A data acquisition and transmission system for safety monitoring of a hydropower station dam

Abstract

The application discloses a kind of hydropower station dam safety monitoring data acquisition and transmission system, it is related to hydropower station dam safety monitoring technical field, including cloud twin brain module and edge neuron.Cloud twin brain module contains sequence neural network engine and reflection core generation module, sequence neural network engine adopts neural network architecture with parallel and cyclic dual representation, including time mixing module and channel mixing module, can learn dam normal operation mode from historical monitoring data;Reflection core generation module compresses reference twin model into lightweight reflection core model and issues to edge device.Edge neuron contains micro-twin body prediction module and hierarchical transmission control module, micro-twin body prediction module real-time predicts dam state theoretical expectation value, calculates and the reflection deviation of actual observation value;Hierarchical transmission control module implements three-level response strategy according to reflection deviation size, abnormal trend transmission abstract information, emergency exception instant upload.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of dam safety monitoring of hydropower stations, and particularly relates to a dam safety monitoring data acquisition and transmission system for hydropower stations. BACKGROUND

[0002] As important water conservancy infrastructure, the safe operation of hydropower dams is related to the safety of people's lives and property and social and economic development downstream. With the rapid development of China's hydropower industry, dam safety monitoring technology has become a key technical means to ensure the long-term safe operation of dams. Traditional dam safety monitoring mainly relies on manual inspection and simple instrument measurement, which is difficult to meet the needs of modern dam safety management.

[0003] In recent years, with the development of Internet of Things, sensor technology and data communication technology, automatic monitoring systems have gradually become the mainstream technology for dam safety monitoring. Existing dam safety monitoring systems usually use distributed sensor networks to collect deformation, seepage, stress and other physical quantity data of various parts of the dam, and transmit the data to the monitoring center through wired or wireless communication networks for centralized analysis and processing. To some extent, this kind of system improves the automation level of monitoring and the real-time nature of data acquisition.

[0004] The existing technology has certain limitations through the traditional data acquisition and transmission method, such as: first, the traditional monitoring system uses a timing full-data transmission strategy, which requires continuous transmission of a large amount of monitoring data regardless of the normal operation state of the dam, resulting in serious waste of network bandwidth resources and increasing the system operation cost; second, the existing system mainly relies on simple threshold comparison for abnormal judgment, lacks deep understanding of the dam operation state, and cannot effectively distinguish between normal changes caused by environmental factors and real abnormal states, resulting in high false positive rate and false negative rate, affecting the reliability of the monitoring system; third, the traditional monitoring system lacks self-learning and evolution ability, and cannot optimize the monitoring strategy and improve the detection accuracy according to the accumulation of dam operation experience, with poor adaptability; fourth, the data processing of the existing system is mainly concentrated in the monitoring center, and the edge device only has simple data acquisition function and cannot perform local intelligent analysis, so the system reliability decreases when the network connection is unstable; fifth, the abnormal detection algorithm of the traditional system is relatively simple, and it is difficult to capture the complex change pattern of the dam operation state, especially for the recognition ability of gradual change and multi-variable coupling.

[0005] Therefore, there is an urgent need for a dam safety monitoring data acquisition and transmission system for hydropower stations with intelligent abnormal detection capability, on-demand data transmission, and self-learning evolution function to improve monitoring accuracy and efficiency, reduce system operation cost, and enhance the reliability and adaptability of the system. SUMMARY

[0006] The water power station dam safety monitoring data acquisition and transmission system of the present application can overcome the defects of the prior art and solve the above problems.

[0007] The water power station dam safety monitoring data acquisition and transmission system of the present application can overcome the defects of the prior art and solve the above problems.

[0008] The twin brain module is deployed in a cloud server and includes:

[0009] The sequence neural network engine adopts a neural network architecture with parallel and cyclic dual equivalent representation, and the architecture includes a time mixing module and a channel mixing module. The time mixing module performs weighted processing on dam multi-source historical monitoring data through a learnable time decay vector, and the channel mixing module is used to model the coupling relationship across variables. The sequence neural network engine processes deformation data, seepage data, stress and strain data, reservoir water level data, rainfall data and environmental temperature data of the dam in parallel mode during training to generate a baseline twin model capable of representing the normal operating state of the dam.

[0010] The reflection core generation module is used to quantitatively compress and optimize the trained baseline twin model, extract core inference logic, generate a lightweight reflection core model with significantly reduced parameter size, and distribute it to the edge neuron through network communication.

[0011] The edge neuron deployed in the dam monitoring site includes:

[0012] The model and data interface module is used to receive and load the reflection core model from the cloud, establish data connection with the on-site sensors, and real-time acquire physical quantity actual observation values and corresponding environmental variable observation values measured by the sensors.

[0013] The data preprocessing module includes an instance normalization unit, a sequence blocking unit and a token generation unit, which is used to standardize the continuous observation value sequence, block it according to the preset length and step, and convert the blocked data into block sequence tokens conforming to the input format of the reflection core model.

[0014] The online twin prediction module calls the cyclic inference mode of the reflection core model, based on the current time environmental variable observation value and historical state information, through the constant level time complexity cyclic inference mode, real-time predicts the theoretical normal value of each monitoring physical quantity under the current working condition, and forms a dynamic reference trajectory with the theoretical normal value sequence.

[0015] The reflection deviation calculation module calculates the Euclidean distance between the actual observation value and the theoretical normal value of the physical quantity as the reflection deviation value, and includes a deviation statistical analysis unit for calculating the statistical characteristics of the reflection deviation value in a sliding time window, including mean, variance, change trend coefficient and cumulative characteristics;

[0016] The hierarchical transmission control module executes a three-level adaptive data transmission strategy according to the amplitude, duration and change rate of the reflection deviation value: the first level local record response is triggered when the reflection deviation value is below the first preset threshold, the second level summary information transmission response is triggered when the reflection deviation value is continuously above the first preset threshold but below the second preset threshold, and the third level immediate early warning information transmission response is triggered when the reflection deviation value is instantaneously above the second preset threshold or the change rate exceeds the preset limit value.

[0017] The local storage module is used for caching observation data, reflection deviation values and system operation logs.

