Industrial equipment state monitoring and hierarchical early warning system and method based on edge computing

Abstract

The application discloses an industrial equipment state monitoring and hierarchical early warning system and method based on edge computing, comprising an edge side multi-source data synchronous acquisition and preprocessing module, an edge side lightweight feature extraction and state representation module, an edge-cloud collaborative state evaluation and fault preliminary judgment module, a hierarchical early warning decision module based on risk entropy, a fault tracing and diagnosis reasoning module and an early warning information distribution and closed-loop disposal module; the application sinks the computing tasks such as data preprocessing, feature extraction and state preliminary judgment to the edge side, uploads only abnormal data to the cloud, filters a large amount of normal data, greatly reduces the data transmission amount, fully utilizes the computing power resources of the edge side, relieves the computing pressure of the cloud server, supports the simultaneous access and monitoring of large-scale industrial equipment and improves the scalability of the system.

Description

Technical Field

[0001] This invention relates to the fields of industrial equipment condition monitoring, edge computing, and fault early warning technology, specifically to an industrial equipment condition monitoring and hierarchical early warning system and method based on edge computing. Background Technology

[0002] Traditional industrial equipment condition monitoring systems mostly adopt a centralized sensor-cloud architecture. All data collected by field sensors is uploaded to a cloud server via an industrial network, where data storage, processing, analysis, and early warning decisions are handled by the cloud. This architecture played a crucial role in the early stages of industrial informatization, but with the surge in the number of industrial devices and the continuous expansion of monitoring dimensions, its inherent shortcomings have become increasingly apparent, mainly in the following aspects: Industrial field sensors generate massive amounts of time-series data every second. Uploading all of this data to the cloud would consume a significant amount of industrial network bandwidth, leading to network congestion and data transmission delays. Simultaneously, cloud servers would need to process data from tens of thousands of devices, severely consuming computing resources and making it difficult to meet the demands of large-scale concurrent monitoring, thus greatly limiting the system's scalability.

[0003] Data collection from edge-side sensors to cloud processing and the return of alert results involves multiple stages, including data transmission, cloud computing, and result feedback, resulting in significant network and processing latency. For critical equipment such as high-speed rotating motors and steam turbines, faults evolve extremely rapidly; even millisecond-level delays can cause a minor anomaly to escalate into a serious accident. The latency characteristics of centralized architectures cannot meet the real-time alert requirements of such equipment.

[0004] Most existing systems use a single threshold method for early warning, triggering an alert when a monitored parameter exceeds a preset threshold. This method fails to comprehensively consider the probability of a fault occurring, its severity, and the potential scope of its impact, resulting in overly coarse alert levels. Minor anomalies may trigger high-level alerts, causing unnecessary production downtime; while for potentially serious faults, since a single parameter may not exceed the threshold, an alert may not be triggered in time, leading to missed incidents.

[0005] The operating status of modern industrial equipment needs to be comprehensively characterized by parameters from multiple dimensions such as vibration, temperature, current, and pressure. However, due to differences in hardware characteristics and deployment locations, different types of sensors collect data with inconsistencies in timestamps, sampling frequencies, and data formats. Existing systems lack effective mechanisms for synchronizing and fusing multi-source data, making it difficult to organically integrate data from different dimensions. This results in an inability to comprehensively and accurately characterize the operating status of equipment, significantly reducing the accuracy of condition assessment and fault diagnosis.

[0006] Most existing systems can only detect equipment anomalies, but cannot quickly pinpoint the root cause and specific reason for the fault. When equipment malfunctions, maintenance personnel need to rely on their personal experience to conduct a comprehensive investigation, which not only consumes a lot of time and manpower, but may also lead to repeated failures due to incomplete investigation, prolonging equipment downtime and reducing production efficiency.

[0007] To address these issues, the industry has begun exploring the application of edge computing technology in industrial equipment condition monitoring. By offloading some computing tasks to the edge, data can be processed and analyzed locally. However, existing edge computing monitoring systems still have several shortcomings: First, the collaboration mechanism between the edge and the cloud is imperfect, failing to fully leverage the advantages of both; second, there is a lack of scientific early warning classification methods, and the accuracy and rationality of early warning decisions need to be improved; third, fault diagnosis and tracing capabilities are insufficient, failing to provide effective support for operation and maintenance work. Summary of the Invention

[0008] The purpose of this invention is to provide an industrial equipment condition monitoring and hierarchical early warning system and method based on edge computing, so as to solve the problems existing in the prior art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an industrial equipment status monitoring and hierarchical early warning system based on edge computing, including a multi-source data synchronous acquisition and preprocessing module on the edge side, a lightweight feature extraction and status characterization module on the edge side, an edge-cloud collaborative status assessment and initial fault judgment module, a hierarchical early warning decision-making module based on risk entropy, a fault tracing and diagnostic reasoning module, and an early warning information distribution and closed-loop handling module. The edge-side multi-source data synchronous acquisition and preprocessing module interfaces with various sensing devices in the industrial field to realize synchronous acquisition and standardized preprocessing of multimodal data, and output time-series datasets in a unified format. The edge-side lightweight feature extraction and state representation module receives the time-series dataset, extracts key features of the device state through a lightweight deep learning model, and generates a multi-dimensional state representation vector. The edge-cloud collaborative status assessment and initial fault judgment module completes the initial judgment of the normal / abnormal status of the device on the edge side, uploads the abnormal data to the cloud for in-depth fault analysis, and the cloud feeds back the model update parameters to the edge side. The risk entropy-based hierarchical early warning decision module receives the status assessment results, calculates the equipment operation risk entropy value, and outputs an early warning signal of the corresponding level according to the risk entropy value range. The fault tracing and diagnosis reasoning module combines historical equipment operation data with fault knowledge graphs to locate the root cause of the fault and generate a diagnostic report. The early warning information distribution and closed-loop handling module distributes early warning information to the corresponding responsible persons according to the early warning level, tracks the handling progress, and updates the system status.

