Edge intelligence early warning node and system with incremental learning and entropy-driven collaborative capability

Abstract

The present application relates to the technical field of structural health monitoring, and in particular to an edge intelligent early warning node and system with incremental learning and entropy-driven collaborative capability. The hardware layer of the node comprises a SoC (System on Chip) and a sensor interface. The firmware layer comprises: a data preprocessing module for data preprocessing; a time series feature fusion network module loaded with a one-dimensional convolutional neural network (1D-CNN) and a gated recurrent unit (GRU) for outputting a probability distribution vector representing different risk categories; a risk level assessment module for assessing the risk level of the monitoring point; an entropy-driven communication decision module for screening effective data tensors as iterative training samples; and an online incremental learning module for locally updating the parameters of the one-dimensional convolutional neural network (1D-CNN) and the gated recurrent unit (GRU) in the time series feature fusion network module. The present application solves the problems of poor adaptability, low resource utilization efficiency, and low efficiency of edge-cloud collaborative learning closed loop of existing edge monitoring node models.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring technology, and particularly to the cross-application field of Internet of Things (IoT), edge computing and artificial intelligence (AI). Specifically, it relates to an edge intelligent early warning node and system with incremental learning and entropy-driven collaborative capabilities for monitoring construction risks in highways, bridges, roadbeds and other projects. Background Technology

[0002] Large infrastructure projects such as roads, bridges, and roadbeds face complex geological environments and dynamic loads during construction and operation, making them prone to safety risks such as slope instability, structural fatigue, and uneven settlement. Traditional monitoring methods often rely on manual inspections or automated monitoring schemes based on fixed thresholds. These schemes often suffer from problems such as high response delays, low early warning accuracy, and difficulty in adapting to dynamic environmental changes.

[0003] With the development of Internet of Things (IoT), edge computing, and artificial intelligence technologies, edge intelligence-based monitoring systems are gradually becoming an important direction for improving the real-time performance and intelligence of security early warnings. In such systems, pre-trained or lightweight AI models are typically deployed on edge nodes close to the sensor data source to enable local analysis and rapid response to structural health conditions.

[0004] For example, Chinese patent CN119249274A discloses an industrial equipment health monitoring and fault early warning system based on edge computing. This system collects industrial equipment monitoring data through edge nodes and uses edge computing resources to execute fault detection models locally, making real-time judgments and early warnings on the health status of the equipment. This helps to reduce data transmission latency and improve the overall system response efficiency.

[0005] Although existing technologies have introduced edge computing capabilities into health monitoring and early warning systems, several problems still need to be solved in practical engineering applications: 1. Poor environmental adaptability and insufficient model updates: Edge nodes typically use offline-trained static models for inference and analysis. When the monitored objects (such as construction environment, material properties, load conditions, etc.) undergo long-term and complex dynamic changes, the static models struggle to maintain high accuracy. Although existing incremental learning methods propose local model update schemes, the coordination of learning new and old knowledge within the entire system and the avoidance of catastrophic forgetting remain insufficiently addressed.

[0006] 2. Low resource utilization efficiency: As long-term unattended IoT edge nodes, their power consumption and resource scheduling strategies have a significant impact on the system lifecycle. Existing scheduling technologies (such as load-based scheduling methods) mostly focus on system resources, but fail to tightly couple resource scheduling with structural security risks, thus failing to significantly reduce power consumption during most "safe interval" periods.

[0007] 3. Blind Spots in Edge-Cloud Collaborative Communication: In an edge-cloud collaborative architecture, the interaction mode between edge nodes and the cloud directly affects learning efficiency and communication costs. Existing technologies often employ periodic or threshold-triggered upload mechanisms, which cannot effectively distinguish between high-value and low-value sample data, thus limiting the efficiency of the edge learning and cloud iterative closed loop.

[0008] Based on the above analysis, existing technologies still lack an edge intelligent early warning system that can continuously adapt to dynamic environmental changes, possess energy-efficient scheduling capabilities, and achieve value-driven edge-cloud collaborative learning. This invention addresses these shortcomings by proposing an edge intelligent early warning node and system with incremental learning and entropy-driven collaborative capabilities. Through an autonomous learning module and intelligent resource scheduling mechanism, it achieves efficient, low-consumption, and real-time response capabilities for structural health monitoring and real-time early warning. Summary of the Invention

[0009] This invention aims to overcome the problems of poor adaptability, low resource utilization efficiency, and inefficient edge-cloud collaborative learning closed loop of existing edge monitoring node models, and to provide an edge intelligent early warning node and system that is faster in response, more accurate in early warning, lower in power consumption, and can continuously self-optimize online.

