameters of the local sub-models, and a secure aggregation rule is used to complete the global update. Let g be the gradient parameter calculated by each local sub-model. i ,i=1,2,…,N, where g i This represents the gradient parameter vector of the i-th local sub-model, and N represents the total number of sub-models participating in the co-training.

[0128] Encryption is achieved by using the homomorphic encryption function E(·) on each gradient parameter vector, resulting in... Where E(·) represents the homomorphic encryption operation function, and the encrypted gradient parameter is denoted as...

[0129] After receiving all the encryption gradients, the cloud server directly aggregates them using the homomorphic property, and the calculation formula is as follows: in, This represents the addition operation under homomorphic encryption, and the aggregation result is the global gradient under the encryption state;

[0130] After aggregation, the result is restored to the global gradient using a threshold decryption scheme:

[0131] Where D(·) represents the decryption function, which restores the encryption result to the true gradient parameters. In order to balance the model convergence speed and data privacy protection, the amplitude of encryption noise and key parameters can be adjusted during the encryption process to ensure both aggregation accuracy and privacy requirements.

[0132] Specifically, by using homomorphic encryption technology, the direct connection between local data and cloud computing is effectively isolated. The encryption and aggregation process does not need to reveal the original gradient, which protects data privacy and maintains the global update of the model. By adopting the method of encryption before aggregation and decryption after aggregation, it not only effectively resists malicious attacks, but also allows the cloud server to complete numerical calculations without decryption, thereby reducing security risks.

[0133] The process of constraining the local sub-model using the Lagrange relaxation method includes:

[0134] A constrained optimization problem is constructed using the Lagrange relaxation method, and the adjustment amount of each valve is denoted as Δv. i , Where, Δv i This represents the opening adjustment amount of the i-th valve, m is the number of valves, and C represents the real-time reception capacity of the emergency pool.

[0135] By introducing the Lagrange multiplier λ, the original constraint is transformed into an unconstrained problem. The Lagrange function is then constructed for the objective function L(Δv):

[0136]

[0137] Where L(Δv) represents the loss function defined during training, and λ is the Lagrange multiplier;

[0138] While solving this optimization problem, λ is dynamically adjusted, and the updated formula is:

[0139]

[0140] Where, λ (t) Let represent the multiplier at the t-th iteration, and η represent the multiplier update step size. + The non-negative part is taken, λ≥0. This update strategy realizes the gradual adjustment of the multiplier according to the default amount, so that the total inflow strictly satisfies the constraint condition.

[0141] Specifically, the Lagrange relaxation method is used to introduce a multiplier to constrain the total inflow flow of the valve. Under the premise of ensuring that the emergency pool's capacity is not exceeded, the system can automatically adjust the valve control command. After constructing the Lagrange function, the system can dynamically adjust the penalty for the part that violates the constraint by iteratively updating the multiplier. This method has good adaptability and stability in practical applications. The update strategy ensures that the multiplier value is adjusted according to the current flow deviation in each training iteration, which can both accelerate model convergence and continuously control the flow within the allowable range.

[0142] The valve actuator includes an electrically operated two-position three-way ball valve installed on the drainage pipes of each bridge deck, and its drive circuit is electrically connected to the local control module.

[0143] The control strategy of the valve actuator includes a graded response mechanism:

[0144] When the emergency pool level reaches the first threshold, the valve priority is dynamically allocated according to the distance between the bridge and the sensitive water body.

[0145] When the liquid level reaches the second threshold, the electric two-position three-way ball valve of the low-priority bridge is forcibly closed.

[0146] Priority weights are dynamically adjusted based on regional hydrological characteristics using an LSTM network;

[0147] In the valve actuator unit, the method for dynamically correcting regional hydrological characteristics is as follows:

[0148] Define a regional hydrological feature vector and use a fully connected layer to linearly encode the regional hydrological features;

[0149] The encoded feature sequence is input into the LSTM network;

[0150] LSTM networks capture the temporal dynamics of hydrological characteristics through gating mechanisms, enabling the system to adjust valve execution strategies in real time, thereby adapting to the ever-changing regional hydrological conditions and ensuring that valve control responds to the actual conditions of the regional environment.

[0151] The emergency pool capacity monitoring device includes an ultrasonic level gauge and a flow switch, which collects real-time data on the liquid level and inflow rate of the emergency treatment pool.

[0152] This embodiment also provides a processing method for the above-mentioned deep learning-based bridge deck runoff real-time monitoring and processing system, including:

[0153] Step S1: Collect water quality parameters and video data through each bridge deck monitoring unit, and generate standardized feature vectors through local control module preprocessing.

[0154] Step S2: Input the feature vector into the federated learning model, and combine it with the real-time capacity of the emergency pool and meteorological forecast data to generate a multi-bridge collaborative valve control instruction set;

[0155] The generation of multi-bridge coordinated valve control instruction sets includes:

[0156] The urgency of pollution is calculated based on the abrupt change characteristics of water quality parameters on each bridge deck.

