An edge-computing-based industrial equipment operation data analysis method and system

Abstract

The application discloses an industrial equipment operation data analysis method based on edge computing, applied to the technical field of edge computing, and through a data acquisition unit, multi-source heterogeneous original operation data can be efficiently acquired, and data reliability is improved through redundant acquisition and data fusion; a data preprocessing unit carries out denoising, normalization and time-frequency domain feature transformation on the original data, and extracts feature information accurately reflecting the equipment operation state; a state evaluation unit calls a dynamic threshold analysis model, realizes real-time and self-adaptive evaluation on the health state of the industrial equipment, and outputs an abnormal early warning result. Therefore, the application has the advantages of realizing local real-time processing of industrial equipment operation data, efficient feature extraction, dynamic and self-adaptive health state evaluation, and improving the timeliness and accuracy of fault early warning.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and in particular to a method and system for analyzing industrial equipment operation data based on edge computing. Background Technology

[0002] In modern industrial production scenarios, with the continuous improvement of industrial automation and intelligence, various industrial equipment generates massive amounts of multi-source, heterogeneous operational data during operation. This data contains key information on equipment operating status, performance, and potential faults.

[0003] However, traditional centralized data processing models typically require transmitting all collected device operation data to a remote cloud platform or central server for unified processing and analysis. This model has several drawbacks: First, the transmission of massive amounts of data over a wide area network (WAN) introduces significant network response latency, especially in industrial control and fault early warning scenarios where real-time performance is critical. This latency can prevent timely identification and response to abnormal device conditions, leading to missed opportunities for optimal intervention. Second, the continuous transmission of massive amounts of data places enormous pressure on network bandwidth, increasing communication costs and system complexity. More importantly, the lag in data processing makes it difficult to detect and warn of potential device faults or anomalies in real time and accurately, potentially causing production interruptions, equipment damage, or even safety accidents, resulting in severe economic losses and personnel risks.

[0004] Existing technologies still face significant challenges in addressing the localized real-time processing of industrial equipment operation data, efficient feature extraction, and dynamic, adaptive health status assessment under complex and variable operating conditions. In particular, there is an urgent need for a more advanced and efficient solution to effectively utilize edge computing capabilities to improve the timeliness and accuracy of fault warnings and to resolve the core contradiction of data processing lag and low resource utilization under centralized architecture. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the present invention aims to provide an industrial equipment operation data analysis method and system based on edge computing. This method enables localized real-time processing of industrial equipment operation data, efficient feature extraction, and dynamic and adaptive health status assessment. It significantly improves the timeliness and accuracy of fault warnings and effectively solves the core contradiction of data processing lag and low resource utilization under centralized architecture.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An edge computing-based method for analyzing industrial equipment operation data is provided for monitoring the status of industrial equipment. The method is applied to an edge processing node deployed near the network edge of the industrial equipment. The edge processing node is communicatively connected to a data acquisition unit and integrates a data preprocessing unit and a status assessment unit. The method includes the following steps: S1: Obtain the raw operating data of the industrial equipment through the data acquisition unit; S2: The data preprocessing unit performs localization processing on the original operating data and extracts feature information reflecting the operating status of the industrial equipment; S3: The dynamic threshold analysis model is invoked through the status assessment unit to assess the health status of the industrial equipment in real time based on the feature information and output the abnormal warning result.

[0007] Preferably, the data acquisition unit includes multiple distributed acquisition modules located in different parts of the industrial equipment, and step S1 includes: S11: The multi-source heterogeneous data generated by the industrial equipment is collected in parallel by the multiple distributed acquisition modules; S12: Perform protocol parsing and format standardization on the multi-source heterogeneous data to generate the original running data in a unified format.

[0008] Preferably, the multiple distributed acquisition modules include at least two redundant acquisition units deployed within the same physically associated area of ​​the industrial equipment, and step S11 includes: S111: Obtain redundant observation data collected by each of the redundant acquisition units for the same physical quantity; S112: Calculate the spatiotemporal correlation index between the redundant observation data, and evaluate the data confidence of each redundant acquisition unit based on the spatiotemporal correlation index; S113: The redundant observation data are weighted and fused according to the data confidence level to generate the original operating data. When the data confidence level of any redundant acquisition unit is lower than a preset confidence threshold, the weight corresponding to the redundant acquisition unit is reset to zero, and a sensor fault self-check command is triggered.

