Intelligent early warning method and system based on fusion of visual features and biochemical parameters of activated sludge in aeration tank

Abstract

The application discloses an intelligent early warning method and system based on multi-dimensional visual features of activated sludge in an aeration tank and fusion of key biochemical parameters. The method realizes real-time collection of core key biochemical parameters and auxiliary biochemical parameters through underwater and water surface multi-channel cameras, carries out deep fusion and intelligent analysis on multi-source heterogeneous data by using a multi-modal deep learning model, and automatically identifies various process abnormal patterns in the operation of the aeration tank. The model is trained in a supervised learning mode and supports incremental learning to adapt to working condition changes. The identified abnormal patterns are intelligently checked and dynamically maintained in combination with an expert rule base constructed based on expert experience, historical data and industry standards, and finally, graded early warning and corresponding operation suggestions are generated. The application overcomes the defects of single data source dependence and subjectivity of manual judgment in traditional early warning methods, improves the accuracy, timeliness and intelligent level of aeration tank abnormal early warning, and effectively reduces the operation risk and energy consumption of a sewage treatment plant.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation monitoring technology for wastewater treatment processes, specifically to an intelligent early warning method and system based on the fusion of visual characteristics and biochemical parameters of activated sludge in aeration tanks. Background Technology

[0002] The activated sludge process is currently the most widely used biological treatment technology in urban wastewater treatment plants. As the core treatment unit of the activated sludge process, the operation of the aeration tank directly determines the effluent quality and overall treatment efficiency of the wastewater treatment system. The operation of the aeration tank is affected by a combination of factors, including fluctuations in influent water quality and quantity, changes in environmental conditions, and adjustments to operating parameters. This can easily lead to various process abnormalities such as sludge bulking, sludge disintegration, over- or under-aeration, organic load shocks, and impaired nitrification. If these issues are not detected and addressed in a timely manner, they can result in effluent quality exceeding standards or even process failure.

[0003] Current early warning methods for aeration tank anomalies mainly rely on on-site inspections and experience-based judgment by operators, periodic water quality testing in laboratories, or fixed threshold alarms based on single sensor parameters. These methods have several limitations in practical applications: First, they suffer from significant warning lag. Long laboratory testing cycles fail to reflect real-time process dynamics, and by the time anomalies are detected, significant water quality deterioration and process imbalances have often occurred, leading to high costs and difficulties in handling them. Second, judgment is highly subjective. Manual inspections and experience-based judgments are greatly influenced by individual factors such as personnel expertise and work status, making it difficult to consistently identify early, subtle anomalies. Third, monitoring dimensions are limited. Sensor-based threshold alarms only set thresholds for individual parameters, failing to fully explore the interrelationships between multiple parameters or consider the visual characteristics of activated sludge, resulting in high false alarm and false negative rates and insufficient accuracy. Fourth, the level of intelligence is low. They cannot effectively integrate unstructured visual information with structured sensor data, lack deep learning and adaptive capabilities for complex process correlations, and are poorly adaptable to dynamic changes in operating conditions and new anomaly patterns. Summary of the Invention

[0004] To address the problems of delayed early warning, strong subjectivity, single monitoring dimensions, and insufficient intelligence in the existing technologies, this invention provides an intelligent early warning method and system based on the fusion of multi-dimensional visual features and key biochemical parameters of activated sludge in aeration tanks. Through deep fusion and intelligent analysis of multi-source heterogeneous data, combined with expert rule verification, early, accurate, and graded early warnings of process anomalies in aeration tanks can be achieved.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: On the one hand, this invention provides an intelligent early warning method based on the fusion of multi-dimensional visual features and key biochemical parameters of activated sludge in an aeration tank, including the following steps: Step 1, data acquisition and preprocessing. Visual data of the aeration tank is acquired in real time using underwater and surface cameras, and time-series data of key biochemical parameters of the aeration tank are acquired in real time using online sensors. The visual data and time-series data are preprocessed separately. Visual data acquisition includes acquiring images of activated sludge flocs using underwater cameras and images of surface foam and bubble distribution using surface cameras. Key biochemical parameters include core key biochemical parameters and optional key auxiliary biochemical parameters. The core key biochemical parameters are dissolved oxygen, oxidation-reduction potential, and pH value; the key auxiliary biochemical parameters are one or more of ammonia nitrogen and nitrate nitrogen. Preprocessing of the visual data includes denoising, image enhancement, format conversion, frame rate normalization, and timestamp alignment. Preprocessing of the time-series data includes outlier removal, smoothing, timestamp alignment, and missing value imputation.