[0010] Furthermore, the edge-side multi-source data synchronous acquisition and preprocessing module includes a data acquisition unit, a time synchronization unit, a data cleaning unit, and a data standardization unit; The data acquisition unit interfaces with vibration, temperature, current, voltage, and pressure sensors via an industrial bus protocol to collect multimodal raw data during device operation. The time synchronization unit uses the system clock of the edge gateway as a reference to perform timestamp calibration and alignment on the data collected by different sensors. The data cleaning unit identifies and repairs missing values, outliers, and noise in the raw data. The data standardization unit converts the cleaned data into a unified numerical range and data format, generating a time-series dataset. The calculation formula for timestamp calibration and alignment is as follows: in, The standard timestamp after data calibration for the i-th sensor device; The local timestamp when the i-th sensor device collected this data; The one-way transmission delay of data from the i-th sensor device to the edge gateway; This is the fixed offset between the local clock of the i-th sensor device and the reference clock of the edge gateway.

[0011] Furthermore, the edge-side lightweight feature extraction and state representation module includes a data framing unit, a lightweight feature extraction unit, and a feature fusion unit; The data framing unit divides the time-series dataset into segments according to fixed time windows to generate a continuous sequence of data frames; the lightweight feature extraction unit uses a deep separable convolutional neural network to extract local temporal and frequency domain features of each data frame; the feature fusion unit performs adaptive weighted fusion of the extracted multi-dimensional features to generate a multi-dimensional feature vector characterizing the device's operating status.

[0012] The formula for adaptive weighted fusion is: in, This is the final feature vector after fusion; Let i be the feature vector of the i-th feature dimension; For the i-th feature dimension, the adaptive weights are used. The total number of feature dimensions; Adaptive weights The calculation method is as follows: in, is the information gain value of the i-th feature dimension, used to measure the ability of this feature to distinguish between normal and abnormal states of the device.

[0013] Furthermore, the edge-cloud collaborative status assessment and initial fault judgment module includes an edge-side initial status judgment unit, a cloud-side deep analysis unit, and a model iteration and update unit; The edge-side initial state judgment unit loads a pre-trained binary classification model, processes the multi-dimensional feature vectors, and outputs a preliminary judgment result of whether the device is normal or abnormal; the cloud-based deep analysis unit receives abnormal feature data and raw data fragments uploaded from the edge side, and uses a complex deep learning model to perform in-depth analysis of fault type and fault degree; the model iteration and update unit uses the fault data accumulated in the cloud to incrementally train the edge-side model and sends the updated model parameters to the edge gateway.

[0014] Furthermore, the risk entropy-based hierarchical early warning decision-making module includes a risk factor calculation unit, a risk entropy calculation unit, and an early warning level classification unit; The risk factor calculation unit extracts three core risk factors—fault occurrence probability, fault impact range, and fault severity—based on the status assessment results; the risk entropy calculation unit calculates the comprehensive risk entropy value of the equipment operation based on the three core risk factors; and the early warning level classification unit classifies the early warning into multiple levels and outputs corresponding early warning signals based on the preset risk entropy value range.

[0015] The formula for calculating the comprehensive risk entropy is: in, The comprehensive risk entropy value for equipment operation; Let be the probability of the k-th type of fault occurring; is the severity coefficient of the k-th type of fault; The influence range coefficient for the k-th type of fault; This represents the total number of possible fault types.

[0016] Furthermore, the fault tracing and diagnosis reasoning module includes a knowledge graph construction unit, a fault reasoning unit, and a diagnosis report generation unit; The knowledge graph construction unit integrates equipment design data, historical fault records, maintenance manuals, and expert experience to construct a fault knowledge graph that includes equipment entities, fault phenomena, fault causes, and maintenance measures. The fault reasoning unit uses a combination of rule-based reasoning and Bayesian network reasoning, combining current equipment state characteristics with the knowledge graph, to locate the root cause of the fault. The diagnostic report generation unit generates a diagnostic report that includes fault phenomena, fault causes, fault locations, and recommended maintenance measures based on the fault reasoning results.

[0017] Furthermore, the early warning information distribution and closed-loop handling module includes an early warning classification and distribution unit, a handling task management unit, and a handling effect evaluation unit; The early warning classification and distribution unit distributes early warning information and diagnostic reports to the responsible persons at the corresponding levels through multiple communication channels according to the early warning level; the handling task management unit creates and tracks fault handling tasks, and records key nodes and operation information in the handling process; the handling effect evaluation unit evaluates the handling effect after the fault handling is completed and feeds back the evaluation results to the system for optimizing the early warning model and inference rules.