[0009] Preferably, step S2 includes: S21: The original operating data is denoised and normalized to obtain standardized operating data; S22: Perform time-frequency domain feature transformation on the standardized operating data to extract the energy distribution characteristics of the industrial equipment within a preset frequency band, which are used as the feature information.

[0010] Preferably, the edge processing node is deployed within the local area network of the production site where the industrial equipment is located, and the process before step S21 includes: S211: Establish a real-time data exchange path within the edge processing node; S212: The raw operating data is directly transmitted from the data acquisition unit to the data preprocessing unit through the real-time data exchange path to reduce the response delay of data during wide area network transmission.

[0011] Preferably, the dynamic threshold analysis model includes a baseline threshold and an environmental compensation factor, and step S3 includes: S31: Obtain the current environmental operating condition parameters of the industrial equipment, wherein the environmental operating condition parameters include at least the ambient temperature and the load fluctuation rate; S32: Determine the corresponding environmental compensation factor based on the environmental operating parameters, and use the environmental compensation factor to correct the benchmark threshold to obtain the target dynamic threshold; S33: Compare the feature information with the target dynamic threshold. When the feature information exceeds the target dynamic threshold, determine that the industrial equipment is in an abnormal state.

[0012] Preferably, the dynamic threshold analysis model further includes a long-term evolution trend term, and step S32 includes: S321: Obtain the feature information evolution sequence of the industrial equipment during its historical healthy operating cycle; S322: Extract the mean drift trajectory that changes with service duration from the feature information evolution sequence, and use it as the long-term evolution trend term; S323: The benchmark threshold is modified by using the environmental compensation factor and the long-term evolution trend term to obtain the target dynamic threshold, so that the target dynamic threshold matches the current service aging degree of the industrial equipment.

[0013] Preferably, step S3 is followed by: S4: Based on the severity of the abnormality corresponding to the abnormality warning result, send a corresponding control command to the control terminal of the industrial equipment. The control command is used to trigger the industrial equipment to perform shutdown protection, frequency reduction operation, or switch to standby mode.

[0014] Preferably, the edge processing nodes are multiple and constitute a decentralized distributed cluster, and the method further includes: S5: Real-time monitoring of resource utilization parameters of each edge processing node; S6: Dynamic load balancing of computing tasks is performed among multiple edge processing nodes based on the resource utilization parameters to eliminate single-point processing bottlenecks in centralized architecture.

[0015] Secondly, this application provides an industrial equipment operation data analysis system based on edge computing, the system comprising: The data acquisition unit is used to acquire the raw operating data of the industrial equipment; The data preprocessing unit is used to perform localization processing on the raw operating data and extract feature information reflecting the operating status of the industrial equipment; The status assessment unit is used to invoke the dynamic threshold analysis model to assess the health status of the industrial equipment in real time based on the feature information and output abnormal warning results.

[0016] Compared to existing technologies, the advantages of this invention are as follows: by deploying data acquisition, preprocessing, and status assessment functions at edge processing nodes close to industrial equipment, localized real-time data processing is achieved, avoiding problems such as high data transmission latency, high network bandwidth pressure, and delayed fault warnings in traditional centralized processing modes. Specifically, the data acquisition unit can efficiently acquire multi-source heterogeneous raw operating data and improve data reliability through redundant acquisition and data fusion; the data preprocessing unit performs noise reduction, normalization, and time-frequency domain feature transformation on the raw data to extract feature information that accurately reflects the operating status of the equipment; the status assessment unit calls a dynamic threshold analysis model, which not only considers environmental operating parameters for dynamic correction but also incorporates long-term evolution trend terms to match the aging degree of the equipment, thereby achieving real-time, adaptive assessment of the health status of industrial equipment and outputting abnormal warning results. In addition, the system can trigger corresponding control commands according to the severity of the abnormality and supports distributed clustering and dynamic load balancing of multiple edge nodes, further improving the robustness and processing capacity of the system. This application effectively solves the core contradiction of data processing lag and low resource utilization under centralized architecture. It has the advantages of realizing localized real-time processing of industrial equipment operation data, efficient feature extraction and dynamic and adaptive health status assessment, as well as improving the timeliness and accuracy of fault early warning. Attached Figure Description

[0017] Figure 1 The flowchart illustrates a method for analyzing industrial equipment operation data based on edge computing, as provided by this invention.