[0006] The second step is visual feature extraction. Using image processing and computer vision techniques, multi-dimensional visual features are extracted and quantified from the preprocessed visual data to form structured visual feature vectors. Visual feature extraction can combine traditional image processing algorithms with deep learning feature encoding techniques. Multi-dimensional visual features include the size, shape, density, compactness, contour complexity, color, motion trajectory, and settling velocity of activated sludge flocs; the quantity, size, color, morphology, stability, and coverage of surface foam; and one or more of the following: bubble size, density, uniformity, and rising speed.

[0007] The third step is deep learning-based anomaly pattern recognition. Visual feature vectors and time-series data of key biochemical parameters are input into a deep neural network model for multimodal feature learning and fusion, outputting preliminary process anomaly categories and their confidence levels. The deep neural network model is trained using supervised learning and supports incremental learning or periodic retraining to adapt to dynamic changes in operating conditions. The deep neural network model employs a two-stream network structure, with one branch using a convolutional neural network or visual transformer to process visual feature data, and the other branch using a long short-term memory network, gated recurrent units, or time-series transformer to process time-series data. Multimodal feature fusion can be achieved through feature splicing, attention-based fusion, or gated unit fusion. Identifiable process anomaly categories include one or more of the following: early signs of activated sludge bulking, signs of activated sludge disintegration and aging, insufficient or excessive aeration, signs of organic load shock, signs of impaired nitrification and denitrification functions, and signs of sludge poisoning.

[0008] The fourth step involves intelligent verification, tiered judgment, and early warning issuance. The system receives the initial anomaly category and confidence level, combines real-time absolute values, trends, and rates of change of key biochemical parameters, and utilizes a pre-defined expert rule base to intelligently verify and correct the initial judgment. Ultimately, the anomaly type and severity level are confirmed, and tiered early warning information and specific operational recommendations are issued. The expert rule base is built based on process expert experience, historical operating data, and industry emission standards, and supports dynamic maintenance and optimization. When the initial judgment of the deep neural network model conflicts with the expert rules, or when the model's confidence level is insufficient, the initial judgment is adopted first, or the final decision is made in conjunction with the expert rule judgment. Tiered early warning information includes three levels: Attention, Warning, and Emergency, corresponding to yellow, orange, and red warning indicators, respectively.

[0009] On the other hand, this invention provides an intelligent early warning system based on the fusion of multi-dimensional visual features and key biochemical parameters of activated sludge in an aeration tank. The system includes a data acquisition and preprocessing module, a visual feature extraction module, a deep learning anomaly pattern recognition module, an intelligent verification, grading, judgment, and early warning release module, and a data interface module. The data acquisition and preprocessing module is signal-connected to both the visual feature extraction module and the deep learning anomaly pattern recognition module, used to transmit preprocessed visual data to the visual feature extraction module and preprocessed key biochemical parameter time series data to the deep learning anomaly pattern recognition module. The visual feature extraction module is signal-connected to the deep learning anomaly pattern recognition module, used to transmit the extracted visual feature vectors to the deep learning anomaly pattern recognition module. The deep learning anomaly pattern recognition module is signal-connected to the intelligent verification, grading, judgment, and early warning release module, used to transmit the output preliminary process anomaly category and confidence level to the intelligent verification, grading, judgment, and early warning release module for verification, correction, and grading determination. The intelligent verification, grading, judgment, and early warning release module is signal-connected to the data interface module, used to transmit grading early warning information and operational suggestions to the data interface module. The data interface module is used to enable communication between the system and the field control system or cloud computing platform. The system can be deployed on edge computing devices or cloud computing platforms and provide users with early warning information and operation suggestions through web interface and mobile application.

[0010] Compared with the prior art, the present invention has the following beneficial effects: First, the timeliness of early warnings is significantly improved. Through real-time fusion analysis of visual features and biochemical parameters, early subtle signs and parameter changes in process anomalies can be captured, significantly advancing the warning time and providing operators with sufficient intervention window to prevent further deterioration of the anomaly.