[0157] Dynamic programming algorithm is used to optimize valve opening combinations, with the emission weight of high-urgency bridge decks being greater than that of low-urgency bridge decks;

[0158] In rainy weather, start the anti-interference mode, and preferentially use the leakage identification results of the video acquisition device to correct the sensor data;

[0159] In step S2, the steps of calculating the pollution urgency include:

[0160] Introduce a quantization method for mutation indicators, and the defined formula is:

[0161]

[0162] Where represents the current observed value of the ι-th type of water quality parameter at time t, represents the average value of the ι-th type of water quality parameter over a past period of time, represents the standard deviation of the ι-th type of water quality parameter within the same historical window, ∈ is a small constant to prevent the denominator from being zero, ι = 1, 2,..., J and J represents the number of types of water quality parameters;

[0163] Based on the mutation indicators of each sensor, calculate the comprehensive pollution urgency by weighted summation, and the formula is:

[0164] Where, E t represents the comprehensive pollution urgency at time t, w ι is the weight coefficient corresponding to the ι-th type of water quality parameter, and satisfies

[0165] The classification logic of the pollution urgency is classified according to the preset threshold:

[0166] When E t < T1, it is classified as low urgency,

[0167] When T1 ≤ E t < T2, it is classified as medium urgency,

[0168] When E t ≥ T2, it is classified as high urgency,

[0169] Where, T1 and T2 are urgency thresholds;

[0170] Step S3, dynamically adjust the opening combination of each bridge deck electric two-way three-way ball valve according to the instruction set, so that the bridge deck runoff with a pollution probability higher than the set threshold is preferentially introduced into the emergency pool;

[0171] Step S4, when a communication interruption is detected, switch to the redundant control strategy generated by the local sub-model to maintain the single-bridge basic drainage function.

[0172] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A real-time monitoring and processing system for bridge deck runoff based on deep learning, characterized in that: include, The bridge deck monitoring unit is deployed at each bridge runoff collection point and includes a pH sensor, an oil sensor, a conductivity sensor, a turbidity sensor, and a video acquisition device. The local control module is connected to the bridge deck monitoring unit via an RS485 bus. It has a built-in PLC controller and edge computing chip, which are used to store sensor data and execute local valve control logic. The collaborative decision-making module includes a cloud server and an optical fiber communication unit. The cloud server integrates a federated learning model and interacts with each local control module through the optical fiber communication unit to exchange gradient parameters. The valve actuator includes an electrically operated two-position three-way ball valve installed on the drainage pipes of each bridge deck, and its drive circuit is electrically connected to the local control module. The emergency pool capacity monitoring device includes an ultrasonic level gauge and a flow switch, which collects the liquid level and inflow flow data of the emergency treatment pool in real time. The federated learning model generates multi-bridge cooperative control strategies in the following ways: Each local control module trains a local sub-model based on a deep reinforcement learning framework, with inputs including the current bridge deck water quality parameter sequence, the historical valve status of adjacent bridges, and the remaining capacity of the emergency pool. The gradient parameters of the local sub-model are encrypted and uploaded to the cloud server for aggregation to generate global model parameters; The global model parameters are sent to each local control module to dynamically adjust the opening combination of the electric two-position three-way ball valve; The training process of the local sub-model includes: A state space is constructed based on historical pollution events, which includes a four-dimensional vector of the current bridge deck water quality parameters, the valve action sequence of adjacent bridges in the past 5 minutes, and the emergency pool capacity change rate. The PPO algorithm is used to optimize the motion space, which is the opening adjustment amount of the electric two-position three-way ball valve on each bridge deck; The total inflow is constrained to not exceed the real-time capacity of the emergency pool by using the Lagrange relaxation method.

2. The bridge deck runoff real-time monitoring and processing system based on deep learning as described in claim 1, characterized in that: The edge computing chip has a built-in sensor data completion module, which is used to perform the following operations when the sensor fails: Extract time-series data and historical records of the same working conditions from other sensors on the current bridge deck; The association weights between missing parameters and contextual data are calculated using a self-attention mechanism; The completed sensor sequence is output to the threshold judgment logic of the PLC controller.