[0010] The detailed technical solution of this invention is as follows: An edge intelligent early warning node with incremental learning and entropy-driven collaborative capabilities, the node comprising a hardware layer and a firmware layer, wherein the hardware layer includes a System-on-a-Chip (SoC) and a sensor interface electrically connected to the SoC; the firmware layer includes: The data preprocessing module is used to preprocess the raw multi-source time-series data of the monitoring points input through the sensor interface to obtain the corresponding data tensors; The temporal feature fusion network module is loaded with a one-dimensional convolutional neural network (1D-CNN) and a gated recurrent unit (GRU). The one-dimensional convolutional neural network (1D-CNN) is used to extract temporal and frequency domain features from the data tensor. The gated recurrent unit (GRU) is used to capture the long-term temporal dependencies between the temporal and frequency domain features and output a probability distribution vector to characterize different risk categories. The risk level assessment module is used to assess the risk level of the monitoring point based on the probability distribution vector and in combination with the dynamic risk rule base, wherein the risk level includes at least safety and alarm. The entropy-driven communication decision module is used to select valid data tensors that meet certain conditions from all data tensors as iterative training samples based on the Shannon entropy value of the probability distribution vector and in combination with the risk level; wherein, the selection conditions are set as the risk level being determined to be alarm and / or the Shannon entropy value being greater than a set threshold. The online incremental learning module performs local incremental updates on the parameters of the one-dimensional convolutional neural network (1D-CNN) and the gated recurrent unit (GRU) in the temporal feature fusion network module based on the received update instructions or iterative training samples.

[0011] According to a preferred embodiment of the present invention, the preprocessing operations of the original multi-source time series data include wavelet denoising, time window slicing, and Z-score normalization to generate standardized data tensors.

[0012] According to a preferred embodiment of the present invention, the Shannon entropy value of the probability distribution vector is calculated as follows: Assume that at time t, the probability distribution vector output by the temporal feature fusion network module is... Where N is the total number of risk categories, It is a hybrid model composed of a one-dimensional convolutional neural network (1D-CNN) and a gated recurrent unit (GRU) that predicts the probability that the current input data belongs to the i-th risk category, and satisfies Then, the Shannon entropy of this probability distribution vector is... The calculation formula is: .

[0013] According to a preferred embodiment of the present invention, the step of selecting valid data tensors from all data tensors as iterative training samples based on the Shannon entropy value of the probability distribution vector and in combination with the risk level specifically includes: When the risk level output by the risk level assessment module is an alarm or warning, the data tensor corresponding to the risk level is determined to be a valid data tensor and used as an iterative training sample. When the Shannon entropy value of the probability distribution vector corresponding to the data tensor is higher than a preset threshold, the data tensor is determined to be a valid data tensor and used as an iterative training sample.

[0014] According to a preferred embodiment of the present invention, the online incremental learning module uses an elastic weight fixation algorithm to locally and incrementally update the parameters of the one-dimensional convolutional neural network (1D-CNN) and the gated recurrent unit (GRU) in the temporal feature fusion network module.

[0015] According to a preferred embodiment of the present invention, the elastic weight solidification algorithm identifies model parameters important to the old task by calculating the Fisher information matrix of the old task, and restricts changes to model parameters important to the old task through a quadratic penalty term. The loss function during its update process... Represented as: ; in: Indicates the current model parameters; It is the standard loss function for the new task B; It is a hyperparameter used to adjust the intensity of the penalty for forgetting old tasks; This indicates that after the model has completed training on the old task A, the i-th parameter... The optimal value; These are the diagonal elements of the Fisher information matrix, representing the i-th parameter. The importance of old tasks.

[0016] According to a preferred embodiment of the present invention, the Fisher information matrix is ​​calculated as follows: ; in: Represents the Fisher information matrix; This represents the entire dataset in the old task A; The gradient operator represents the gradient over all model parameters. Find the partial derivatives; This represents the model's predicted probability of the correct answer; Represents the input sample The correct corresponding label; Indicates the use of old task datasets An input sample extracted from the sample; This represents the optimal parameters of the model after completing the training for the old task A.

[0017] According to a preferred embodiment of the present invention, the firmware layer of the edge intelligent early warning node further includes a dynamic resource scheduling module, which is used to control the power consumption mode of the entire node based on the risk level output by the risk level assessment module.