[0011] Second, the accuracy of early warnings has been significantly improved. Deep learning models can uncover complex nonlinear relationships between multi-source data. Combined with expert rule verification and correction mechanisms and dynamic optimization mechanisms, the probability of false alarms and missed alarms has been effectively reduced, improving the reliability of early warning results and their adaptability to different operating conditions.

[0012] Third, the system is highly intelligent. It eliminates the heavy reliance on manual inspection experience and achieves fully automated intelligent judgment from data collection and feature analysis to decision-making suggestions. At the same time, through incremental learning and dynamic maintenance of the rule base, it ensures that the system can be continuously iterated and optimized as the working conditions change.

[0013] Fourth, the early warning coverage is wider. By integrating multi-dimensional visual data, it can identify operating conditions such as changes in floc morphology and abnormal foaming that are difficult to judge using sensor data alone, thus expanding the early warning coverage and identification capabilities for process anomalies.

[0014] Fifth, it helps to save energy and reduce consumption during operation. Timely and accurate early warnings and targeted operational guidance can avoid problems such as over-aeration and decreased treatment efficiency caused by process deterioration, ensuring the stable and efficient operation of aeration tanks and achieving energy saving and consumption reduction in wastewater treatment plants. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the workflow of the intelligent early warning method of the present invention; Figure 2 This is a schematic diagram of the structure of the deep learning anomaly pattern recognition unit of the present invention; Figure 3 This is a schematic diagram illustrating the construction and maintenance mechanism of the expert rule base of this invention.

[0016] The diagram shows the data acquisition and preprocessing module 1, the visual feature extraction module 2, the deep learning anomaly pattern recognition module 3, the intelligent verification, grading judgment and early warning release module 4, and the data interface module 5. Detailed Implementation

[0017] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to specific embodiments.

[0018] like Figure 1As shown, the intelligent early warning system of this invention is arranged sequentially according to the data acquisition, feature extraction, anomaly identification, and verification and early warning chain. It includes a data acquisition and preprocessing module 1, a visual feature extraction module 2, a deep learning anomaly pattern recognition module 3, an intelligent verification and grading judgment and early warning release module 4, and a data interface module 5. The data acquisition and preprocessing module 1 serves as the system's data input, connecting to the on-site image acquisition equipment and parameter sensing equipment. Its output is connected to the input of the visual feature extraction module 2 and the deep learning anomaly pattern recognition module 3, respectively. The output of the visual feature extraction module 2 is connected to the visual feature input of the deep learning anomaly pattern recognition module 3. The output of the deep learning anomaly pattern recognition module 3 is connected to the input of the intelligent verification and grading judgment and early warning release module 4. The output of the intelligent verification and grading judgment and early warning release module 4 is connected to the data interface module 5, ultimately outputting early warning information and operation commands through the data interface module 5.

[0019] In actual engineering deployment, industrial-grade underwater high-definition cameras and surface high-definition cameras are installed in the aeration tank. The underwater camera shell is made of corrosion-resistant material and equipped with a self-cleaning function. It is submerged below the surface of the aeration tank, facing the flow area of ​​activated sludge flocs, and is used to collect microscopic images of the activated sludge flocs. The surface camera is mounted on the inspection bridge above the surface of the aeration tank, taking a downward view of the aeration tank surface, and is used to collect macroscopic images of the foam morphology and bubble distribution. The image resolution of both types of cameras is no less than 1080P, and the frame rate is set to 1 to 5 frames per second, which can be flexibly adjusted according to the on-site computing power and monitoring needs. An online sensor group is also deployed in the aeration tank, equipped with an SC200 controller, a dissolved oxygen probe, an oxidation-reduction potential probe, a pH probe, and an optional online ammonia nitrogen and nitrate analyzer. The sensor data acquisition frequency is once every 1 to 5 minutes. All sensors are calibrated regularly according to the operation and maintenance procedures to ensure data accuracy. All acquired visual and sensor data are transmitted via industrial Ethernet and aligned with a unified timestamp using the Network Time Protocol to ensure that the synchronization error between visual and sensor data is less than 50 milliseconds, providing a timing alignment basis for subsequent multimodal fusion.