[0018] The system also includes a closed-loop model evolution mechanism, which receives uploaded summary information and early warning information data through the incremental learning module, performs incremental training and parameter fine-tuning on the baseline twin model, and maintains iteration records through the model version management module. After the model update is completed, the generation and distribution of the new version of the reflection kernel model are automatically triggered, realizing the continuous self-evolution ability of the system.

[0019] The time mixing module of the sequence neural network engine assigns different decay weights to data at different time steps in the historical monitoring data sequence, so that the influence weight of historical data farther away from the current time is smaller, effectively capturing long-term dependencies. The channel mixing module realizes the modeling of the correlation between different monitoring variables by information interaction and fusion in the feature dimension. The sequence neural network engine adopts a parallel computing mode in the training stage to improve training efficiency, and switches to a loop computing mode in the inference stage to realize efficient real-time inference.

[0020] The workflow of the reflection kernel generation module includes: model pruning and parameter quantization of the trained baseline twin model to remove redundant parameters and calculation paths; extracting the core inference logic and key parameters of the baseline twin model to generate a lightweight reflection kernel model suitable for edge computing environment; performing edge device compatibility test on the generated reflection kernel model to ensure its stable operation in resource-constrained environment; and securely distributing the reflection kernel model to each edge neuron through encrypted transmission.

[0021] The prediction mechanism of the online twin prediction module includes: receiving the reservoir water level observation value, the environmental temperature observation value and the rainfall observation value at the current time as input conditions; based on the environmental variable observation value and the historical state information, the expected value of each monitoring physical quantity at the next time is calculated through the cycle inference of the reflection kernel model; the expected value sequence of continuous prediction is composed into a dynamic reference trajectory, which reflects the normal behavior pattern of the dam under the current environmental conditions; the prediction process has constant level time complexity, which ensures real-time response requirements.

[0022] The reflection deviation calculation module quantifies the degree of system deviation by calculating the Euclidean distance between the actual observation value vector and the theoretical normal value vector; the deviation statistical analysis unit continuously tracks the change of the reflection deviation value within the preset sliding time window, calculates its statistical characteristics including the average deviation, the deviation variance, the change trend coefficient and the cumulative deviation amount, which are used to evaluate the stability and abnormal development trend of the system state.

[0023] The three-level response strategy of the hierarchical transmission control module is as follows: the first level local recording response, when the reflection deviation value is lower than the first preset threshold, the reflection deviation value and the corresponding observation value are stored in the local storage module, and no data transmission is performed to save communication resources; the second level summary transmission response, when the reflection deviation value continuously exceeds the first preset threshold but is lower than the second preset threshold, a data summary containing the deviation statistical characteristics, the change trend analysis and the time mark is generated, and the network transmission is performed in a low priority mode; the third level early warning transmission response, when the reflection deviation value exceeds the second preset threshold or the deviation change rate exceeds the preset safety limit, the complete data record of the abnormal period is immediately intercepted, and the real-time transmission is performed in the highest priority.

[0024] The instance normalization unit of the data preprocessing module independently normalizes the observation value sequence of each sensor channel, eliminating the differences in dimension and numerical range between different sensors; the sequence blocking unit splits the long time sequence according to fixed block length and overlap step, generating short sequence blocks suitable for reflection kernel model processing; the token generation unit converts each sequence block into a standardized input token containing position encoding and feature encoding.

[0025] The closed-loop model evolution mechanism realizes continuous learning through the following steps: the hierarchical transmission control module uploads the data generated by the second and third level responses to the twin brain module; the incremental learning module performs quality evaluation and label annotation on the collected new data, distinguishing normal changes, abnormal events and equipment failures; an incremental training algorithm is used to update the parameters of the baseline twin model, maintaining the memory of historical knowledge while learning new behavior patterns; the model version management module records the performance indicators and change contents of each update, and triggers version release when the performance of the new model reaches the preset improvement standard.

[0026] The incremental learning module adopts an experience replay strategy, mixes the newly collected data with historical representative data according to a preset proportion for training, prevents overfitting of the model to new data, adopts a contrast learning method to enhance the identification ability of the model to abnormal patterns, and improves the discrimination performance of the model by constructing positive and negative sample pairs; a multi-level model performance evaluation system is set up, including prediction accuracy indicators, abnormal detection rate indicators and false alarm rate indicators, to ensure the quality and reliability of model updating.

[0027] The hardware architecture of the edge neuron includes: an embedded computing unit adopting a low-power high-performance processor and a preset capacity of memory, used for running a reflex core model and data processing tasks; a multi-channel data acquisition unit supporting simultaneous access of analog signals, digital signals and pulse signals; a local storage unit providing a preset capacity of data cache space, supporting local backup and historical query of data; a communication unit supporting wireless network communication and wired network communication, having communication redundancy and fault switching capability; a clock synchronization unit achieving high-precision time synchronization through a satellite positioning system, ensuring the time consistency of distributed monitoring data.

[0028] The beneficial effects of the present application are:

[0029] By constructing a cloud twin brain module, deep intelligent modeling of the dam operation state is realized. The twin brain module adopts a sequence neural network engine with parallel and cyclic dual representation capabilities, which can efficiently process massive historical monitoring data, and fully utilizes cloud parallel computing resources in the training stage, significantly improving the model training efficiency and convergence speed. Compared with the traditional single architecture neural network, the dual representation mechanism of the present application avoids the contradiction between training efficiency and inference efficiency, and realizes the unified optimization of both.

[0030] The edge neuron adopts a micro-twin prediction mechanism, which can generate the theoretical expected value of the dam state in real time on site, and realizes intelligent anomaly detection through calculation of the reflection deviation. This design fundamentally changes the simple threshold alarm mode of the traditional monitoring system, enabling the system to have real intelligent judgment capability. The edge neuron can autonomously distinguish between normal changes caused by environmental factors and real abnormal states, significantly reducing false positives and false negatives.

[0031] The hierarchical transmission control mechanism dynamically adjusts the data transmission strategy according to the size and change trend of the reflection deviation, realizing intelligent management of network resources. When the dam is operating normally, the system hardly produces data transmission demand, greatly reducing network bandwidth occupation and communication cost. When potential abnormalities are detected, the system can automatically extract key information for compressed transmission. When an emergency anomaly is found, the system immediately triggers high-priority transmission to ensure timely communication of key information. This intelligent transmission strategy not only ensures the real-time performance of monitoring, but also realizes efficient use of communication resources.