[0018] In another aspect of the present invention, an edge intelligent early warning system is provided, comprising an edge intelligent early warning node as described above with incremental learning and entropy-driven collaborative capabilities, the system further comprising a sensor and a cloud platform respectively communicatively connected to the edge intelligent early warning node; wherein, the sensor is used to acquire raw multi-source time-series data of the monitoring point and transmit it to the edge intelligent early warning node through a sensor interface; The edge intelligent early warning node is used for: The raw multi-source time-series data of the monitoring points input through the sensor interface are preprocessed to obtain the corresponding data tensor; The temporal and frequency domain features of the data tensor are extracted using a one-dimensional convolutional neural network (1D-CNN), and the long-term temporal dependencies between the temporal and frequency domain features are captured using a gated recurrent unit (GRU), outputting a probability distribution vector to characterize different risk categories. Based on the probability distribution vector and combined with the dynamic risk rule base, the risk level of the monitoring point is evaluated, and the risk level includes at least safety and alarm. Based on the Shannon entropy value of the probability distribution vector and combined with the risk level, valid data tensors that meet the conditions are selected from all data tensors as iterative training samples and uploaded to the cloud platform; wherein, the selection conditions are set as the risk level is determined to be alarm and / or the Shannon entropy value is greater than a set threshold. Based on update instructions or iterative training samples issued by the cloud platform, the parameters of the one-dimensional convolutional neural network (1D-CNN) and the gated recurrent unit (GRU) are updated locally and incrementally.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The edge intelligent early warning node of the present invention has stronger environmental adaptability and robustness: the online incremental learning module enables the node to continuously self-optimize locally, and can adapt to changes in the construction environment in real time, effectively avoiding model failure caused by environmental drift and catastrophic forgetting problems in the update process, and ensuring the stability and reliability of the long-term operation of the monitoring system.

[0020] (2) The edge intelligent early warning node of the present invention has significantly optimized system power consumption: the dynamic resource scheduling module directly links the power consumption of the node with the actual risk level, operates at the lowest power consumption during most "safe" periods, and only increases resource consumption when the risk increases, thereby realizing refined energy management and greatly extending the on-site deployment life of the node.

[0021] (3) The edge intelligent early warning node of the present invention has efficient and intelligent edge-cloud collaboration: an entropy-driven communication decision-making mechanism realizes the transformation from "blind reporting" to "value-driven accurate reporting". It ensures that only the highest value data (high risk or high uncertainty) is uploaded, which greatly saves communication bandwidth and provides high-quality, information-rich training samples for the rapid iteration of cloud models, thus accelerating the intelligent evolution of the entire system.

[0022] (4) The edge intelligent early warning node of the present invention has higher early warning accuracy: through a specially designed time-series feature fusion network module and combined with continuous online learning optimization, the system can extract key risk features more effectively from strong noise and non-stationary monitoring signals, thereby improving the sensitivity and accuracy of risk identification. Attached Figure Description

[0023] Figure 1 This is a functional block diagram of the edge intelligent early warning node described in this invention.

[0024] Figure 2 This is a schematic diagram of the architecture of the edge intelligent early warning system described in this invention.

[0025] Figure 3 This is a schematic diagram of the workflow of the edge intelligent early warning node described in this invention. Detailed implementation manners

[0026] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0028] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary implementation manners of the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0029] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0030] Embodiment 1 Refer Figure 1 , this embodiment provides an edge intelligent warning node 30 with the ability of incremental learning and entropy-driven collaboration. The edge intelligent warning node 30 includes a hardware layer 300 and a firmware layer 400.

[0031] Among them, the hardware layer 300 includes a system-on-chip (SoC) 320 and a sensor interface 310 electrically connected to the SoC 320. The SoC 320 serves as a computing core and integrates a low-power CPU and a dedicated neural processing unit (NPU). The sensor interface 310 is used to connect multiple sensors to collect raw multi-source time-series data of monitoring points, such as vibration, strain and other data.

[0032] Furthermore, the hardware layer 300 further includes a wireless communication module 330 that supports a low-power wide area network (such as NB-IoT / LTE-M), and a power management unit 340 responsible for fine-grained power supply control. The wireless communication module 330 is used to implement functions such as wireless communication and data transmission, and the power management unit 340 is used for self-power supply of the node.

[0033] The firmware layer 400 runs on the SoC 320 and includes: A data preprocessing module 410, which is used to preprocess the raw multi-source time-series data of the monitoring points input through the sensor interface 310 to obtain corresponding data tensors; The temporal feature fusion network module 420 is loaded with a one-dimensional convolutional neural network 1D-CNN and a gated recurrent unit GRU. The one-dimensional convolutional neural network 1D-CNN is used to extract temporal and frequency domain features from the data tensor. The gated recurrent unit GRU is used to capture the long-term temporal dependencies between the temporal and frequency domain features and output a probability distribution vector to characterize different risk categories. The risk level assessment module 440 is used to assess the risk level of the monitoring point based on the probability distribution vector and in combination with the dynamic risk rule base, wherein the risk level includes at least safety and alarm. The entropy-driven communication decision module 460 is used to select valid data tensors that meet certain conditions from all data tensors as iterative training samples based on the Shannon entropy value of the probability distribution vector and in combination with the risk level; wherein the selection conditions are set as the risk level is determined to be alarm and / or the Shannon entropy value is greater than a set threshold. The online incremental learning module 430 performs local incremental updates on the parameters of the one-dimensional convolutional neural network 1D-CNN and the gated recurrent unit GRU in the temporal feature fusion network module 420 based on the received update instructions or iterative training samples.