[0020] The collected raw data is first sent to the data acquisition and preprocessing module 1 for preprocessing. For visual data, Gaussian filtering for noise reduction and histogram equalization for image enhancement are performed sequentially to improve image clarity and contrast. Then, a semantic segmentation network trained on the U-Net architecture is used to accurately segment the activated sludge floc region and foam region in the image, removing the water background and irrelevant interference regions. For sensor time series data, moving average filtering is used to smooth data noise, and the Laida criterion is used to remove outliers. For occasional missing sensor data, linear interpolation is used to complete the data, ensuring the continuity and integrity of the time series data. After preprocessing, the data acquisition and preprocessing module 1 transmits the segmented visual data to the visual feature extraction module 2, and directly transmits the preprocessed biochemical parameter time series data to the deep learning anomaly pattern recognition module 3.

[0021] After receiving the visual data, the visual feature extraction module 2 comprehensively utilizes traditional image processing algorithms and deep learning feature encoding techniques to extract multi-dimensional quantitative features. For the segmented activated sludge floc region, traditional image processing methods such as contour detection and connected component analysis are used to extract basic morphological features such as floc quantity, average size, perimeter, area ratio, and morphological moments. Simultaneously, a ResNet-50 model pre-trained on a large dataset of activated sludge images is used as a feature encoder to extract deep semantic features of the flocs, quantifying their compactness and structural complexity. For the water surface foam region, morphological opening and closing operations and clustering algorithms are used to analyze the number, average size, color space parameters, morphological stability, and water surface coverage of the foam. For the bubble region, background subtraction is used to extract bubble contours, analyze bubble size, density, and uniformity, and optical flow is used to track and calculate the bubble's rising speed. All extracted visual features are standardized to form a structured visual feature vector, which is then output to the deep learning anomaly pattern recognition module 3.

[0022] like Figure 2As shown, the deep learning anomaly pattern recognition module 3 employs a dual-stream deep neural network architecture, divided into two parallel processing paths: a visual branch and a temporal branch. The visual branch uses a ResNet-50 pre-trained network on the ImageNet dataset as its backbone, receiving visual feature vector input and outputting high-order visual feature vectors through a global average pooling layer. The temporal branch uses a two-layer bidirectional long short-term memory network, combined with an attention mechanism, to process time-series data of biochemical parameters, automatically capturing the long-term temporal dependencies and trend characteristics of key parameters. The features output from both branches are fed into a multimodal fusion layer. This fusion layer uses a transformer encoder structure, employing self-attention and cross-attention mechanisms to capture feature correlations within a single modality and coupling correlations between different modalities, achieving deep fusion of visual and temporal features. The fused features are then input into a fully connected classification output layer, which outputs the initially identified process anomaly categories and their corresponding confidence scores. Identifiable process anomaly categories include early signs of activated sludge bulking, signs of activated sludge disintegration and aging, insufficient or excessive aeration, signs of organic load shock, signs of impaired nitrification and denitrification functions, and signs of sludge poisoning.

[0023] The deep neural network model is trained using supervised learning. The training dataset is collected from over three years of historical operating data from aeration tanks in a large wastewater treatment plant, covering complete data samples of various common abnormal operating conditions. All samples are precisely labeled by senior process experts. The dataset is divided into training, validation, and test sets in a 7:1.5:1.5 ratio. The training process uses the Adam optimizer with an initial learning rate of 0.001 and a batch size of 32. An early stopping strategy is employed: training stops when the validation set loss no longer decreases after ten consecutive training epochs to prevent overfitting. Simultaneously, data augmentation techniques such as random pruning, rotation, and brightness adjustment are used to expand the visual dataset and improve the model's generalization ability. After the system goes live, new labeled abnormal data is collected quarterly for incremental learning or periodic retraining. Transfer learning reuses the weight parameters of the old model to quickly adapt to changes in operating conditions caused by fluctuations in influent water quality and adjustments in process conditions, ensuring the long-term stability of the model's recognition performance.