[0032] The closed-loop model evolution mechanism enables the system to continuously self-learn and optimize. Through intelligent analysis and model updating of newly collected data by the incremental learning module, the system can continuously adapt to changes in the dam operating state, learn new abnormal patterns, and improve detection accuracy. The introduction of experience replay strategy and contrastive learning method effectively prevents the catastrophic forgetting phenomenon, ensuring that the system maintains memory of historical experience while learning new knowledge. The multi-level model performance evaluation system provides reliable quality assurance for model updating, ensuring the directionality and stability of system evolution.

[0033] The reflection kernel generation and delivery mechanism enables efficient transfer of cloud intelligence to edge devices. Through model compression, parameter quantization, and other techniques, complex cloud models are converted into lightweight versions suitable for edge device operation, enabling edge devices to have independent intelligent analysis capabilities. The incremental update mechanism only transmits the changed part of the model parameters, significantly reducing the network transmission volume of model delivery and improving the efficiency and real-time performance of system updates.

[0034] The intelligent hierarchical transmission strategy enables on-demand data transmission, minimizing network resource consumption while ensuring monitoring quality. To address the high false alarm rate and lack of intelligent judgment in existing systems, the invention uses micro-twin prediction and reflection deviation calculation to enable the system to have a deep understanding of dam state changes and accurately distinguish between normal changes and abnormal states. To address the problem of traditional systems being unable to evolve and having poor adaptability, a complete closed-loop learning mechanism is constructed to enable the system to continuously optimize performance as operating experience accumulates.

[0035] The local intelligent analysis capability of edge neurons enables the system to function normally even in unstable network connections, improving the system's robustness and usability. The cloud-edge collaborative architecture design fully leverages the rich cloud computing resources and rapid edge response, optimizing the allocation of computing resources. Model version management and automated deployment mechanisms ensure the safety and traceability of system updates, providing reliable guarantees for long-term stable operation.

[0036] Not only suitable for hydropower dam monitoring, but also applicable to other large infrastructure safety monitoring fields such as bridge monitoring, tunnel monitoring, high-rise building structure monitoring, etc., with broad application prospects and important engineering practical value. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 System architecture of the invention Figure One ;

[0038] Figure 2 System architecture of the invention Figure Two . Detailed Implementation

[0039] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] It should be noted that the directional concepts of "left", "right", "up", "down", "front", "back", "inner", and "outer" in the following scheme are all relative directions, and will not be listed one by one here.

[0041] Example 1: Construction and Training of Twin Brain Modules

[0042] This embodiment details the specific structure and implementation of a twin brain module in a hydropower station dam safety monitoring data acquisition and transmission system. Deployed on a cloud server, this twin brain module serves as the intelligent core of the entire monitoring system. It possesses the ability to learn from multi-source historical monitoring data of the dam and establish a normal operation status representation model. It can also provide lightweight intelligent inference models for edge computing devices deployed on-site, thereby realizing a cloud-edge collaborative intelligent monitoring architecture. The design and implementation of this twin brain module provide strong technical support for the system described in this invention, enabling the entire monitoring system to possess highly intelligent data processing and analysis capabilities.

[0043] This twin brain module adopts a modular design, mainly comprising two core components: a sequence neural network engine and a reflection kernel generation module. These two components work together to achieve a complete process from learning from historical data to generating a lightweight model. The sequence neural network engine, as the core of intelligent learning, employs a neural network architecture with both parallel and recurrent equivalent representations. This architecture has significant technical advantages: during model training, it can fully utilize parallel computing resources, simultaneously inputting dam monitoring data from multiple time steps, including deformation data, seepage data, and stress-strain data reflecting the dam's structural response, as well as reservoir water level data, rainfall data, and ambient temperature data reflecting external environmental conditions. Parallel processing significantly improves training efficiency. During model inference, it automatically switches to a recurrent computing mode, recursively processing the input data step by step, achieving efficient real-time inference with constant-time complexity. This design makes the generated model particularly suitable for deployment and operation in resource-constrained edge computing environments.

[0044] The internal structure of the sequence neural network engine is carefully designed, including a special time mixing module and a channel mixing module, which are responsible for processing the time-dependent relationship in time series data and the coupling relationship between multiple variables, respectively. The time mixing module implements an advanced time information fusion mechanism, which intelligently weights the dam multi-source historical monitoring data by maintaining a learnable time decay vector. Each element of the vector corresponds to a different time offset, which can adaptively learn the importance distribution of historical data, so that the historical data farther away from the current time is given a smaller influence weight, effectively capturing the long-term dependence in dam monitoring data, while avoiding the common gradient vanishing problem in traditional recurrent neural networks.

[0045] The time mixing module adopts an adaptive decay mechanism in specific implementation, which not only considers the influence of time distance on data importance, but also fully considers the periodicity characteristics inherent in dam monitoring data. The calculation of its time decay weight adopts the following mathematical formula:

[0046] ;

[0047] Among them: represents the decay weight of the time offset ; is the basic decay coefficient, which controls the overall decay speed; is the decay exponent, which determines the shape of the decay function; is the periodic modulation coefficient, which controls the strength of the periodic modulation; is the periodic frequency parameter, which determines the frequency of periodic changes; is the time offset, which represents the time difference between historical data and the current time. This formula introduces a periodic modulation term in the form of a sine function, so that the time mixing module can effectively identify and utilize seasonal change patterns, day-night change rules, and other periodic physical phenomena in dam monitoring data, which is of great significance for accurately modeling the normal behavior patterns of the dam.

[0048] The time mixing module adopts a feature fusion mechanism when processing historical monitoring data, which combines data at different time steps according to their time decay weights to generate a feature representation containing rich time information. The mathematical expression of its feature fusion process is:

[0049] ;

[0050] Among them: is the feature representation vector after time mixing; is the input feature vector of the i-th time step; and are learnable weight matrices, which are adaptively adjusted through training; is the bias vector; represents the element-wise multiplication operation; T is the total length of the time series. This design draws on the idea of attention mechanism by introducing the concepts of query and key, enabling the model to automatically learn the importance of data at different time steps, thus achieving more intelligent fusion of time information.