[0034] Specifically, the data preprocessing module 410 is used to perform noise reduction (e.g., using wavelet thresholding), time window slicing (e.g., in 2-second units), and Z-score normalization on the raw multi-source time-series data (such as vibration and strain) acquired through the sensor interface 310, generating a standard data tensor that can be used by the model.

[0035] Temporal Feature Fusion Network Module 420: A lightweight, edge-deployable hybrid model combining a one-dimensional convolutional neural network (1D-CNN) and a gated recurrent unit (GRU), deployed on an NPU. The front-end 1D-CNN layer automatically extracts multi-scale local temporal and frequency domain features from high-frequency temporal signals; the back-end GRU network captures the long-term temporal dependencies between these features, ultimately outputting a probability distribution vector representing different risk categories. In this embodiment, risk categories can include safety, minor subsidence, structural resonance, etc.

[0036] Risk level assessment module 440: Based on the probability vector output by the time-series feature fusion network module 420, and combined with a pre-built dynamic risk rule base containing historical status and trend analysis, it comprehensively assesses and quantifies the risk level of the current monitoring point. In this embodiment, the risk level can include four types: safe, attention, early warning, and alarm.

[0037] For example, the rule base can define: if the probability of structural resonance in the risk category is greater than the set threshold of 0.7, then its risk level is immediately determined to be an alarm; if the probability of minor settlement in the risk category is greater than the set threshold of 0.4 for 5 consecutive periods, then its risk level is determined to be a warning, thus achieving a more reliable risk assessment than a single threshold.

[0038] Entropy-Driven Communication Decision Module 460: This module is the core of value-driven communication. It quantifies the uncertainty of the mixture model's prediction of the current input by calculating the Shannon entropy of the probability distribution vector output by the temporal feature fusion network module 420; the specific calculation formula is as follows: Suppose that at time t, the probability distribution vector output by the temporal feature fusion network module 420 is... Where N is the total number of risk categories, It is the probability that the model predicts the current data belongs to the i-th risk category, and satisfies... .

[0039] Then, the Shannon entropy of this probability distribution vector The calculation formula is: ; in: It represents the logarithm to the base 2, and the unit of entropy is bits.

[0040] Understandably, when the mixture model's predictions are highly deterministic, for example, when the output probability distribution vector is... When the probability distribution is concentrated on a certain category, the entropy value is... It will be very low, close to 0.

[0041] When the model prediction is highly uncertain, for example, if there are 3 risk categories, and the output probability distribution vector is... When the probability is uniformly distributed across all categories, the entropy value is... It will reach its maximum value. This indicates that the model is "confused" by the current input data and is unable to make a clear classification.

[0042] By calculating this entropy value, this module can accurately quantify the "degree of confusion" of the model, thus providing a reliable mathematical basis for subsequent screening of high-value samples.

[0043] A high entropy value indicates that the mixture model is confused by the current data and unable to make a definitive judgment. This entropy-driven communication decision module 460 decides whether to trigger communication based on a dual criterion of risk level and Shannon entropy value. Criterion 1 (High-Risk Trigger): When the risk level assessment module 440 outputs an "alarm" or "warning" level, regardless of the Shannon entropy value, data upload must be triggered immediately. This data will then be used as iterative training samples to optimize the hybrid model in the temporal feature fusion network module 420. These samples are positive samples for the model, representing the most crucial data for training the model to correctly identify dangerous patterns. Using these samples for model iteration strengthens the model's memory and recognition of known risk features, essentially "reviewing" the model to ensure it doesn't miss any critical moments. Therefore, the data uploaded during alarm / warning events is a core component of high-value new samples.