[0018] Figure 2 This invention provides a structural diagram of an industrial equipment operation data analysis system based on edge computing.

[0019] Figure 3 This is a schematic diagram of an industrial equipment operation data analysis system based on edge computing provided by the present invention.

[0020] In the diagram: 201, data acquisition unit; 202, data preprocessing unit; 203, status assessment unit. Detailed Implementation

[0021] The present invention will now be further described with reference to the accompanying drawings and specific embodiments: Please refer to Figure 1 An edge computing-based method for analyzing industrial equipment operation data is disclosed for monitoring the status of industrial equipment. The method is applied to edge processing nodes deployed near the network edge of the industrial equipment. The edge processing nodes are communicatively connected to a data acquisition unit and integrate a data preprocessing unit and a status assessment unit. The method includes the following steps: S1: Acquire raw operating data of industrial equipment through the data acquisition unit; S2: The data preprocessing unit performs localized processing on the raw operating data and extracts characteristic information reflecting the operating status of industrial equipment; S3: By calling the dynamic threshold analysis model through the status assessment unit, the health status of industrial equipment is assessed in real time based on feature information, and abnormal warning results are output.

[0022] In modern industrial production environments, large industrial equipment, such as CNC machine tools, stamping equipment, or chemical reaction vessels, are at the core of the production process. These devices generate massive amounts of multi-dimensional data during operation, including vibration, temperature, pressure, current, and noise. Traditional equipment condition monitoring methods typically employ a centralized data processing architecture, transmitting data collected from all field devices via a wide area network to a remote cloud data center for centralized analysis. This model has several inherent technical bottlenecks. First, data transmission latency is high; when equipment exhibits early signs of sudden failure, the long-distance transmission and centralized processing of data can take seconds or even longer, missing the optimal intervention window. Second, bandwidth pressure is enormous; the continuous uploading of massive amounts of raw data places a heavy burden on the factory's outbound bandwidth, resulting in high costs and potential network congestion. Third, the centralized processing model struggles to cope with complex changes in field operating conditions, limiting the accuracy of early warning models.

[0023] To overcome the aforementioned technical challenges, this application proposes a solution that brings data analytics capabilities down to the production floor. The core idea is to deploy one or more edge processing nodes near industrial equipment. This edge processing node can be an industrial computer, embedded system, or dedicated edge server with strong computing and storage capabilities, physically installed in a control cabinet next to the equipment or located within the factory's local area network. This near-source deployment fundamentally changes the data flow path and processing mode.

[0024] The method works as follows: First, the data acquisition unit deployed on the edge processing node begins operation. This unit is responsible for acquiring the raw operational data from various sensors installed on the industrial equipment. Subsequently, this raw data is sent directly to the data preprocessing unit within the edge processing node without needing to be transmitted remotely. The data preprocessing unit, acting as a localized data processing engine, performs a series of operations on the raw data, including cleaning, transformation, and feature extraction, refining the complex raw data into significantly reduced feature information that accurately characterizes the equipment's operating status. Finally, the status assessment unit invokes a dynamic threshold analysis model to analyze this feature information in real time. This model intelligently determines the current health status of the equipment, and upon detecting any abnormal signs, immediately generates and outputs an anomaly warning result locally. This warning result can be directly sent to the on-site equipment control system or the operator's monitoring terminal, achieving a millisecond-level response.