[0051] Corresponding to the time mixing module, the channel mixing module is specifically responsible for handling the complex coupling relationship between multiple different physical quantities in the dam monitoring system. Since the safety state of the dam is influenced by a variety of factors, including the response characteristics of the structure itself and the changes in external environmental conditions, there is a complex nonlinear coupling relationship between these different monitoring variables. The channel mixing module uses a multi-head coupled attention mechanism to model these cross-variable correlations, which can automatically learn and capture the potential correlation patterns between deformation data, seepage data, stress-strain data, reservoir water level data, rainfall data, and environmental temperature data.

[0052] The channel mixing module uses a multi-head coupled attention mechanism to model the coupling relationship between variables. The core idea of this mechanism is to automatically discover and utilize the correlation between different monitoring variables by learning their attention weights. Specifically, this mechanism calculates the attention weight of each monitoring variable to all other variables, thereby achieving information interaction and fusion. The formula for calculating the attention weight is:

[0053] ;

[0054] where: represents the attention weight of variable i to variable j in the hth attention head, reflecting the influence degree of variable j on variable i; and are the query vector and the key vector, respectively, which are generated from the input features through linear transformation; is a learnable coupling relationship bias matrix, used to encode prior physical coupling relationships; C is the total number of monitoring variables; is the dimension of the key vector. The coupling relationship bias matrix introduced in this formula is an important feature that allows the model to combine prior knowledge from domain experts to better learn the correlation between physically related variables.

[0055] The output feature calculation formula of the channel mixing module is:

[0056]

[0057] where: the calculation formula of each attention head is:

[0058] ;

[0059] Cross-variable interaction function is defined as:

[0060]

[0061] wherein: is an output weight matrix; is an output bias vector; H is the number of attention heads; is the h-th value weight matrix; [;] represents vector concatenation operation; is an activation function; is an interaction function weight matrix. The interaction function can capture linear relationships, nonlinear interactions and difference features between variables.

[0062] The sequence neural network engine fuses time and channel features using a gating mechanism:

[0063] ;

[0064] ;

[0065] wherein: is a gating vector; is a gating weight matrix; is a gating bias vector; is the final output feature.

[0066] The sequence neural network engine has a dual computing mode: in training, it uses a parallel mode to process historical monitoring data, improving training efficiency; in inference, it switches to a loop mode to achieve real-time prediction with constant time complexity. The state update formula of the loop mode is:

[0067] ; wherein: is the current state; is a dynamic forgetting factor.

[0068] Through deep learning of the deformation data, seepage data, stress and strain data, reservoir water level data, rainfall data and environmental temperature data of the dam, the sequence neural network engine generates a benchmark twin model capable of representing the normal operating state of the dam.

[0069] The training of the benchmark twin model uses a multi-task learning loss function:

[0070]

[0071] wherein: the prediction loss ; the consistency loss ; the physical constraint loss is the weighting coefficient; N is the number of training samples; M is the number of consistency verification samples; K is the number of physical constraints; This is a predicted value; This is the actual value; For disturbance; This is the k-th physical constraint function; This is the constraint threshold.

[0072] The baseline twin model is trained using supervised learning, with monitoring data from historical normal operation as training samples. The model parameters are optimized by minimizing the error between the predicted value and the actual observed value.

[0073] The reflection kernel generation module is used to quantize and compress the trained benchmark twin model and optimize its parameters, extract the core inference logic, and generate a lightweight reflection kernel model with significantly reduced parameter size.

[0074] The reflection kernel generation module employs an adaptive importance pruning algorithm, and the parameter importance scoring formula is as follows:

[0075]

[0076] The formula for calculating the parameter retention probability is:

[0077]

[0078] in: The importance score for parameter p; p is a single parameter in the model; In order to validate the dataset Expectation operation on; The loss function; Let |p| be the gradient of the loss function with respect to parameter p; |p| is the absolute value of parameter p. Let p be the retention probability of parameter p; This is the pruning threshold coefficient; The median of the importance scores for all parameters; The importance score for parameter q; This is the set of all model parameters.

[0079] The lightweight objective function of the reflection kernel model is:

[0080]

[0081] in: The overall lightweight loss for the reflection kernel model; For the sake of accuracy loss; This is due to dimensional loss; For delayed losses; This is the size loss weighting coefficient; is a delay loss weight coefficient; precision loss wherein is an original model output, is a compressed model output, is a square of L2 norm; size loss wherein is a compressed model size, is a target model size; delay loss wherein is a compressed model inference delay, is a target inference delay.

[0082] The reflection core generation module specifically implements steps including:

[0083] Model pruning and parameter quantization are performed on the benchmark twin model to remove redundant parameters and calculation paths;

[0084] The core inference logic and key parameters of the benchmark twin model are extracted to generate a lightweight reflection core model suitable for edge computing environments;

[0085] The generated reflection core model is tested for edge device compatibility to ensure stable operation in resource-constrained environments;

[0086] The reflection core model is securely distributed to each edge neuron through an encrypted transmission method.

[0087] The reflection core model retains the temporal mixing mechanism and channel mixing mechanism of the benchmark twin model, but optimizes and simplifies the network structure, adopts a recurrent inference mode, and realizes real-time inference capability with constant time complexity.

[0088] Through the above technical solution, the twin brain module realizes complete conversion from dam historical monitoring data to a lightweight reflection core model, providing intelligent support for the entire monitoring system and realizing the organic combination of cloud intelligence and edge computing.

[0089] Embodiment Two: Deployment and Online Monitoring of Edge Neurons

[0090] On the basis of embodiment one, this embodiment describes in detail the specific structure and implementation method of edge neuron in the dam safety monitoring data acquisition and transmission system of a hydropower station. As an intelligent terminal device deployed in the dam monitoring site, the edge neuron undertakes key functions such as receiving lightweight inference models generated by the cloud, real-time processing of on-site monitoring data, intelligent prediction analysis, calculation of state deviation indicators, and implementation of adaptive data transmission control. The design of the edge neuron fully considers the complexity and resource constraints of the on-site environment, and achieves efficient edge intelligent computing capability through a modular hardware and software architecture. The edge neuron includes model and data interface module, data preprocessing module, online twin prediction module, reflection deviation calculation module, hierarchical transmission control module, and local storage module, and other core functional components. The modules work together to form a complete edge intelligent monitoring system, which can realize highly intelligent data analysis and decision-making processing in resource-constrained edge environments, providing real-time, accurate, and reliable technical support for dam safety monitoring.