[0044] Criterion Two (High Uncertainty Trigger): When the risk level is "Safe" or "Concern," but the Shannon entropy value is higher than a preset threshold (e.g., 0.8), data upload will be actively triggered. This type of "high entropy, low risk" data is also an "unknown but important" sample for the hybrid model, serving as the most valuable nourishment for driving cloud-based model iteration. It will also be used as training samples to optimize the hybrid model in the temporal feature fusion network module 420. These samples represent "unknown" or "fuzzy" areas outside the model's current cognitive boundaries. While the model doesn't classify them as high-risk, it is "confused" about their classification. This usually indicates the emergence of new operating conditions, new sources of interference, or insufficient generalization ability of the model. Using these samples for model iteration helps the model explore new knowledge, expand its cognitive boundaries, and fill gaps in its capabilities. They are the key nourishment driving the model from the "known" to the "unknown," achieving continuous self-optimization.

[0045] Online incremental learning module 430: Employs the Elastic Weight Consolidation (EWC) algorithm to locally and incrementally update the model parameters of the temporal feature fusion network module 420 when receiving update instructions or iterative training samples from the cloud platform.

[0046] EWC identifies weights crucial to prior knowledge by calculating the Fisher Information Matrix for old tasks and applies a secondary penalty to changes in these weights when training new tasks. This allows the hybrid model to effectively avoid catastrophic forgetting of already mastered risk patterns (old knowledge) when learning new situations (new knowledge).

[0047] Furthermore, the elastic weight fixation algorithm employed by the online incremental learning module 430 is based on the core idea of ​​using a quadratic penalty term to limit changes to model parameters (weights) important for older tasks when learning a new task. Its loss function during the update process... It can be represented as: ; in: Indicates the current model parameters; It is the standard loss function for the new task (Task B). In this scheme, it is the cross-entropy loss calculated on the new high-value samples (e.g., the 5 "new construction disturbance" samples issued by the cloud platform). It is a hyperparameter used to adjust the strength of the penalty for forgetting old tasks. The larger the value, the more the model tends to retain old knowledge; This indicates that after the model has completed training on the old task (Task A), the i-th parameter... The optimal value; These are the diagonal elements of the Fisher information matrix, representing the i-th parameter. The importance of old tasks.

[0048] Fisher Information Matrix The calculation is as follows: The Fisher information matrix measures the impact of small changes in model parameters on the model's output distribution. For old task A, it is calculated as the expectation over all old task data: ; in: Represents the Fisher information matrix; This represents the entire dataset in the old task A; The gradient operator represents the gradient over all model parameters. Find the partial derivatives; This represents the model's predicted probability of the correct answer; Represents the input sample The correct corresponding label; Indicates the use of old task datasets An input sample extracted from the sample; This represents the optimal parameters of the model after completing the training for the old task A.

[0049] In practice, the diagonal approximation is usually used, that is, only each parameter is calculated. . The larger the value, the more significant the parameter. The more significant the impact on the prediction results of old tasks, the more important they are.

[0050] The algorithm process based on the sample data of this scheme is as follows: 1) Initial state (learning from old task A): It is assumed that the model has been trained on a large amount of historical data (old task A, such as identifying "structural resonance") and the optimal parameters have been obtained. .

[0051] 2) Calculate Importance (FIM): The online incremental learning module 430 uses the data of the old task A to calculate the Fisher information corresponding to each parameter corresponding to the Fisher information . For example, it is calculated that the values corresponding to the weights of some CNN convolutional kernels and the weights of GRU gated units are very high, indicating that these parameters are crucial for identifying "structural resonance".

[0052] 3) Receive the new task B: The node receives 5 "new construction disturbance" samples sent by the cloud platform, forming the new task B.

[0053] 4) Incremental update: The module starts training on these new samples. When calculating the loss, the above combined loss function is used : ① This term drives the model parameters to update in the direction of correctly classifying the "new construction disturbance" samples; ② The penalty term will also take effect simultaneously: a) For those parameters with very high values (such as the key parameters for identifying "structural resonance"), any deviation from their old optimal values will be penalized, so as to keep them stable; b) For those parameters with very low values and non-importance, the penalty term is very small, allowing them to freely adjust to adapt to the new task B.

[0054] 5) Complete the update: After the training is completed, the obtained new parameters can identify the new "construction disturbance" and basically retain the ability to identify "structural resonance", thus effectively avoiding catastrophic forgetting.

[0055] Furthermore, the firmware layer 400 also includes a dynamic resource scheduling module 450: a policy engine based on the finite state machine FSM. This module switches the power consumption mode of the entire node according to the four states of "safe", "concerned", "warning", and "alarm" output by the risk level assessment module 440.

[0056] For example, in the "safe" state, the dynamic resource scheduling module 450 sends instructions to the power management unit 340 through the SoC system-on-chip 320, reduces the sampling frequency of the sensor to the lowest, reduces the SoC main frequency to the sleep level, and turns off the radio frequency unit of the wireless communication module 330 to achieve extreme power saving.