[0025] By completing the three core steps of data acquisition, preprocessing, and condition assessment at edge processing nodes close to the data source, the method proposed in this application constructs a localized closed-loop analysis system. This architecture significantly reduces the time delay from data generation to decision response, alleviates dependence on network bandwidth, and, because the processing is performed locally, can better integrate with real-time on-site conditions, thereby significantly improving the timeliness, reliability, and accuracy of industrial equipment condition monitoring.

[0026] Preferably, the data acquisition unit includes multiple distributed acquisition modules located in different parts of the industrial equipment, and step S1 includes: S11: Collects multi-source heterogeneous data generated by industrial equipment in parallel through multiple distributed acquisition modules; S12: Perform protocol parsing and format standardization on multi-source heterogeneous data to generate raw runtime data in a unified format.

[0027] In specific industrial applications, a single sensor or data source often cannot comprehensively reflect the overall operating status of a complex industrial machine. For example, for a large gantry milling machine, the health of its spindle, the smoothness of its feed system, the efficiency of its cooling system, and the load on its electrical system are all key aspects for evaluating its overall performance. Therefore, to achieve comprehensive and accurate monitoring, the data acquisition unit is usually designed as a distributed system, consisting of multiple acquisition modules deployed in different key parts of the equipment.

[0028] Taking this gantry milling machine as an example, the following deployment can be made: a three-axis accelerometer is installed near the spindle bearing housing to collect vibration signals, which is one of the distributed acquisition modules; a thermocouple temperature sensor is installed on the spindle motor housing to monitor the motor temperature, which is the second acquisition module; a Hall current sensor is connected to the power line of the servo motor driving the feed screw to monitor the load current, which is the third acquisition module; and a pressure sensor is installed on the main pipeline of the hydraulic cooling system, which is the fourth acquisition module.

[0029] These distributed acquisition modules work in parallel and continuously, each collecting data on different physical quantities. This results in so-called multi-source heterogeneous data. "Multi-source" refers to the data originating from different parts of the equipment, while "heterogeneous" refers to the vastly different types, formats, and generation methods of the data. For example, an accelerometer outputs a high-frequency analog voltage waveform, which needs to be sampled by an analog-to-digital converter to form time-series data; a temperature sensor may output a digitized temperature value via the Modbus industrial bus protocol; a current sensor also outputs an analog signal; and a pressure sensor may communicate via the CAN bus protocol. The sampling frequency, data accuracy, physical units, and communication protocols of these data all differ.

[0030] To enable subsequent data preprocessing units to handle this diverse range of data, protocol parsing and format standardization are essential. This process is completed at the aggregation node or the access layer of the edge processing node within the data acquisition unit. For each data source, the system invokes the corresponding protocol parser. For example, for temperature data accessed via the Modbus bus, the system parses it according to the Modbus protocol frame format, extracting the address, function code, and data field to ultimately obtain the specific temperature value. For raw sampled data streams from accelerometers, the system organizes them into an array with timestamps.

[0031] After parsing, all data is formatted into a unified internal data format. This format is typically a structured data object, such as JSON or a similar data structure. This structure contains standard fields, such as timestamps accurate to milliseconds, unique sensor identifiers, physical quantity names, standardized values, and uniform physical units. For example, data from different sensors will ultimately be converted to a format like {timestamp:"2023-10-27T10:00:00.123Z",sensor_id:"spindle_vibration_x", value: 0.05, unit: "g"}. Here, the timestamp field represents the data acquisition timestamp, formatted according to the ISO 8601 standard, accurate to milliseconds, and includes time zone information. The sensor_id field represents the unique identifier of the sensor, used to identify the specific sensor from which the data originated. The value field represents the physical quantity value acquired by the sensor. The unit field indicates the unit of the physical quantity corresponding to the value field. Through this parallel acquisition and format formatting, the system can efficiently acquire comprehensive, clean, and directly usable raw operational data streams.