[0091] The model and data interface module, as the core interface component of the edge neuron, undertakes dual key functions: on the one hand, it is responsible for communication with the cloud twin brain module, receiving and managing lightweight reflection kernel models; on the other hand, it establishes data connections with various types of sensor devices on site, realizing real-time acquisition and preprocessing of multi-source monitoring data. This module uses high-reliability communication protocols and data processing mechanisms to ensure the security of model transmission and the accuracy of data acquisition.

[0092] In terms of model reception and management, this module realizes complete model lifecycle management functions. When the cloud twin brain module generates a new version of the reflection kernel model, the model and data interface module receives the model file through a secure network connection. To ensure the integrity and authenticity of the received model file, this module uses a set of strict integrity verification algorithms, whose mathematical expression is:

[0093]

[0094] This verification algorithm combines a cryptographic hash function and a cyclic redundancy check, generating an integrity verification value by calculating the XOR result of the SHA256 hash value of the received file and the CRC32 check value with a timestamp. The judgment condition of the verification process is:

[0095] ;

[0096] Where: is the received model file data; is the reception timestamp, used to prevent replay attacks; represents the bitwise XOR operation; is the expected verification value calculated and sent by the cloud in advance; to receive the difference between the time of reception and the time of transmission; is a preset timeout threshold for preventing the loading of expired models. Only when the verification value matches and the time difference is within the allowed range, the model file is considered valid and safe.

[0097] The rigorously verified reflective kernel model is safely loaded into the memory space of the edge neuron and completes the initialization configuration. This module is also responsible for version management of the model, maintaining information of the current active model and historical versions, supporting dynamic updating and rollback operations of the model. In terms of data interfaces, this module establishes reliable connections with various types of sensors on site, supporting multiple communication protocols and interface standards, including analog signal interfaces, digital communication interfaces, industrial bus interfaces, etc. Through these interfaces, this module can obtain real-time physical quantity actual observation values and corresponding environmental variable observation values of the dam structure. Physical quantity actual observation values mainly include key parameters such as deformation, seepage flow, stress value, and strain value that reflect the response characteristics of the dam structure, which directly reflect the current structural state and safety level of the dam. Environmental variable observation values include external condition parameters that affect the behavior of the dam, such as reservoir water level observation values, environmental temperature observation values, and rainfall observation values, which provide necessary environmental condition information for the intelligent prediction module.

[0098] The data preprocessing module is a key component in the edge neuron responsible for data cleaning, standardization, and format conversion. This module ensures that raw monitoring data from different sensors can be correctly understood and processed by the reflective kernel model. Due to the variety of sensors on site, the measured physical quantities have different dimensions, value ranges, and sampling characteristics. The data preprocessing module converts these heterogeneous raw data into standardized inputs of a unified format through a series of carefully designed processing steps. This module includes three core processing units: instance normalization unit, sequence blocking unit, and token generation unit. These three units work collaboratively in a pipeline manner to realize the complete conversion process from raw observation data to model input tokens.

[0099] The instance normalization unit, as the first link of data preprocessing, is specifically responsible for solving the problem of inconsistent data scales between different sensors. This unit independently normalizes the observation value sequence of each sensor channel, effectively eliminating the dimension difference and value range difference between different sensors, ensuring that data from different physical quantities can be processed on the same numerical scale. The instance normalization unit adopts a dynamic window adaptive strategy, which can automatically adjust the normalization parameters according to real-time data changes. Its core algorithm formula is: ;

[0100] ;

[0101] The normalized output is: ;

[0102] wherein: is the dynamic mean of the ith sensor at the tth time; is the dynamic mean of the ith sensor at the kth time; is the dynamic mean of the ith sensor at the kth time; is the dynamic update coefficient, controlling the retention degree of historical information; W is the sliding window length; is the summation operation from the kth time to the tth time; is the observation value of the ith sensor at the kth time; is the dynamic standard deviation of the ith sensor at the tth time; is the dynamic standard deviation of the ith sensor at the kth time; is the dynamic standard deviation of the ith sensor at the kth time; is the dynamic standard deviation of the ith sensor at the kth time; is the normalized output value of the ith sensor at the tth time; is the original observation value of the ith sensor at the tth time; is a small constant to prevent division by zero; is the learnable scaling parameter of the ith sensor; is the learnable translation parameter of the ith sensor.

[0103] The sequence blocking unit splits the long time sequence according to a fixed block length and an overlap step, generating short sequence blocks suitable for processing by the reflection kernel model. The token generation unit converts each sequence block into a standardized input token containing position encoding and feature encoding.

[0104] The online twin prediction module calls the recurrent inference mode of the reflection kernel model, and based on the observation value of the environmental variable at the current time and the historical state information, it predicts the theoretical normal value of each monitored physical quantity under the current working condition in real time through the recurrent inference mode with constant time complexity, and forms a dynamic reference trajectory with the sequence of the theoretical normal value.

[0105] The prediction algorithm of the online twin prediction module adopts an environment-adaptive state update mechanism:

[0106]

[0107] The prediction output calculation formula is:

[0108]

[0109] wherein: is the predicted state; is the dynamic forgetting coefficient; is the environmental variable vector; is the environment-state mapping function; is the theoretical normal value; is a weight matrix; is a bias vector.

[0110] The reflection deviation calculation module calculates the Euclidean distance between the physical quantity actual observation value and the theoretical normal value as the reflection deviation value, and includes a deviation statistical analysis unit for calculating the statistical characteristics of the reflection deviation value in a sliding time window, including mean, variance, change trend coefficient and cumulative characteristics.

[0111] The reflection deviation calculation uses the Euclidean distance algorithm:

[0112] Wherein: is a reflection deviation value; is a physical quantity actual observation value vector; is a theoretical normal value vector.

[0113] The deviation statistical analysis unit calculates the statistical characteristics in a preset sliding time window:

[0114] Average deviation:

[0115] Deviation variance:

[0116] Wherein: W is the length of the sliding window; is the average deviation; is the deviation variance.