[0018] Furthermore, the time synchronization unit adopts a one-way broadcast time synchronization mechanism. The edge gateway periodically sends a reference time synchronization packet to all sensing devices. The sensing devices update their local clock offset according to the received synchronization packet and perform real-time timestamp calibration on the subsequently collected data.

[0019] Furthermore, the lightweight feature extraction unit employs a deep separable convolutional neural network composed of multiple deep separable convolutional blocks stacked together. Each deep separable convolutional block includes a deep convolutional layer, a pointwise convolutional layer, a batch normalization layer, and an activation function layer, thereby achieving real-time feature extraction at the edge by reducing the number of model parameters and computational cost.

[0020] A method for monitoring and classifying early warning of industrial equipment status based on edge computing includes the following steps: Step 1: System Initialization and Data Acquisition The edge-side multi-source data synchronous acquisition and preprocessing module is activated to connect with various sensing devices in the industrial field, collect multimodal raw data of device operation, and generate time-series datasets after time synchronization, data cleaning and standardization processing. Step 2: Edge-side feature extraction and state representation The time-series dataset is input into the edge-side lightweight feature extraction and state representation module. Through data framing, lightweight feature extraction and adaptive feature fusion, a multi-dimensional feature vector representing the device's operating state is generated. Step 3: Edge-Cloud Collaboration Status Assessment The edge-side state initial judgment unit processes the multi-dimensional feature vector to determine the device's operating status. If the status is determined to be normal, the feature data is stored only on the edge side. If the status is determined to be abnormal, the abnormal feature data and the original data fragments are uploaded to the cloud for in-depth fault analysis. Step 4: Comprehensive Risk Assessment and Tiered Early Warning The risk entropy-based hierarchical early warning decision module receives the status assessment results, extracts the core risk factors and calculates the comprehensive risk entropy value, and outputs the corresponding level of early warning signal according to the risk entropy value range. Step 5: Fault Origin and Diagnosis The fault tracing and diagnosis reasoning module combines the fault knowledge graph with the current equipment status characteristics to locate the root cause of the fault and generate a diagnostic report. Step 6: Early Warning Distribution and Closed-Loop Response The early warning information distribution and closed-loop handling module distributes early warning information and diagnostic reports according to the early warning level, creates and tracks fault handling tasks, evaluates the handling effect after the handling is completed, and updates the system status.

[0021] Compared with the prior art, the beneficial effects of the present invention are: This invention offloads computational tasks such as data preprocessing, feature extraction, and initial state assessment to the edge, uploading only abnormal data to the cloud. This filters out a large amount of normal data, significantly reducing data transmission volume, making full use of edge computing resources, alleviating the computational pressure on cloud servers, supporting simultaneous access and monitoring of large-scale industrial equipment, and improving system scalability.

[0022] The edge device performs a preliminary assessment of its status, enabling millisecond-level response to abnormal conditions. This avoids the delay in early warnings caused by cloud transmission latency and allows for timely detection of early device failures. Simultaneously, the precise time synchronization and adaptive feature fusion of multi-source data comprehensively and accurately characterize the device's operating status, improving the accuracy of status assessment and fault identification, and effectively reducing false alarms and missed alarms.

[0023] The risk entropy-based hierarchical early warning mechanism comprehensively considers three core risk factors: the probability of failure, the severity, and the scope of impact. It can comprehensively and objectively assess the operational risks of equipment, and the early warning level classification is scientific and reasonable. It provides accurate decision support for operation and maintenance personnel, making it easier for them to take corresponding measures according to the early warning level and rationally allocate operation and maintenance resources.

[0024] A fault knowledge graph containing knowledge of the entire equipment lifecycle has been constructed. A hybrid reasoning approach combining rule-based reasoning and Bayesian network reasoning is adopted, which can quickly and accurately locate the root cause of the fault and generate detailed diagnostic reports. This provides clear maintenance guidance for operation and maintenance personnel, significantly shortens the fault diagnosis and downtime, and improves operation and maintenance efficiency. Attached Figure Description

[0025] Figure 1 This is a system module diagram of the present invention; Figure 2 This is a schematic diagram of the edge-side multi-source data synchronous acquisition and preprocessing module of the present invention; Figure 3 This is a schematic diagram of the edge-side lightweight feature extraction and state characterization module of the present invention; Figure 4 This is a flowchart of the method of the present invention. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0027] Please see Figure 1-4 This invention provides an industrial equipment status monitoring and hierarchical early warning system based on edge computing, including a multi-source data synchronous acquisition and preprocessing module on the edge side, a lightweight feature extraction and status characterization module on the edge side, an edge-cloud collaborative status assessment and initial fault judgment module, a hierarchical early warning decision module based on risk entropy, a fault tracing and diagnosis reasoning module, and an early warning information distribution and closed-loop handling module. The edge-side multi-source data synchronous acquisition and preprocessing module is responsible for acquiring various status data during the operation of industrial equipment and performing preliminary processing on the raw data to eliminate noise and inconsistencies, providing a high-quality data foundation for subsequent feature extraction and status assessment. This module establishes communication connections with various field sensors via industrial bus protocols, supports multiple mainstream industrial bus protocols, and is compatible with different manufacturers and types of sensors, exhibiting good versatility and scalability.