[0032] Preferably, the multiple distributed acquisition modules include at least two redundant acquisition units deployed within the same physically associated area of ​​the industrial equipment, and step S11 includes: S111: Obtain redundant observation data collected by each redundant acquisition unit for the same physical quantity; S112: Calculate the spatiotemporal correlation index between redundant observation data, and evaluate the data confidence of each redundant acquisition unit based on the spatiotemporal correlation index; S113: The redundant observation data are weighted and fused according to the data confidence level to generate the original operation data. When the data confidence level of any redundant acquisition unit is lower than the preset confidence threshold, the weight corresponding to the redundant acquisition unit is reset to zero and a sensor fault self-check command is triggered.

[0033] In harsh industrial environments, sensors may experience data drift, distortion, or even complete failure due to vibration, contamination, aging, or circuit faults. Relying entirely on readings from a single sensor can lead to erroneous analysis results if that sensor malfunctions, potentially causing false or missed alarms and resulting in unnecessary downtime or safety incidents. To address this issue, this application introduces a redundant acquisition and data fusion mechanism.

[0034] Specifically, for some critical measuring points on the equipment, multiple data acquisition units with the same function will be deployed. For example, at the critical position of the spindle bearing housing of the gantry milling machine mentioned earlier, three independent acceleration sensors can be installed side by side. These three sensors constitute a redundant data acquisition unit, and their goal is to collect the same physical quantity, namely the vibration at that position.

[0035] During operation, the three sensors simultaneously collect data, generating three sets of redundant observation data sequences. Upon receiving these three sets of data, the edge processing node does not simply select one; instead, it executes a data verification and fusion process. First, it acquires the redundant observation data collected by the three sensors within the same time window. For example, at a certain moment, the readings of the three sensors are v1 = 0.05g, v2 = 0.052g, and v3 = 0.15g, respectively.

[0036] Next, the system calculates the spatiotemporal correlation index among these redundant observations. Temporal correlation can be assessed by analyzing whether the trends of the three data sequences are consistent over a past period, for example, by calculating their cross-correlation coefficient. Spatial correlation is based on the prior knowledge that their physical locations are close, assuming that their readings should be very similar under normal circumstances. A simple spatiotemporal correlation index can be calculated by measuring the standard deviation or the absolute difference between the three data points at the current time. In the example above, v1 and v2 are very close, while v3 is a clear outlier.

[0037] Based on this correlation index, the system evaluates a data confidence level for each redundant acquisition unit. The confidence level is a value between 0 and 1, reflecting the degree of consistency between the sensor data and other redundant sensor data. In the example above, the confidence levels for v1 and v2 might be rated as 0.98, while the confidence level for v3 might only be 0.2.

[0038] Finally, the system performs weighted fusion of redundant observation data based on the assessed data confidence level to generate a more reliable original operational data. The fusion calculation formula can be V_fused = (w1*v1 + w2*v2 + w3*v3) / (w1 + w2 + w3), where the weight w is a function of the data confidence level. During this process, the system pre-sets a confidence threshold, for example, 0.5. When the confidence level of a sensor is lower than this threshold, it means that its data is highly likely to be unreliable. At this time, the system will forcibly set its corresponding weight w to zero. In the example above, since the confidence level of v3 is 0.2, which is lower than 0.5, w3 is set to 0. The final fused vibration value will be determined only by the data from v1 and v2, thus eliminating the interference of abnormal data v3. Simultaneously, setting w3 to zero triggers a sensor fault self-check command. This command sends an alarm to the maintenance system, indicating that the third accelerometer sensor may have malfunctioned and needs to be checked or calibrated. This mechanism not only improves the quality of input data, but also enables online self-diagnosis of sensor health.

[0039] Preferably, step S2 includes: S21: Denoise and normalize the raw operating data to obtain standardized operating data; S22: Perform time-frequency domain feature transformation on standardized operating data to extract the energy distribution characteristics of industrial equipment within a preset frequency band as feature information.

[0040] After acquiring high-quality raw operational data, the data preprocessing unit begins in-depth processing, aiming to extract characteristic information from the raw data that can sensitively reflect changes in the equipment's health status. This process typically involves two main stages.