[0117] The hierarchical transmission control module executes a three-level adaptive data transmission strategy according to the magnitude, duration and change rate of the reflection deviation value:

[0118] First level local record response: triggered when the reflection deviation value is below the first preset threshold, the reflection deviation value and the corresponding observation value are stored in the local storage module;

[0119] Second level abstract information transmission response: triggered when the reflection deviation value is continuously higher than the first preset threshold but lower than the second preset threshold, a data abstract containing deviation statistical characteristics, change trend analysis and time mark is generated for transmission;

[0120] Third level immediate early warning information transmission response: triggered when the reflection deviation value is instantaneously higher than the second preset threshold or the change rate exceeds the preset limit, the complete data record of the abnormal period is immediately intercepted for real-time transmission.

[0121] The local storage module is used to cache observation data, reflection deviation value and system operation log. This module uses a multi-level storage architecture, including cache, local database and backup storage, to ensure the reliability and traceability of data.

[0122] The hardware architecture of the edge neuron comprises:

[0123] An embedded computing unit: a low-power high-performance processor and a preset capacity of memory are used to run the reflection kernel model and data processing tasks;

[0124] A multi-channel data acquisition unit: supporting simultaneous access of analog signals, digital signals and pulse signals;

[0125] A local storage unit: providing a preset capacity of data cache space, supporting local backup and historical query of data;

[0126] A communication unit: supporting wireless network communication and wired network communication, with communication redundancy and fault switching capability;

[0127] A clock synchronization unit: high-precision time synchronization is realized through a satellite positioning system to ensure the time consistency of distributed monitoring data.

[0128] Through the above technical solutions, the edge neuron realizes intelligent data analysis and decision-making in a resource-limited edge environment, can calculate the reflection deviation value in real time and execute a hierarchical transmission strategy, providing reliable protection for dam safety monitoring.

[0129] Example three: implementation of closed-loop model evolution mechanism

[0130] This embodiment describes in detail the specific structure and implementation method of a closed-loop model evolution mechanism in a hydropower station dam safety monitoring data acquisition and transmission system. As the intelligent core component of the entire monitoring system, the closed-loop model evolution mechanism has the ability of autonomous learning and continuous optimization, can continuously learn new knowledge and patterns from the actual data generated during system operation, and feed these knowledge back to the system to realize continuous improvement of the performance of the monitoring system. The core idea of this mechanism is to establish a complete closed-loop feedback system from edge perception to cloud intelligence to edge deployment, realizing self-evolution and performance optimization of the system through data-driven way.

[0131] The design of the closed-loop model evolution mechanism fully considers the characteristics of long-term and complexity of dam safety monitoring. In the long-term monitoring process, the behavior pattern of the dam may change slowly, new abnormal patterns may appear, and external environmental conditions may change, which requires the monitoring system to have adaptability and learning ability. The mechanism establishes a complete data backflow channel to collect various data generated by the edge neuron in actual operation, including normal operation data, abnormal detection data and system response data, and then uses advanced machine learning technology to deeply analyze and mine these data, extracting valuable information for model optimization.

[0132] The closed-loop model evolution mechanism includes four main components: data collection and feedback subsystem, incremental learning processing subsystem, model version management subsystem, and automated deployment subsystem. The data collection and feedback subsystem is responsible for receiving operation data from various edge neurons. When the hierarchical transmission control module of the edge neuron detects the need to upload, it will transmit the corresponding data to the cloud end twin brain module through the network. These data mainly include two categories: one is the data summary generated by the second level summary information transmission response, which contains bias statistical characteristics, trend analysis and time marker, etc. The other is the complete data record generated by the third level immediate warning information transmission response, which contains detailed monitoring data and system response information during the abnormal period.

[0133] The incremental learning processing subsystem is the core component of the entire closed-loop mechanism. This subsystem receives the feedback data and processes it intelligently. First, the subsystem conducts a comprehensive quality assessment of the new data collected to ensure that the data used for learning has sufficient quality and reliability. Then, the system intelligently labels the collected data according to the event types it represents, mainly classified into three categories: normal changes, abnormal events, and device failures. Next, the system uses advanced incremental training algorithms to update the parameters of the existing baseline twin model and optimize its performance, while maintaining the memory of historical knowledge and learning new behavior patterns and abnormal features.

[0134] The incremental learning module, as the intelligent processing core of the closed-loop model evolution mechanism, is responsible for extracting valuable information from raw feedback data and using it for model optimization. This module first conducts a comprehensive quality assessment of the data returned from the edge neurons, establishing a complete multi-dimensional data quality inspection system. The data quality assessment process includes four main dimensions: integrity check to ensure that the uploaded data contains all necessary sensor readings and environmental variable information, avoiding the impact of missing data on learning effectiveness; consistency verification to identify potential sensor failures or abnormal readings by cross-verifying the data correlation between different sensors; time sequence continuity analysis to check whether the timestamp sequence of the data is missing, repeated or abnormally jumping, ensuring the correct time logic of the data; noise level evaluation to quantify the random disturbance component in the data through statistical analysis methods to filter out low-quality data with excessive noise.

[0135] The comprehensive score of data quality uses a multi-dimensional weighted evaluation algorithm, and its mathematical expression is:

[0136] ;

[0137] Where: is the comprehensive quality score, with a value range of 0 to 1; is the data set is the score on the i-th quality dimension; is the weight coefficient of the i-th dimension, reflecting the importance of this dimension to the overall quality; is the adjustment parameter, controlling the degree of punishment for deviating from the reference value; is the reference value of the i-th dimension, representing the expected quality level. This formula ensures that any serious quality problem in any one dimension will significantly affect the overall score, while the exponential term provides additional punishment for deviating from the reference value.

[0138] After the data quality assessment, the incremental learning module performs intelligent label annotation, which is a key step in data preprocessing. The label annotation process uses a semi-supervised learning method, combining domain expert knowledge and automated analysis algorithms, to accurately classify the collected data according to the event types it represents. The system classifies events into three main categories: normal change category, including changes in dam behavior caused by environmental conditions such as seasonal temperature changes, water level fluctuations, and climate changes, which are within the expected range and conform to physical laws; abnormal event category, including events that may affect dam safety but do not pose an immediate threat, such as minor structural deformation acceleration, abnormal increase in seepage flow, and stress distribution changes; and equipment failure category, including sensor faults or measurement system problems, such as sensor drift, communication interruption, and data acquisition anomalies.