[0028] During the data acquisition phase, the data acquisition unit simultaneously collects multimodal data, including vibration, temperature, current, voltage, and pressure. These data reflect the operating status of the equipment from different physical dimensions. For example, for rotating equipment, vibration data contains rich fault characteristics, reflecting mechanical faults such as bearing wear, rotor imbalance, and shaft misalignment; temperature data reflects the thermal state of the equipment, enabling timely detection of overheating faults caused by friction, overload, etc.; current data reflects the electrical state of the equipment, identifying electrical faults such as stator winding short circuits and rotor bar breakage. Through the comprehensive acquisition of multimodal data, the operating status of the equipment can be characterized in a comprehensive and three-dimensional way, avoiding missed faults caused by monitoring a single parameter.

[0029] Time synchronization is a prerequisite for multi-source data fusion. Due to the different hardware characteristics of various sensing devices, their local clocks inevitably deviate, and network latency during data transmission causes data collected by different sensing devices to be out of sync in time. For example, abnormal equipment vibration and temperature rise occurring at the same time may be mistakenly identified as two independent events due to asynchronous timestamps, affecting the accuracy of fault diagnosis. The time synchronization unit adopts a one-way broadcast time synchronization mechanism. The edge gateway periodically sends a reference time synchronization packet to all sensing devices. After receiving the synchronization packet, the sensing device calculates the offset between its local clock and the edge gateway's reference clock, and uses this information, combined with data transmission latency, to perform real-time calibration of subsequently collected data. In this way, the timestamps of all sensing devices can be unified to the edge gateway's reference clock, achieving precise time synchronization of multi-source data.

[0030] The calculation formula for multi-source data time synchronization calibration is as follows: in, The standard timestamp after data calibration for the i-th sensor device; The local timestamp when the i-th sensor device collected this data; The one-way transmission delay of data from the i-th sensor device to the edge gateway; This is a fixed offset between the local clock of the i-th sensor and the reference clock of the edge gateway. One-way transmission delay. The fixed offset is obtained by sending synchronization packets multiple times and calculating the average round-trip time. It is dynamically updated during each time synchronization process to ensure the long-term accuracy of time calibration.

[0031] Data cleaning is a crucial step in improving data quality. The complex environment of industrial sites leads to a significant amount of noise and outliers in the raw data. For example, electromagnetic interference can cause spikes in current data, mechanical shocks can cause abnormally large values ​​in vibration data, and poor contact in sensing devices can result in missing data. The data cleaning unit first uses statistical methods to identify outliers, marking data points outside the normal statistical range as outliers. Then, interpolation methods are used to repair missing and outlier values. Finally, filtering algorithms are used to remove high-frequency noise and retain the effective features in the data. After data cleaning, the signal-to-noise ratio of the data is significantly improved, laying a solid foundation for subsequent feature extraction.

[0032] The purpose of data standardization is to eliminate differences in dimensions and numerical ranges between different features, ensuring that different features have equal importance in model training. The data standardization unit uses the Z-score standardization method to transform the data for each feature dimension into a standard normal distribution with a mean of 0 and a variance of 1. Standardized data not only facilitates model processing but also improves the model's convergence speed and accuracy.

[0033] The lightweight feature extraction and state representation module at the edge fully utilizes the computing resources of the edge gateway to extract features locally when the data is generated, transforming the raw data into high-dimensional feature vectors, significantly reducing the amount of data that needs to be uploaded to the cloud. The module employs a deep separable convolutional neural network, a lightweight deep learning model specifically designed for edge devices. Compared to traditional convolutional neural networks, it reduces the number of parameters and computational cost by more than an order of magnitude, enabling real-time operation in resource-constrained edge gateways.

[0034] The data framing unit divides continuous time-series data into fixed-length data frames, each representing the device's operating status over a period of time. The size of the time window needs to be reasonably selected based on the device's operating characteristics and fault evolution patterns. If the time window is too small, the data frames will not contain enough information to effectively characterize the device's status; if the time window is too large, fault features will be submerged in a large amount of normal data, reducing the sensitivity of fault detection.

[0035] The lightweight feature extraction unit processes each data frame, extracting its time-domain and frequency-domain features. Time-domain features include mean, variance, peak value, kurtosis, and margin, reflecting the statistical and impact characteristics of the data. Frequency-domain features include fundamental amplitude, harmonic amplitude, frequency center, and spectral entropy, reflecting the frequency distribution characteristics of the data. By simultaneously extracting time-domain and frequency-domain features, the fault characteristics of the equipment can be comprehensively captured, improving the accuracy of fault identification.

[0036] The feature fusion unit adaptively weights and fuses the extracted multi-dimensional features. Different features have varying abilities to distinguish device states; for example, kurtosis is highly sensitive to impact failures, while the mean is more sensitive to slowly changing failures. Traditional equal-weight fusion methods do not consider this difference, resulting in low accuracy in state representation. This invention employs an adaptive weighted fusion method based on information gain, determining the weight of each feature according to its information gain value; features with higher information gain values ​​have higher weights. This approach highlights the role of highly discriminative features, improving the accuracy of state representation.

[0037] The formula for adaptive weighted fusion is: in, This is the final feature vector after fusion; Let i be the feature vector of the i-th feature dimension; For the i-th feature dimension, the adaptive weights are used. The total number of feature dimensions; Adaptive weights The calculation method is as follows: in, Let be the information gain value of the i-th feature dimension, used to measure the feature's ability to distinguish between normal and abnormal device states. A larger information gain value indicates that the feature provides more information about the device state, contributing more to state differentiation, and thus a larger adaptive weight. Through adaptive weighted fusion, the role of high-contribution features can be effectively highlighted, the interference of redundant features can be suppressed, and the accuracy of state representation can be improved.