[0041] The first stage is data cleaning and standardization. Even after redundancy fusion, raw operational data may still contain noise from sources such as electromagnetic interference and environmental vibration. Therefore, data denoising is necessary. For example, for collected vibration time-series data, a digital filter, such as a bandpass filter, can be applied to remove frequency components unrelated to equipment failure, or more advanced methods like wavelet denoising can be used to smooth the signal. After denoising, normalization is required to eliminate differences in dimensions and numerical ranges between different physical quantities. For example, temperature data might range from 20 to 100 degrees Celsius, while current data might range from 5 to 50 amperes. If these raw values ​​are used directly for analysis, features with large numerical ranges will disproportionately affect the model results. Normalization, such as using a min-max normalization method, linearly maps all data to the interval between 0 and 1. After this processing, all features are on the same starting line, facilitating fair evaluation of the model later. After this stage, we obtain clean and scale-uniform standardized operational data.

[0042] The second stage is the core feature extraction. For dynamic signals like vibration and acoustics, simply observing their waveforms in the time domain often fails to detect subtle signs of early faults. A tiny bearing crack might appear as negligible fluctuations in the time domain waveform, but it will produce a very specific frequency component in the frequency domain. Therefore, time-frequency domain feature transformation is needed for standardized operating data. Commonly used methods include Short-Time Fourier Transform (STFT) or wavelet transform. These transforms can reveal how the frequency components of a signal change over time. Taking the STFT as an example, it divides long time series data into many short, overlapping time windows and performs a Fourier transform on the data within each window, thus obtaining a spectrum. The horizontal axis of this graph represents time, the vertical axis represents frequency, and the color intensity represents the energy intensity at that time and frequency point.

[0043] Based on engineering experience and equipment failure mechanism analysis, those skilled in the art generally know that specific failure modes correspond to energy anomalies in specific frequency bands. For example, for a rolling bearing, damage to its inner and outer rings, balls, or cage will generate energy peaks at specific frequencies in the vibration spectrum. These frequencies can be precisely calculated based on the bearing's geometry and rotational speed. Therefore, the final step in feature extraction is to extract the energy distribution characteristics within these preset key frequency bands from the time-frequency domain analysis results. For example, the system can calculate the average energy value or the height of the energy peak in the 245 Hz to 255 Hz range, which is the characteristic frequency band of the inner ring failure. These calculated values, such as the energy in the inner ring failure frequency band, constitute the final feature information. Compared to the original time series containing thousands of data points, this single energy value has a very small data volume but extremely high information content, directly and sensitively reflecting the health status of the equipment.

[0044] Preferably, the edge processing node is deployed within the local area network of the production site where the industrial equipment is located, and the process prior to step S21 includes: S211: Establish a real-time data exchange path within the edge processing node; S212: The raw operating data is directly transmitted from the data acquisition unit to the data preprocessing unit through a real-time data exchange path to reduce the response delay during data transmission over the wide area network.

[0045] To maximize the real-time advantages of edge computing, in addition to deploying processing nodes close to the devices on a macro level, it is also necessary to optimize the data flow efficiency within the nodes on a micro level. Deploying edge processing nodes within the local area network of the production site, such as the Ethernet in the workshop, ensures that the data transmission latency from sensors to edge nodes is controlled at an extremely low level, avoiding the uncertainty and high latency of public network transmission.

[0046] However, within the hardware entity of the edge processing node, there is still an internal transmission process from the data acquisition interface to its use by the preprocessing software module. While using traditional operating system-based network protocol stacks for data exchange, such as communication between different processes via TCP / IP, is feasible, it still introduces processing overhead and latency. To achieve ultimate real-time performance, this application proposes a more direct data exchange method.

[0047] In the software architecture design of edge processing nodes, a dedicated real-time data exchange path can be pre-established. This path bypasses conventional, non-real-time data transmission mechanisms. There are several implementation methods. One method utilizes shared memory technology. After receiving sensor data, the driver for the data acquisition unit does not encapsulate it into network packets, but instead writes it directly to a physical memory area shared by the data acquisition unit and the data preprocessing unit. The data preprocessing unit then monitors this shared memory in real time using polling or interrupt methods. Once new data is detected, it immediately reads it and begins processing. This method virtually eliminates the overhead of data copying and protocol encapsulation, achieving latency at the microsecond level.