[0057] Example 2 See Figure 2This embodiment provides an edge intelligent early warning system 100, which includes an edge intelligent early warning node 30 with incremental learning and entropy-driven collaborative capabilities as described above. The edge intelligent early warning system 100 also includes a sensor 20 and a cloud platform 40 that are respectively communicatively connected to the edge intelligent early warning node 30.

[0058] The sensor 20 is used to acquire the original multi-source time-series data of the monitoring point and transmit it to the edge intelligent early warning node 30 through the sensor interface 310. The edge intelligent early warning node 30 is used for: The raw multi-source time-series data of the monitoring points input through the sensor interface 310 are preprocessed to obtain the corresponding data tensor; The temporal and frequency domain features of the data tensor are extracted using a one-dimensional convolutional neural network (1D-CNN), and the long-term temporal dependencies between the temporal and frequency domain features are captured using a gated recurrent unit (GRU), outputting a probability distribution vector to characterize different risk categories. Based on the probability distribution vector and combined with the dynamic risk rule base, the risk level of the monitoring point is evaluated, and the risk level includes at least safety and alarm. Based on the Shannon entropy value of the probability distribution vector and combined with the risk level, valid data tensors that meet the conditions are selected from all data tensors as iterative training samples and uploaded to the cloud platform 40; wherein, the selection conditions are set as the risk level is determined to be alarm and / or the Shannon entropy value is greater than a set threshold. Based on the update instructions or iterative training samples issued by the cloud platform 40, the parameters of the one-dimensional convolutional neural network 1D-CNN and the gated recurrent unit GRU are updated locally and incrementally.

[0059] The effectiveness of the present invention will be verified by using specific application examples below.

[0060] The application scenario is: intelligent risk early warning during bridge construction.

[0061] In this application scenario, the edge intelligent early warning system 100 is deployed on a bridge monitoring object 10 undergoing cantilever construction. Multiple sensors 20 (including MEMS accelerometers and fiber Bragg grating strain gauges) are deployed on the piers and key parts of the completed beam segments. These sensors transmit the collected vibration and strain data as sensor data stream 51 in real time to the edge intelligent early warning nodes 30 deployed near the piers via wired connections.

[0062] Edge intelligent early warning node 30 is responsible for local real-time analysis and risk assessment of data. When the edge intelligent early warning node 30 needs to interact with the cloud platform 40, it will send alarm data and high-entropy data upload data stream 52 to the remote cloud platform 40 via the 5G / NB-IoT network. After analyzing massive amounts of data and retraining the model, the cloud platform 40 can remotely optimize the edge intelligent early warning node 30 by issuing model update and control commands via data stream 53.

[0063] The following combination Figure 2 and Figure 3 This paper details a complete working cycle of the edge intelligent early warning node 30.

[0064] Step S10: System initialization.

[0065] After the edge intelligent early warning node 30 is powered on, the SoC on-chip system 320 executes the firmware program. The system initializes, including: loading the initial weights of the temporal feature fusion network module 420 (i.e., a hybrid model of 1D-CNN+GRU) deployed on the NPU; configuring the default sampling frequency of the sensor interface 310 to 100Hz; and putting the wireless communication module 330 into a low-power standby mode through the power management unit 340.

[0066] Steps S20-S40: Data acquisition, preprocessing, and model inference.

[0067] Steps S20-S30: The SoC on-chip system 320 acquires data from multiple sensors through the sensor interface 310. The data preprocessing module 410 takes over the raw data, performs wavelet denoising, slices it into 2-second time windows, performs Z-score normalization, and finally generates a standardized data tensor.

[0068] Step S40: The tensor is fed into the temporal feature fusion network module 420 for inference. After NPU acceleration, the temporal feature fusion network module 420 outputs a probability distribution vector, such as [0.95, 0.04, 0.01], which represent the probabilities of the data at the current monitoring point belonging to the three categories of "safe", "concern (minor settlement)" and "alarm (structural resonance)".

[0069] Steps S50-S63: Risk-driven parallel decision-making and resource scheduling.

[0070] Step S50: The risk level assessment module 440 receives the probability distribution vector [0.95, 0.04, 0.01]. Based on the internal dynamic risk rule base, for example, if the prediction results for three consecutive periods are all "safe", the risk level assessment module 440 will finally output a clear risk level: "safe".

[0071] Parallel decision-making begins, and the process is divided into two parts: Resource scheduling flow step S61: The dynamic resource scheduling module 450 receives the "safety" level. Its internal finite state machine switches to the "low-power state". The dynamic resource scheduling module 450 immediately generates control commands and sends them to the power management unit 340 through the SoC on-chip system 320. Accordingly, the power management unit 340 reduces the main frequency of the SoC on-chip system 320 to the minimum and reduces the sampling frequency of the sensor interface 310 to 50Hz to save power to the maximum extent.