3. The bridge deck runoff real-time monitoring and processing system based on deep learning as described in claim 2, characterized in that: In the local control module, the method for calculating the association weight between the missing parameter and the context data is as follows: Position encoding fusion is performed on the input sensor data, so that... ,in, This represents the input data sequence after fusion and position encoding. Represents the original sensor data sequence. Represents the position encoding matrix; Position coding is defined using sine and cosine functions, with the following formula: ， in, Indicates the position index in the data sequence. The baseline index representing the feature dimension. Represents the total feature dimension of the input data sequence, a constant. Used to construct periodic scales and These correspond to the index positions used in the calculation of the sine and cosine functions, respectively. Data after fusion location encoding Perform a linear mapping to generate a vector of queries, keys, and values, using the following formula: ， in, This represents the generated query vector sequence. This represents the generated key vector sequence. This represents the generated value vector sequence. , and These are the corresponding linear mapping weight matrices; A multi-head attention mechanism is employed to enhance model capabilities. Let H be the number of attention heads and the feature dimension of each head be . : For the first The formula for calculating attention is: ,in, Indicates the first The output of each attention head, , and The first Queries, keys, and values ​​after linear projection of the head express transpose, The scaling factor is used to balance the magnitude of the vector inner product. This indicates a normalization operation; The formula for generating the vector projections of each head is as follows: , among which, among which, , and They represent the first Linear mapping weights for each size; After concatenating the outputs of all heads, the multi-head attention result is obtained through output transformation: ,in, This represents the final output of the multi-head self-attention mechanism. This indicates that the outputs of each attention head are concatenated according to their dimensions. For the first Output per unit size To output the fused linear transformation matrix, Indicates the number of attention heads; The number of attention heads is determined based on the input feature dimension and the minimum requirement per head, expressed as: ， in, This represents the minimum feature dimension required for each attention head. This indicates a round-down operation.

4. The bridge deck runoff real-time monitoring and processing system based on deep learning as described in claim 1, characterized in that: In the collaborative decision-making module, the gradient parameters of the local sub-model are encrypted and then uploaded to the cloud server for aggregation in the following way: Homomorphic encryption is used to encrypt the gradient parameters of the local sub-models, and a secure aggregation rule is used to complete the global update. Let the gradient parameters calculated by each local sub-model be denoted as... ,in, Indicates the first The gradient parameter vector of each local sub-model This represents the total number of sub-models participating in the collaborative training; Apply a homomorphic encryption function to each gradient parameter vector To achieve encryption, ,in, This represents the homomorphic encryption operation function, and the gradient parameters after encryption are denoted as... ; After receiving all the encryption gradients, the cloud server directly aggregates them using the homomorphic property, and the calculation formula is as follows: ,in, This represents the addition operation under homomorphic encryption, and the aggregation result is the global gradient under the encryption state; After aggregation, the result is restored to the global gradient using a threshold decryption scheme: ,in, This represents the decryption function, which restores the encrypted result to the true gradient parameters.

5. The bridge deck runoff real-time monitoring and processing system based on deep learning as described in claim 4, characterized in that: During the training of the local sub-model, the process of constraining using the Lagrange relaxation method includes: A constrained optimization problem is constructed using the Lagrange relaxation method, and the adjustment amount of each valve is denoted as . ,in, Indicates the first The valve opening adjustment amount For the number of valves, This indicates the real-time capacity of the emergency pool; Introducing Lagrange multipliers Then, the original constrained problem is transformed into an unconstrained problem, and the objective function is then... Construct the Lagrangian function: ， in, This represents the loss function defined during training. For Lagrange multipliers; While solving this optimization problem, The formula is dynamically adjusted and updated as follows: ， in, Indicates the first The multiplier in the next iteration Indicates the step size of the multiplier update. This indicates taking the non-negative part. .

6. The bridge deck runoff real-time monitoring and processing system based on deep learning as described in claim 1, characterized in that: The control strategy of the valve actuator includes a graded response mechanism: When the emergency pool level reaches the first threshold, the valve priority is dynamically allocated according to the distance between the bridge and the sensitive water body. When the liquid level reaches the second threshold, the electric two-position three-way ball valve of the low-priority bridge is forcibly closed. Priority weights are dynamically adjusted based on regional hydrological characteristics using an LSTM network.

7. A method for real-time monitoring and processing of bridge deck runoff based on deep learning, based on the system described in any one of claims 1 to 6, characterized in that, include: Step S1: Collect water quality parameters and video data through each bridge deck monitoring unit, and generate standardized feature vectors through local control module preprocessing. Step S2: Input the feature vector into the federated learning model, and combine it with the real-time capacity of the emergency pool and meteorological forecast data to generate a multi-bridge collaborative valve control instruction set; Step S3: Dynamically adjust the opening combination of each bridge deck electric two-position three-way ball valve according to the instruction set, so that bridge deck runoff with a pollution probability higher than the set threshold is preferentially introduced into the emergency pool. Step S4: When a communication interruption is detected, switch to the redundant control strategy generated by the local sub-model to maintain the single-bridge foundation drainage function.

8. The method for real-time monitoring and processing of bridge deck runoff based on deep learning as described in claim 7, characterized in that, The generation of the multi-bridge cooperative valve control instruction set includes: The urgency of pollution is calculated based on the abrupt change characteristics of water quality parameters on each bridge deck. Dynamic programming algorithm is used to optimize valve opening combinations, with the emission weight of high-urgency bridge decks being greater than that of low-urgency bridge decks; In heavy rain, the anti-interference mode is activated, and the leakage identification results of the video acquisition device are used first to correct the sensor data.

Images