[0038] The edge-cloud collaborative status assessment and initial fault diagnosis module constructs a collaborative architecture of edge initial judgment and cloud in-depth analysis, fully leveraging the advantages of the edge's high real-time performance and the cloud's powerful computing capabilities. The edge side is responsible for quickly and initially determining the device status and promptly detecting abnormal states; the cloud side is responsible for in-depth analysis of abnormal data to accurately identify the fault type and severity. This architecture ensures both real-time early warning and accurate fault diagnosis.

[0039] The binary classification model loaded by the edge-side state initial judgment unit is pre-trained in the cloud and then deployed to the edge gateway. This model has a simple structure, low computational cost, and can achieve millisecond-level inference at the edge. When the model determines that the device is in a normal state, it only stores the feature data locally at the edge, without uploading it to the cloud. This filters over 90% of normal data at the edge, significantly reducing data transmission volume and cloud computing load. When the model determines that the device is in an abnormal state, it immediately uploads the abnormal feature data and corresponding original data fragments to the cloud for further in-depth analysis.

[0040] The cloud-based deep analytics unit runs sophisticated deep learning models, enabling multi-dimensional and in-depth analysis of anomaly data. The cloud server possesses ample computing resources, capable of simultaneously processing anomaly data from multiple edge gateways and running complex models such as convolutional neural networks and recurrent neural networks. These models have powerful feature learning and pattern recognition capabilities, accurately identifying various types of faults and assessing their severity. Simultaneously, the cloud stores all fault data and analysis results in a database, accumulating a large-scale fault dataset to support iterative model updates.

[0041] The model iterative update unit employs incremental learning to update the edge-side model. Traditional model update methods require retraining the model using all historical data, which is not only computationally intensive but also time-consuming. Incremental learning, on the other hand, trains the model using only newly accumulated fault data. While retaining existing model knowledge, it learns new fault characteristics, enabling rapid adaptation to changes in equipment operating characteristics and the emergence of new fault types. The updated model parameters are distributed to the edge gateway via the network, and the edge gateway automatically completes the model update without manual intervention, achieving online iterative optimization of the model.

[0042] The risk entropy-based hierarchical early warning decision-making module solves the problems of false alarms, missed alarms, and unreasonable early warning levels in traditional single-threshold early warning methods. This module comprehensively considers three core risk factors: the probability of failure, severity, and scope of impact. It assesses the operational risk of equipment by calculating a comprehensive risk entropy value and classifies early warnings according to the risk entropy value range.

[0043] The risk factor calculation unit extracts three core risk factors from the condition assessment results: failure probability reflects the likelihood of a failure occurring, failure severity reflects the potential losses once a failure occurs, and failure impact range reflects the areas and equipment that may be affected by the failure. These three factors reflect the operational risks of the equipment from different perspectives; only by organically combining them can the risk level be comprehensively and accurately assessed.

[0044] The risk entropy calculation unit calculates a comprehensive risk entropy value based on three risk factors. The concept of risk entropy originates from information theory and measures the degree of uncertainty of a system. In this invention, a higher risk entropy value indicates higher uncertainty in equipment operation and a more severe level of risk. Compared with the traditional weighted summation method, the risk entropy method can better handle uncertain information and comprehensively reflect the combined effect of multiple risk factors.

[0045] The formula for calculating the comprehensive risk entropy is: in, The comprehensive risk entropy value for equipment operation; Let be the probability of the k-th type of fault occurring; is the severity coefficient of the k-th type of fault; The influence range coefficient for the k-th type of fault; This represents the total number of possible failure types. The higher the overall risk entropy value, the greater the uncertainty in equipment operation, the more severe the risk, and the more urgent the early warning measures required.

[0046] The early warning level classification unit divides early warnings into four levels based on a preset risk entropy value range, with different levels corresponding to different handling measures. This scientific early warning classification enables operations and maintenance personnel to take appropriate measures according to the warning level, rationally allocate operations and maintenance resources, and improve operational efficiency. For example, for a level one early warning, operations and maintenance personnel can conduct checks during routine inspections; for a level four early warning, immediate shutdown and emergency repairs are required.

[0047] The fault tracing and diagnostic reasoning module can quickly locate the root cause of a fault, providing maintenance personnel with accurate repair guidance and significantly shortening the troubleshooting time. This module uses a fault knowledge graph for reasoning, which integrates equipment design data, historical fault records, maintenance manuals, and expert experience, forming a structured fault knowledge base.

[0048] The knowledge graph construction unit first builds the ontology of the fault knowledge graph, defining entity types and relationships between entities. For example, entity types include equipment, components, fault phenomena, fault causes, and maintenance measures; relationships include equipment containing components, components causing fault phenomena, fault phenomena caused by fault causes, and maintenance measures corresponding to fault causes. Then, natural language processing techniques are used to extract entities and relationships from unstructured documents, converting them into triples in the knowledge graph. The knowledge graph supports dynamic updates, continuously incorporating new fault cases and expert experience to improve the knowledge system.