[0024] like Figure 3As shown, the intelligent verification, grading, judgment, and early warning release module 4 incorporates an expert rule base, serving as the basis for verifying and correcting anomaly identification results. The expert rule base is built upon the operational experience of senior process experts, historical operational data mining results, national and industry emission standards, and equipment operation manuals, stored in structured JSON format. Each rule includes a unique identifier, priority, judgment conditions, execution action, version information, and associated expert information. Judgment conditions can integrate real-time parameter thresholds, parameter change trends, preliminary model judgment results, and specific visual features. Execution actions correspond to early warning levels, anomaly type descriptions, and specific operational suggestions. The rule base uses a version control system to manage all modification records, ensuring traceability of rule changes. Before a rule takes effect, the system automatically performs conflict detection, identifies contradictory rules, and prompts experts for correction, ensuring the consistency and effectiveness of the rule set. The platform provides historical data backtracking analysis functionality, allowing simulation testing of new rules or optimization of threshold parameters on historical anomaly event datasets. Rule effectiveness is verified by calculating hit rate, false alarm rate, and false negative rate. Newly added or modified rules must undergo simulation testing and expert review before being put into use, supporting continuous dynamic maintenance and optimization of the rule base.

[0025] The intelligent verification, grading, judgment, and early warning module 4 receives the preliminary anomaly judgment results output by the deep learning model and the rule triggering status from the expert rule base in parallel. It then makes a final judgment according to a preset decision logic: when the model outputs a high-confidence judgment and the expert rule result supports it, the system directly confirms the anomaly type and issues an early warning of the corresponding level; when the model judgment result conflicts with the expert rule, the system prioritizes the higher-priority expert rule or the hard threshold rule based on environmental emission standards, while recording the model's misjudgment data for subsequent model optimization; when the model outputs a low-confidence judgment but the expert rule is triggered, the system adopts the expert rule judgment and issues an early warning, while recording the corresponding data for model iteration training; when the model fails to identify an anomaly but the expert rule is triggered, the system directly executes the rule judgment result to compensate for the model's knowledge gaps and insufficient adaptation to new anomalies. The final early warning results are divided into three levels: Attention, Warning, and Emergency, corresponding to yellow, orange, and red indicators, respectively. Targeted operational suggestions are also generated, including adjusting aeration rate, checking influent water quality, adjusting sludge discharge rate and return ratio, and activating backup aeration equipment, to assist operators in quickly and accurately implementing process interventions.

[0026] The entire system can be flexibly deployed, either on an edge computing gateway for local real-time data analysis, feature extraction, and low-latency early warning, or customized based on the NVIDIA-Jetson-AGX-Xavier development kit. The edge gateway uploads pre-processed data and initial early warning information to the cloud platform via message queue telemetry transmission protocol or expressive state transition interface. The cloud platform is responsible for full model training, unified management and dynamic updates of the expert rule base, multi-plant data aggregation analysis, and remote monitoring visualization. The system provides standardized industrial communication interfaces, supporting seamless integration with existing SCADA or DCS systems. Early warning information and operational suggestions can be directly pushed to operator workstations, and real-time alarm information can be pushed to authorized users through a web interface and mobile application, enabling multi-terminal human-machine collaborative operation and maintenance.

[0027] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An intelligent early warning method based on the fusion of multi-dimensional visual features and key biochemical parameters of activated sludge in aeration tanks, characterized in that, Includes the following steps: a. Data acquisition and preprocessing: Visual data of the aeration tank is acquired in real time through underwater and surface cameras, and time series data of key biochemical parameters of the aeration tank are acquired in real time through online sensors. The visual data and the time series data are preprocessed respectively. b. Visual feature extraction: Using image processing and computer vision techniques, multi-dimensional visual features are extracted and quantified from preprocessed visual data to form a visual feature vector. c. Deep learning anomaly pattern recognition: The visual feature vector and the time series data of the key biochemical parameters are used as inputs and fed into a deep neural network model for multimodal feature learning and fusion. The model outputs a preliminary process anomaly category and its confidence level. The deep neural network model is trained using supervised learning and supports incremental learning or periodic retraining. d. Intelligent verification, graded judgment and early warning issuance: Receive the preliminary anomaly category and confidence level, combine the absolute value, trend and rate of change of key biochemical parameters in real time, and the preset expert rule base, to intelligently verify and correct the preliminary judgment, confirm the anomaly type and determine the severity level, and issue graded early warning information and specific operation suggestions. The expert rule base is built based on the experience of process experts, historical operating data and industry emission standards, and supports dynamic maintenance and optimization.