[0072] Communication decision flow steps S62-S63: At the same time, the entropy-driven communication decision module 460 calculates the Shannon entropy of the probability distribution vector [0.95, 0.04, 0.01] to obtain a Shannon entropy value, for example, 0.25.

[0073] Step S62: The risk level assessment module 440 determines whether the risk level (known as the safety level) is "alarm," and if not, proceeds to step S63. The entropy-driven communication decision module 460 determines whether the calculated Shannon entropy value of 0.25 is greater than the preset entropy threshold (e.g., 0.8), and if not, proceeds to step S70. Decision result: Neither of the two communication trigger conditions is met, therefore step S70 will not be executed in this cycle, i.e., no data will be uploaded to the cloud platform.

[0074] Steps S80-S90: Monitoring and online incremental learning.

[0075] Steps S80-S81: The process converges to step S80. In low-power mode, the edge intelligent early warning node 30 periodically and briefly wakes up the wireless communication module 330 to listen for downlink instructions (data stream 53) from the cloud platform 40. In step S81, if no instruction is received, the process directly returns to step S20 via the "No" branch to start the next monitoring cycle in low-power mode.

[0076] Scenario Assumption: Suppose that during a certain period, due to the passage of large construction machinery, the edge intelligent early warning node 30 detects an unprecedented vibration pattern. At this time, the output of the temporal feature fusion network module 420 becomes a probability distribution vector [0.4, 0.3, 0.3], and the hybrid model exhibits great uncertainty. The risk level remains "safe," but the entropy value calculated by the entropy-driven communication decision module 460 (e.g., 1.58) is much higher than the entropy threshold of 0.8. Therefore, in step S63, the judgment is "yes," triggering step S70, which uploads this "high entropy" data to the cloud platform 40.

[0077] Cloud-based response and learning loop: After aggregating such "high-entropy" data uploaded by multiple edge intelligent early warning nodes 30, the cloud platform 40 identifies it as a "new type of construction disturbance." The cloud platform 40 can choose to package the five most representative samples and issue incremental learning instructions through data stream 53.

[0078] Step S90: The edge intelligent early warning node 30 receives the instruction in step S81, determines it as "yes," and the process proceeds to step S90. The online incremental learning module 430 is activated. It uses the EWC algorithm to fine-tune the weights of the hybrid model of the local temporal feature fusion network module 420 using these 5 new samples. The EWC algorithm ensures that while learning to identify "new types of construction disturbances," the model will not forget how to identify previously learned risk patterns such as "structural resonance." After learning is complete, the updated hybrid model is solidified. The process returns to step S20, and the edge intelligent early warning node 30 will continue its monitoring task with a more powerful model that is more adapted to the current environment.

[0079] Through the above detailed implementation methods, the system proposed in this invention demonstrates a complete, intelligent, and efficient closed-loop workflow, accurately solving technical problems such as poor model adaptability, high power consumption, and inefficient edge-cloud collaboration in existing technologies.

[0080] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. An edge intelligent early warning node with incremental learning and entropy-driven collaborative capabilities, characterized in that, The node includes a hardware layer and a firmware layer, wherein the hardware layer includes a SoC (System-on-a-Chip) and a sensor interface electrically connected to the SoC; the firmware layer includes: The data preprocessing module is used to preprocess the raw multi-source time-series data of the monitoring points input through the sensor interface to obtain the corresponding data tensors; The temporal feature fusion network module is loaded with a one-dimensional convolutional neural network (1D-CNN) and a gated recurrent unit (GRU). The one-dimensional convolutional neural network (1D-CNN) is used to extract temporal and frequency domain features from the data tensor. The gated recurrent unit (GRU) is used to capture the long-term temporal dependencies between the temporal and frequency domain features and output a probability distribution vector to characterize different risk categories. The risk level assessment module is used to assess the risk level of the monitoring point based on the probability distribution vector and in combination with the dynamic risk rule base. The risk level includes at least safety and alarm. The entropy-driven communication decision module is used to select valid data tensors that meet certain conditions from all data tensors as iterative training samples based on the Shannon entropy value of the probability distribution vector and in combination with the risk level; wherein, the selection conditions are set as the risk level being determined to be alarm and / or the Shannon entropy value being greater than a set threshold. The online incremental learning module performs local incremental updates on the parameters of the one-dimensional convolutional neural network (1D-CNN) and the gated recurrent unit (GRU) in the temporal feature fusion network module based on the received update instructions or iterative training samples.