[0049] The fault reasoning unit employs a hybrid reasoning approach, combining the advantages of rule-based reasoning and Bayesian network reasoning. For known and explicit fault patterns, rule-based reasoning can quickly pinpoint the root cause; for uncertain and complex fault patterns, Bayesian network reasoning can determine the most probable cause through probability calculation. The combination of these two reasoning methods improves the accuracy and efficiency of fault tracing.

[0050] The diagnostic report generation unit generates a structured diagnostic report based on the fault reasoning results. The report includes information such as fault symptoms, fault causes, fault locations, and recommended maintenance measures. The diagnostic report is presented to maintenance personnel in a clear and easy-to-understand manner, enabling them to quickly understand the fault situation and take appropriate maintenance measures.

[0051] The early warning information distribution and closed-loop handling module enables accurate distribution of early warning information and closed-loop management of fault handling, ensuring that early warning information is delivered in a timely manner and effectively addressed. This module constructs a complete closed-loop process from early warning issuance to task tracking and effect evaluation, improving the efficiency and quality of fault handling.

[0052] The early warning tiered distribution unit distributes early warning information to the responsible personnel at the corresponding level based on the early warning level. Different early warning levels correspond to different responsible personnel, ensuring that early warning information reaches the most appropriate people and avoiding information overload or omissions. Early warning information is distributed through multiple channels, including SMS, email, mobile application push notifications, and industrial dashboards, ensuring that responsible personnel receive early warning information in a timely manner.

[0053] The task management unit automatically creates fault handling tasks and tracks their execution throughout the process. Each task clearly defines the responsible person, task content, handling requirements, and completion deadline, ensuring the responsible person clearly understands their duties. The task management unit tracks task progress in real time and automatically sends a reminder to the responsible supervisor if the task is not completed on time, ensuring timely fault handling. Simultaneously, it records key milestones and operational information during the handling process, creating a complete fault handling file.

[0054] After a fault is resolved, the fault handling effectiveness evaluation unit assesses the effectiveness of the handling and feeds the evaluation results back to the system. By monitoring subsequent equipment operating data, it determines whether the fault has been completely resolved. If the handling is effective, the evaluation results are fed back to the system for optimizing the model and knowledge graph; if the handling is ineffective, the system will re-perform fault diagnosis and early warning to ensure that the fault is ultimately resolved.

[0055] Example: Taking the condition monitoring and graded early warning of asynchronous motors widely used in industrial production as an example, the specific implementation process of this invention is described in detail: After system initialization, the edge-side multi-source data synchronous acquisition and preprocessing module interfaces with vibration sensors, temperature sensors, current sensors, and voltage sensors installed on the asynchronous motor. The data acquisition unit collects raw data such as motor vibration acceleration, stator temperature, three-phase current, and three-phase voltage at a preset sampling frequency. The time synchronization unit uses the edge gateway's system clock as a reference to perform timestamp calibration and alignment on the data collected by the four sensors, ensuring that the data from different sensors are synchronized in the time dimension. The data cleaning unit identifies and repairs missing values, outliers, and noise in the raw data. For example, for current spikes caused by electromagnetic interference, a sliding window averaging method is used for filtering; for missing temperature data caused by poor sensor contact, linear interpolation is used for filling. The data standardization unit converts the cleaned data into a unified numerical range and data format, generating a time-series dataset containing four dimensions.

[0056] The time-series dataset is input into the lightweight feature extraction and state representation module on the edge side. The data framing unit divides the time-series dataset into a continuous sequence of data frames according to a fixed time window. The lightweight feature extraction unit uses a deep separable convolutional neural network to extract the temporal and frequency domain features of each data frame. The temporal features include mean, variance, peak value, kurtosis, etc., and the frequency domain features include fundamental amplitude, harmonic amplitude, frequency center, etc. The feature fusion unit calculates the information gain value of each feature dimension. For example, the kurtosis feature is very sensitive to bearing faults, so its information gain value is large, and its corresponding weight is also large; while the mean feature has a weak ability to distinguish bearing faults, so its information gain value is small, and its corresponding weight is also small. Adaptive weights are calculated based on the information gain values, and the multi-dimensional features are weighted and fused to generate a multi-dimensional feature vector representing the motor's operating state.

[0057] The edge-side initial state judgment unit loads a pre-trained binary classification model, processes the multi-dimensional feature vectors, and outputs a preliminary judgment result indicating whether the motor is normal or abnormal. If the motor is judged to be in a normal state, the feature data is stored in the local storage of the edge gateway; if it is judged to be in an abnormal state, the abnormal feature data and the corresponding original data fragments are immediately uploaded to the cloud. After receiving the abnormal data, the cloud-based deep analysis unit uses a convolutional neural network model for deep analysis, identifies a bearing wear fault in the motor, and assesses the fault severity as moderate. The model iteration and update unit adds the fault data to the training set, incrementally trains the edge-side binary classification model, and sends the updated model parameters to the edge gateway, which automatically completes the model update.

[0058] The risk entropy-based hierarchical early warning decision-making module receives the status assessment results. The risk factor calculation unit extracts the probability of occurrence, severity coefficient, and impact range coefficient of bearing wear failure. The risk entropy calculation unit calculates the comprehensive risk entropy value of motor operation based on the three risk factors. The early warning level classification unit determines the current early warning as a Level 3 early warning (orange) based on the preset risk entropy value range.