[0139] The incremental learning module uses an advanced incremental training algorithm to intelligently update the parameters of the baseline twin model. The core goal of this algorithm is to learn new knowledge while maintaining the memory of historical knowledge, effectively avoiding catastrophic forgetting. The incremental training process uses a flexible weight consolidation strategy, which sets different update constraints for different parameters by calculating the importance of model parameters to historical task performance. The mathematical expression of its incremental training loss function is:

[0140] where: is the supervised learning loss of new data; is the importance weight of parameter i, controlling the update amplitude of the parameter; is the i-th diagonal element of the Fisher information matrix, quantifying the sensitivity of parameter i to the model output; is the i-th parameter of the current model; is the important parameter value obtained by training on historical data. This loss function ensures that important parameters do not deviate too far from their historical optimal values through the regularization term, thereby maintaining the model's performance on historical tasks.

[0141] ​The incremental learning module also adopts an experience replay strategy to further enhance the learning effect. This strategy mixes the newly collected data with samples selected from the historical representative database for intelligent training in a certain proportion, effectively preventing the model from overfitting to new data. The construction of the mixed training batch uses the following formula:

[0142]

[0143] Wherein: is the final mixed data batch for training; is the training batch composed of newly collected data; is the maintained historical representative database; is the number of replay samples selected from the historical database. The experience replay strategy selects the most representative and important historical samples for replay through an intelligent sampling algorithm, ensuring that the model learns new knowledge without forgetting important historical experience.

[0144] To further enhance the model's ability to identify abnormal patterns, the incremental learning module uses a contrast learning method to improve the model's discriminative performance by constructing positive and negative sample pairs. Contrast learning enables the model to better distinguish between normal and abnormal states, improving the accuracy and robustness of anomaly detection. The module also sets up a multi-level model performance evaluation system, including three main levels of prediction accuracy indicators, anomaly detection rate indicators, and false positive rate indicators, each level containing multiple specific evaluation indicators to ensure the quality and reliability of model updates.

[0145] The model version management module is responsible for maintaining complete model iteration records. This module uses a distributed version control design concept to generate a unique identifier for each model version and records detailed change information, including model structure parameters, performance indicators, training data summaries, and change descriptions. When the comprehensive score of the newly trained model in the multi-level performance evaluation system reaches the pre-set improvement standard, the system automatically triggers the version release process. The comprehensive performance evaluation uses a hierarchical weighted scoring mechanism, with the formula being:

[0146] Wherein: is the comprehensive performance score; is the weight coefficient of the lth layer; is the score of the lth layer. The three levels correspond to the prediction accuracy indicator, anomaly detection rate indicator, and false positive rate indicator, respectively, and the weight of each level is adjusted according to actual application requirements.

[0147] When the model version management module confirms the release of the new version, the automated deployment subsystem starts the generation and delivery process of the reflection core model. This process first calls the reflection core generation module to perform quantization compression and parameter optimization on the updated baseline twin model, generating a lightweight reflection core model suitable for edge computing environments. The model delivery process uses an incremental update mechanism, which only transmits the changed parameter part by calculating the parameter difference between the new version model and the current deployed version, significantly reducing the data volume of network transmission and improving the update efficiency. The delivery process uses encryption transmission and digital signature technology to ensure the security and integrity of the model file during transmission.

[0148] To ensure the effectiveness and safety of the system's continuous self-evolution ability, the closed-loop model evolution mechanism also implements a number of safeguard mechanisms. The data quality monitoring mechanism continuously monitors the quality trend of uploaded data, identifies and filters low-quality data, and prevents poor-quality data from negatively affecting model training. The model performance monitoring mechanism continuously tracks the performance of deployed models in actual operation, analyzes the performance trend of performance indicators, and promptly detects performance degradation problems. When performance abnormalities are detected, the model rollback and retraining process is automatically triggered. The security safeguard mechanism ensures that model updates do not introduce security risks or privacy leakage problems through multiple verification and audit processes, including model behavior consistency checks, security vulnerability scans, and privacy protection assessments.

[0149] Through the above complete technical solution, the closed-loop model evolution mechanism successfully realizes the continuous self-evolution ability of the hydropower dam safety monitoring system, enabling the entire system to continuously learn new knowledge and patterns and continuously improve monitoring accuracy, early warning ability, and system reliability during long-term operation, providing strong technical support for the long-term safe operation of the dam.

[0150] The above only describes the preferred embodiments of the present application, and it should be understood that the present application is not limited to the forms disclosed herein, and should not be considered as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concepts described herein, by the above teachings or related art or knowledge. Any modification and change made by those skilled in the art without departing from the spirit and scope of the present application shall be within the protection scope of the appended claims of the present application.

Claims

1. A data acquisition and transmission system for safety monitoring of hydropower station dams, characterized in that, include: The twin brain module, deployed on a cloud server, includes: The sequence neural network engine adopts a neural network architecture with dual equivalent representations of parallelism and recurrence. The architecture includes a time mixing module and a channel mixing module. The time mixing module performs weighted processing on the multi-source historical monitoring data of the dam through a learnable time decay vector. The channel mixing module is used to realize cross-variable coupling relationship modeling. During training, the sequence neural network engine uses a parallel mode to process the deformation data, seepage data, stress and strain data of the dam, as well as reservoir water level data, rainfall data, and ambient temperature data, to generate a benchmark twin model that can characterize the normal operating state of the dam. The reflection kernel generation module is used to quantize and compress the trained benchmark twin model and optimize its parameters, extract the core inference logic, generate a lightweight reflection kernel model with significantly reduced parameter size, and distribute it to the edge neurons via network communication. And at least one edge neuron deployed at the dam monitoring site, including: The model and data interface module is used to receive and load the reflection kernel model from the cloud, establish a data connection with the field sensors, and obtain the actual observed values ​​of physical quantities measured by the sensors and the corresponding environmental variable observation values ​​in real time. The data preprocessing module includes an instance normalization unit, a sequence segmentation unit, and a token generation unit. It is used to standardize the continuous observation sequence, segment it into blocks according to a preset length and step size, and convert the segmented data into block sequence tokens that conform to the input format of the reflection kernel model. The online twin prediction module calls the cyclic inference mode of the reflection kernel model. Based on the environmental variable observation values ​​and historical state information at the current moment, it predicts the theoretical normal values ​​of each monitored physical quantity under the current working condition in real time through a cyclic inference method with constant time complexity, and constructs the sequence of theoretical normal values ​​into a dynamic reference trajectory. The reflection deviation calculation module calculates the Euclidean distance between the actual observed value and the theoretical normal value of the physical quantity as the reflection deviation value, and includes a deviation statistical analysis unit for calculating the statistical characteristics of the reflection deviation value within a sliding time window, including mean, variance, trend coefficient and cumulative characteristics. The hierarchical transmission control module executes a three-level adaptive data transmission strategy based on the magnitude, duration, and rate of change of the reflection deviation value: when the reflection deviation value is lower than the first preset threshold, a first-level local recording response is triggered; when the reflection deviation value is continuously higher than the first preset threshold but lower than the second preset threshold, a second-level summary information transmission response is triggered; and when the reflection deviation value is instantaneously higher than the second preset threshold or the rate of change exceeds a preset limit, a third-level immediate warning information transmission response is triggered. The local storage module is used to cache observation data, reflection deviation values, and system operation logs; The system also includes a closed-loop model evolution mechanism. The incremental learning module receives the uploaded summary information and early warning information data, performs incremental training and parameter fine-tuning on the benchmark twin model, and maintains the iteration record through the model version management module. After the model is updated, the generation and distribution of the new version of the reflection kernel model are automatically triggered, realizing the system's continuous self-evolution capability.