2. The edge intelligent early warning node with incremental learning and entropy-driven collaborative capabilities according to claim 1, characterized in that, The preprocessing operations of the original multi-source time series data include wavelet denoising, time window slicing, and Z-score normalization to generate standardized data tensors.

3. The edge intelligent early warning node with incremental learning and entropy-driven collaborative capabilities according to claim 1, characterized in that, The Shannon entropy of the probability distribution vector is calculated as follows: Assume that at time t, the probability distribution vector output by the temporal feature fusion network module is... Where N is the total number of risk categories, It is a hybrid model composed of a one-dimensional convolutional neural network (1D-CNN) and a gated recurrent unit (GRU) that predicts the probability that the current input data belongs to the i-th risk category, and satisfies ; Then, the Shannon entropy of this probability distribution vector The calculation formula is: 。 4. The edge intelligent early warning node with incremental learning and entropy-driven collaborative capabilities according to claim 1, characterized in that, The process of selecting valid data tensors from all data tensors as iterative training samples based on the Shannon entropy value of the probability distribution vector and in conjunction with the risk level includes: When the risk level output by the risk level assessment module is an alarm or warning, the data tensor corresponding to the risk level is determined to be a valid data tensor and used as an iterative training sample. When the Shannon entropy value of the probability distribution vector corresponding to the data tensor is higher than a preset threshold, the data tensor is determined to be a valid data tensor and used as an iterative training sample.

5. The edge intelligent early warning node with incremental learning and entropy-driven collaborative capabilities according to claim 1, characterized in that, The online incremental learning module uses an elastic weight fixation algorithm to locally and incrementally update the parameters of the one-dimensional convolutional neural network (1D-CNN) and gated recurrent unit (GRU) in the temporal feature fusion network module.

6. The edge intelligent early warning node with incremental learning and entropy-driven collaborative capabilities according to claim 5, characterized in that, The elastic weight solidification algorithm identifies model parameters important to the old task by calculating the Fisher information matrix of the old task, and restricts changes to these parameters through a quadratic penalty term. Its loss function during the update process... Represented as: ； in: Indicates the current model parameters; It is the standard loss function for the new task B; It is a hyperparameter used to adjust the intensity of the penalty for forgetting old tasks; This indicates that after the model has completed training on the old task A, the i-th parameter... The optimal value; These are the diagonal elements of the Fisher information matrix, representing the i-th parameter. The importance of old tasks.

7. The edge intelligent early warning node with incremental learning and entropy-driven collaborative capabilities according to claim 6, characterized in that, The Fisher information matrix is ​​calculated as follows: ； in: Represents the Fisher information matrix; This represents the entire dataset in the old task A; The gradient operator represents the gradient over all model parameters. Find the partial derivatives; This represents the model's predicted probability of the correct answer; Represents the input sample The correct corresponding label; Indicates the use of old task datasets An input sample extracted from the sample; This represents the optimal parameters of the model after completing the training for the old task A.

8. The edge intelligent early warning node with incremental learning and entropy-driven collaborative capabilities according to claim 1, characterized in that, The firmware layer of the edge intelligent early warning node also includes a dynamic resource scheduling module, which is used to control the power consumption mode of the entire node based on the risk level output by the risk level assessment module.

9. An edge intelligent early warning system, comprising an edge intelligent early warning node possessing incremental learning and entropy-driven collaborative capabilities as described in any one of claims 1 to 8, characterized in that, The system also includes sensors and a cloud platform that are communicatively connected to the edge intelligent early warning node; wherein, the sensors are used to acquire raw multi-source time-series data of the monitoring points and transmit them to the edge intelligent early warning node through the sensor interface; The edge intelligent early warning node is used for: The raw multi-source time-series data of the monitoring points input through the sensor interface are preprocessed to obtain the corresponding data tensor; The temporal and frequency domain features of the data tensor are extracted using a one-dimensional convolutional neural network (1D-CNN), and the long-term temporal dependencies between the temporal and frequency domain features are captured using a gated recurrent unit (GRU), outputting a probability distribution vector to characterize different risk categories. Based on the probability distribution vector and combined with the dynamic risk rule base, the risk level of the monitoring point is evaluated, and the risk level includes at least safety and alarm. Based on the Shannon entropy value of the probability distribution vector and combined with the risk level, valid data tensors that meet the conditions are selected from all data tensors as iterative training samples and uploaded to the cloud platform; wherein, the selection conditions are set as the risk level is determined to be alarm and / or the Shannon entropy value is greater than a set threshold. Based on update instructions or iterative training samples issued by the cloud platform, the parameters of the one-dimensional convolutional neural network (1D-CNN) and the gated recurrent unit (GRU) are updated locally and incrementally.

Images