[0059] The fault tracing and diagnostic reasoning module receives three levels of early warning signals. The fault reasoning unit matches bearing wear fault phenomena in the fault knowledge graph and filters out possible fault causes through rule-based reasoning, including poor lubrication, bearing overload, improper installation, and bearing aging. Then, Bayesian network reasoning is used to calculate the probability of occurrence for each fault cause, with poor lubrication having the highest probability, thus identifying poor lubrication as the root cause. The diagnostic report generation unit generates a diagnostic report, which includes the fault phenomenon, fault cause, fault location, and recommended maintenance measures.

[0060] The early warning grading and distribution unit distributes level-three early warning information and diagnostic reports to on-site maintenance personnel and team leaders via SMS and mobile application. The task management unit automatically creates fault handling tasks, assigning on-site maintenance personnel as the responsible parties, with a 24-hour completion deadline. Upon receiving the task, the maintenance personnel go to the site for inspection, confirming poor lubrication of the motor drive bearing. They replenish grease according to recommended maintenance measures and check the bearing seals, finding slight aging and replacing them. After handling, the maintenance personnel submit a handling completion report via mobile application. The handling effectiveness evaluation unit monitors the motor's operating data for the following 24 hours, finding that the motor's vibration and temperature values ​​have returned to normal ranges, determining the handling effective. The handling effectiveness evaluation results are fed back to the system, the model iteration and update unit optimizes the model parameters based on this data, and the knowledge graph construction unit improves the relevant inference rules for bearing lubrication problems. The system returns to the data acquisition state and continues to monitor the motor's operating status in real time.

[0061] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An industrial equipment condition monitoring and hierarchical early warning system based on edge computing, characterized in that: It includes a multi-source data synchronous acquisition and preprocessing module for the edge side, a lightweight feature extraction and state characterization module for the edge side, a collaborative state assessment and initial fault judgment module for the edge-cloud, a hierarchical early warning decision-making module based on risk entropy, a fault tracing and diagnostic reasoning module, and an early warning information distribution and closed-loop handling module. The edge-side multi-source data synchronous acquisition and preprocessing module interfaces with various sensing devices in the industrial field to realize synchronous acquisition and standardized preprocessing of multimodal data, and output time-series datasets in a unified format. The edge-side lightweight feature extraction and state representation module receives the time-series dataset, extracts key features of the device state through a lightweight deep learning model, and generates a multi-dimensional state representation vector. The edge-cloud collaborative status assessment and initial fault judgment module completes the initial judgment of the normal / abnormal status of the device on the edge side, uploads the abnormal data to the cloud for in-depth fault analysis, and the cloud feeds back the model update parameters to the edge side. The risk entropy-based hierarchical early warning decision module receives the status assessment results, calculates the equipment operation risk entropy value, and outputs an early warning signal of the corresponding level according to the risk entropy value range. The fault tracing and diagnosis reasoning module combines historical equipment operation data with fault knowledge graphs to locate the root cause of the fault and generate a diagnostic report. The early warning information distribution and closed-loop handling module distributes early warning information to the corresponding responsible persons according to the early warning level, tracks the handling progress, and updates the system status.

2. The industrial equipment condition monitoring and hierarchical early warning system based on edge computing according to claim 1, characterized in that: The edge-side multi-source data synchronous acquisition and preprocessing module includes a data acquisition unit, a time synchronization unit, a data cleaning unit, and a data standardization unit; The data acquisition unit connects to vibration, temperature, current, voltage, and pressure sensors via an industrial bus protocol to collect multimodal raw data during device operation. The time synchronization unit uses the system clock of the edge gateway as a reference to perform timestamp calibration and alignment on the data collected by different sensors. The data cleaning unit identifies and repairs missing values, outliers, and noise in the raw data. The data standardization unit converts the cleaned data into a unified numerical range and data format to generate a time-series dataset. The formula for calculating timestamp calibration and alignment is: in, The standard timestamp after data calibration for the i-th sensor device; The local timestamp when the i-th sensor device collected this data; The one-way transmission delay of data from the i-th sensor device to the edge gateway; This is the fixed offset between the local clock of the i-th sensor device and the reference clock of the edge gateway.

3. The industrial equipment condition monitoring and hierarchical early warning system based on edge computing according to claim 1, characterized in that: The edge-side lightweight feature extraction and state representation module includes a data framing unit, a lightweight feature extraction unit, and a feature fusion unit; The data framing unit divides the time-series dataset into segments according to fixed time windows to generate a continuous sequence of data frames; the lightweight feature extraction unit uses a deep separable convolutional neural network to extract local temporal and frequency domain features of each data frame; the feature fusion unit performs adaptive weighted fusion of the extracted multi-dimensional features to generate a multi-dimensional feature vector characterizing the device's operating status. The formula for adaptive weighted fusion is: in, This is the final feature vector after fusion; Let i be the feature vector of the i-th feature dimension; For the i-th feature dimension, the adaptive weights are used. The total number of feature dimensions; Adaptive weights The calculation method is as follows: in, is the information gain value of the i-th feature dimension, used to measure the ability of this feature to distinguish between normal and abnormal states of the device.