2. The system according to claim 1, characterized in that, The temporal mixing module of the sequence neural network engine assigns different decay weights to data at different time steps in the historical monitoring data sequence, so that the influence weight of historical data further away from the current time is smaller, thereby effectively capturing long-term dependencies. The channel mixing module achieves correlation modeling between different monitoring variables by interacting and fusing information on the feature dimension; the sequence neural network engine adopts a parallel computing mode to improve training efficiency during the training phase and switches to a loop computing mode during the inference phase to achieve efficient real-time inference.

3. The system according to claim 1, characterized in that, The workflow of the reflection kernel generation module includes: pruning and parameter quantization of the trained baseline twin model to remove redundant parameters and computational paths; extracting the core inference logic and key parameters of the baseline twin model to generate a lightweight reflection kernel model suitable for edge computing environments; performing edge device compatibility testing on the generated reflection kernel model to ensure its stable operation in resource-constrained environments; and securely distributing the reflection kernel model to each edge neuron via encrypted transmission.

4. The system according to claim 1, characterized in that, The prediction mechanism of the online twin prediction module includes: receiving the current reservoir water level observation value, ambient temperature observation value, and rainfall observation value as input conditions; calculating the expected value of each monitored physical quantity at the next moment through cyclical inference of the reflection kernel model based on the environmental variable observation value and historical state information; forming a dynamic reference trajectory from the continuously predicted expected value sequence, which reflects the normal behavior pattern of the dam under the current environmental conditions; the prediction process has constant time complexity to ensure real-time response requirements.

5. The system according to claim 1, characterized in that, The reflection deviation calculation module quantifies the degree of system deviation by calculating the Euclidean distance between the actual observed value vector and the theoretical normal value vector; the deviation statistical analysis unit continuously tracks the change of reflection deviation value within a preset sliding time window and calculates its statistical characteristics, including average deviation, deviation variance, trend coefficient, and cumulative deviation, to assess the stability and abnormal development trend of the system state.

6. The system according to claim 1, characterized in that, The three-level response strategy of the hierarchical transmission control module is as follows: The first level is local recording response. When the reflection deviation value is lower than the first preset threshold, the reflection deviation value and the corresponding observation value are stored in the local storage module without data transmission to save communication resources; the second level is summary transmission response. When the reflection deviation value continuously exceeds the first preset threshold but is lower than the second preset threshold, a data summary containing deviation statistical characteristics, trend analysis and time stamp is generated and transmitted over the network in a low-priority manner. The third-level early warning transmission response immediately intercepts the complete data record of the abnormal period when the reflection deviation value exceeds the second preset threshold or the deviation change rate exceeds the preset safety limit, and transmits it in real time with the highest priority.

7. The system according to claim 1, characterized in that, The instance normalization unit of the data preprocessing module performs independent normalization processing on the observation sequence of each sensor channel to eliminate the differences in dimensions and numerical ranges between different sensors. The sequence segmentation unit divides the long sequence into short sequence blocks suitable for reflection kernel model processing by dividing the long sequence into blocks with fixed block length and overlap step size; the token generation unit converts each sequence block into a standardized input token containing position encoding and feature encoding.

8. The system according to claim 1, characterized in that, The closed-loop model evolution mechanism achieves continuous learning through the following steps: the hierarchical transmission control module uploads the data generated by the second and third level responses to the twin brain module; the incremental learning module performs quality assessment and labeling on the collected new data, distinguishing between normal changes, abnormal events, and equipment failures; and the incremental training algorithm is used to update the parameters of the baseline twin model, maintaining the memory of historical knowledge while learning new behavioral patterns. The model version management module records the performance metrics and changes for each update, and triggers a version release when the performance of the new model reaches the preset improvement standard.

9. The system according to claim 8, characterized in that, The incremental learning module employs an experience replay strategy, mixing newly collected data with historical representative data in a preset ratio for training to prevent the model from overfitting to new data; it uses a contrastive learning method to enhance the model's ability to identify abnormal patterns, improving the model's discriminative performance by constructing positive and negative sample pairs; and it sets up a multi-level model performance evaluation system, including prediction accuracy indicators, anomaly detection rate indicators, and false alarm rate indicators, to ensure the quality and reliability of model updates.

10. The system according to claim 1, characterized in that, The hardware architecture of the edge neuron includes: an embedded computing unit, employing a low-power, high-performance processor and a preset memory capacity, for running the reflection kernel model and data processing tasks; a multi-channel data acquisition unit, supporting simultaneous access of analog, digital, and pulse signals; a local storage unit, providing a preset data cache space to support local data backup and historical retrieval; a communication unit, supporting wireless network communication and wired network communication, with communication redundancy and fault switching capabilities; and a clock synchronization unit, achieving high-precision time synchronization through a satellite positioning system to ensure the time consistency of distributed monitoring data.

Images