2. The intelligent early warning method based on the fusion of multi-dimensional visual features and key biochemical parameters of activated sludge in an aeration tank according to claim 1, characterized in that, In step a, the visual data acquisition includes acquiring images of activated sludge flocs through an underwater camera, and acquiring images of water surface foam and bubble distribution through a surface camera.

3. The intelligent early warning method based on the fusion of multi-dimensional visual features and key biochemical parameters of activated sludge in an aeration tank according to claim 1, characterized in that, In step b, the multi-dimensional visual features include the size, shape, density, compactness, contour complexity, color, movement trajectory, and settling velocity of activated sludge flocs; the quantity, size, color, shape, stability, and coverage of surface foam; and one or more of the size, density, uniformity, and rising speed of bubbles.

4. The intelligent early warning method based on the fusion of multi-dimensional visual features and key biochemical parameters of activated sludge in an aeration tank according to claim 1, characterized in that, In step a, the key biochemical parameters include core key biochemical parameters and optional key auxiliary biochemical parameters. The core key biochemical parameters are dissolved oxygen, redox potential, and pH value, and the key auxiliary biochemical parameters are one or more of ammonia nitrogen and nitrate nitrogen.

5. The intelligent early warning method based on the fusion of multi-dimensional visual features and key biochemical parameters of activated sludge in an aeration tank according to claim 1, characterized in that, In step c, the deep neural network model adopts a two-stream network structure, in which one branch uses a convolutional neural network or a visual transformer to process visual feature data, and the other branch uses a long short-term memory network, a gated recurrent unit, or a time series transformer to process time series data.

6. The intelligent early warning method based on the fusion of multi-dimensional visual features and key biochemical parameters of activated sludge in an aeration tank according to claim 5, is characterized in that, In step c, the multimodal feature fusion adopts feature splicing, attention-based fusion, or gating unit fusion methods.

7. The intelligent early warning method based on the fusion of multi-dimensional visual features and key biochemical parameters of activated sludge in an aeration tank according to claim 1, characterized in that, In step c, the process anomaly category includes one or more of the following: early signs of activated sludge bulking, signs of activated sludge disintegration and aging, insufficient or excessive aeration, signs of organic load shock, signs of impaired nitrification and denitrification functions, and signs of sludge poisoning.

8. The intelligent early warning method based on the fusion of multi-dimensional visual features and key biochemical parameters of activated sludge in an aeration tank according to claim 1, characterized in that, In step d, when the initial judgment of the deep neural network model conflicts with the expert rules or the model confidence is insufficient, the final decision is made by prioritizing or combining the judgment of the expert rules.

9. The intelligent early warning method based on the fusion of multi-dimensional visual features and key biochemical parameters of activated sludge in an aeration tank according to claim 1, characterized in that, In step d, the graded early warning information includes three levels: attention, warning, and emergency, which correspond to yellow, orange, and red warning icons, respectively.

10. An intelligent early warning system based on the fusion of multi-dimensional visual characteristics and key biochemical parameters of activated sludge in an aeration tank, characterized in that, It includes a data acquisition and preprocessing module (1), a visual feature extraction module (2), a deep learning abnormal pattern recognition module (3), an intelligent verification, hierarchical judgment and early warning release module (4), and a data interface module (5). The data acquisition and preprocessing module (1) is connected to the visual feature extraction module (2) and the deep learning abnormal pattern recognition module (3) respectively, and is used to transmit the preprocessed visual data to the visual feature extraction module (2) and transmit the preprocessed key biochemical parameter time series data to the deep learning abnormal pattern recognition module (3). The visual feature extraction module (2) is signal-connected to the deep learning abnormal pattern recognition module (3) and is used to transmit the extracted visual feature vector to the deep learning abnormal pattern recognition module (3). The deep learning anomaly pattern recognition module (3) is connected to the intelligent verification, grading judgment and early warning release module (4) by signal, and is used to transmit the output preliminary process anomaly category and confidence level to the intelligent verification, grading judgment and early warning release module (4) for verification correction and grading judgment. The intelligent verification, hierarchical judgment and early warning release module (4) is connected to the data interface module (5) by signal, and is used to transmit hierarchical early warning information and operation suggestions to the data interface module (5). The data interface module (5) is used to realize communication between the system and the field control system or cloud computing platform. The system can be deployed on edge computing devices or cloud computing platforms and provide users with early warning information and operation suggestions through Web interface and mobile application.

Images