4. The industrial equipment condition monitoring and hierarchical early warning system based on edge computing according to claim 1, characterized in that: The edge-cloud collaborative status assessment and initial fault judgment module includes an edge-side initial status judgment unit, a cloud-side deep analysis unit, and a model iteration and update unit. The edge-side initial state judgment unit loads a pre-trained binary classification model, processes the multi-dimensional feature vectors, and outputs a preliminary judgment result of whether the device is normal or abnormal; the cloud-based deep analysis unit receives abnormal feature data and raw data fragments uploaded from the edge side, and uses a complex deep learning model to perform in-depth analysis of fault type and fault degree; the model iteration and update unit uses the fault data accumulated in the cloud to incrementally train the edge-side model and sends the updated model parameters to the edge gateway.

5. The industrial equipment condition monitoring and hierarchical early warning system based on edge computing according to claim 1, characterized in that: The risk entropy-based hierarchical early warning decision-making module includes a risk factor calculation unit, a risk entropy calculation unit, and an early warning level classification unit. The risk factor calculation unit extracts three core risk factors—fault occurrence probability, fault impact range, and fault severity—based on the status assessment results. The risk entropy calculation unit calculates the comprehensive risk entropy value of the equipment operation based on three core risk factors; The warning level classification unit divides the warning into multiple levels and outputs corresponding warning signals according to a preset risk entropy value range; The formula for calculating the comprehensive risk entropy is: in, The comprehensive risk entropy value for equipment operation; The probability of the k-th type of fault occurring; is the severity coefficient of the k-th type of fault; The influence range coefficient for the k-th type of fault; This represents the total number of possible fault types.

6. The industrial equipment condition monitoring and hierarchical early warning system based on edge computing according to claim 1, characterized in that: The fault tracing and diagnosis reasoning module includes a knowledge graph construction unit, a fault reasoning unit, and a diagnosis report generation unit. The knowledge graph construction unit integrates equipment design data, historical fault records, maintenance manuals, and expert experience to construct a fault knowledge graph that includes equipment entities, fault phenomena, fault causes, and maintenance measures. The fault reasoning unit uses a combination of rule-based reasoning and Bayesian network reasoning, combining current equipment state characteristics with the knowledge graph, to locate the root cause of the fault. The diagnostic report generation unit generates a diagnostic report that includes fault phenomena, fault causes, fault locations, and recommended maintenance measures based on the fault reasoning results.

7. The industrial equipment condition monitoring and hierarchical early warning system based on edge computing according to claim 1, characterized in that: The early warning information distribution and closed-loop handling module includes an early warning hierarchical distribution unit, a handling task management unit, and a handling effect evaluation unit; The early warning classification and distribution unit distributes early warning information and diagnostic reports to the responsible persons at the corresponding levels through multiple communication channels according to the early warning level; the handling task management unit creates and tracks fault handling tasks, and records key nodes and operation information in the handling process; the handling effect evaluation unit evaluates the handling effect after the fault handling is completed and feeds back the evaluation results to the system for optimizing the early warning model and inference rules.

8. The industrial equipment condition monitoring and hierarchical early warning system based on edge computing according to claim 2, characterized in that: The time synchronization unit adopts a one-way broadcast time synchronization mechanism. The edge gateway periodically sends a reference time synchronization packet to all sensing devices. The sensing devices update their local clock offset according to the received synchronization packet and perform real-time timestamp calibration on the subsequently collected data.

9. The industrial equipment condition monitoring and hierarchical early warning system based on edge computing according to claim 3, characterized in that: The lightweight feature extraction unit uses a deep separable convolutional neural network composed of multiple deep separable convolutional blocks stacked together. Each deep separable convolutional block contains a deep convolutional layer, a pointwise convolutional layer, a batch normalization layer, and an activation function layer. This reduces the number of model parameters and computational cost, enabling real-time feature extraction at the edge.

10. A method for industrial equipment condition monitoring and hierarchical early warning based on edge computing, characterized in that: Includes the following steps: Step 1: System Initialization and Data Acquisition The edge-side multi-source data synchronous acquisition and preprocessing module is activated to connect with various sensing devices in the industrial field, collect multimodal raw data of device operation, and generate time-series datasets after time synchronization, data cleaning and standardization processing. Step 2: Edge-side feature extraction and state representation The time-series dataset is input into the edge-side lightweight feature extraction and state representation module. Through data framing, lightweight feature extraction and adaptive feature fusion, a multi-dimensional feature vector representing the device's operating state is generated. Step 3: Edge-Cloud Collaboration Status Assessment The edge-side state initial judgment unit processes the multi-dimensional feature vector to determine the device's operating status; if the status is determined to be normal, the feature data is stored only on the edge side. If the condition is determined to be abnormal, the abnormal feature data and the original data fragments will be uploaded to the cloud for in-depth fault analysis. Step 4: Comprehensive Risk Assessment and Tiered Early Warning The risk entropy-based hierarchical early warning decision module receives the status assessment results, extracts the core risk factors and calculates the comprehensive risk entropy value, and outputs the corresponding level of early warning signal according to the risk entropy value range. Step 5: Fault Origin and Diagnosis The fault tracing and diagnosis reasoning module combines the fault knowledge graph with the current equipment status characteristics to locate the root cause of the fault and generate a diagnostic report. Step 6: Early Warning Distribution and Closed-Loop Response The early warning information distribution and closed-loop handling module distributes early warning information and diagnostic reports according to the early warning level, creates and tracks fault handling tasks, evaluates the handling effect after the handling is completed, and updates the system status.

Images