[0048] Another approach is to utilize inter-process communication (IPC) mechanisms provided by real-time operating systems (RTOS), such as message queues or semaphores, which are designed for high determinism and low latency scenarios. Through this established real-time data exchange path, raw operational data acquired from the data acquisition unit can flow instantly to the data preprocessing unit with near-zero copying. This optimization ensures that the starting point of the data processing flow—the connection between data acquisition and preprocessing—is as efficient as possible, laying a solid foundation for low-latency performance throughout the entire analysis chain, thus truly achieving instantaneous response to the status of industrial equipment.

[0049] Preferably, the dynamic threshold analysis model includes a baseline threshold and an environmental compensation factor, and step S3 includes: S31: Obtain the current environmental operating condition parameters of the industrial equipment. The environmental operating condition parameters shall include at least the ambient temperature and the load fluctuation rate. S32: Determine the corresponding environmental compensation factor based on the environmental operating parameters, and use the environmental compensation factor to correct the benchmark threshold to obtain the target dynamic threshold; S33: Compare the feature information with the target dynamic threshold. When the feature information exceeds the target dynamic threshold, determine that the industrial equipment is in an abnormal state.

[0050] After extracting key feature information, the condition assessment unit needs a reliable benchmark to determine whether this feature information is within the normal range. Using a fixed, unchanging threshold is not feasible in complex industrial scenarios. This is because the operating status of industrial equipment is significantly affected by its environment and workload. For example, a compressor operating in the high temperatures of summer will have higher internal temperatures and vibrations than in the low temperatures of winter; similarly, its vibration energy at full load is much higher than at no-load. If a fixed threshold is used, it is likely to generate a large number of false alarms in summer or at full load, while failing to detect true early faults in winter or at light load.

[0051] To address this issue, this application employs a dynamic threshold analysis model. The core idea of ​​this model is to allow the alarm threshold to adaptively adjust based on the device's current operating conditions. The model first includes a baseline threshold. This baseline threshold is determined during the device's learning phase, i.e., by allowing the device to operate healthily under standard, ideal conditions (e.g., standard load, standard ambient temperature) for a period of time, collecting and calculating the statistical distribution of its characteristic information, for example, taking its mean plus three standard deviations as an initial baseline.

[0052] In actual monitoring, the status assessment unit not only focuses on the feature information transmitted from the data preprocessing unit, but also simultaneously acquires the environmental operating parameters of the equipment. These parameters can be obtained through additional sensors, such as installing a temperature and humidity sensor next to the equipment to obtain the ambient temperature, or directly reading information such as the current motor load, speed, and production cycle from the equipment's control system (such as a PLC). Load fluctuation rate is also an important operating parameter, which can be obtained by calculating the range of load value changes over a short period of time.

[0053] Next, the model determines a corresponding environmental compensation factor based on real-time acquired environmental operating parameters. This compensation factor is not a fixed value, but a pre-established mapping relationship, which can be a multidimensional lookup table or a mathematical function. This mapping relationship is obtained through extensive historical data analysis or mechanistic modeling. For example, the model may know that for every 10-degree Celsius increase in ambient temperature, the normal baseline of vibration characteristics will rise by 5%; for every 20% increase in load rate, the normal baseline will rise by 15%. Based on the currently acquired ambient temperature and load rate, the model calculates a comprehensive environmental compensation factor.

[0054] Then, this environmental compensation factor is used to correct the baseline threshold, resulting in a target dynamic threshold under the current operating conditions. The correction can be done through multiplication or addition; for example, target dynamic threshold = baseline threshold × environmental compensation factor. This calculated target dynamic threshold is the normal upper limit of the equipment under the current specific operating conditions.

[0055] In the final step, the condition assessment unit compares the real-time extracted feature information with this newly generated target dynamic threshold. If the feature value exceeds the target dynamic threshold, even if it may be lower than the threshold under other operating conditions, the system will accurately determine that the equipment is currently in an abnormal state. Conversely, if the feature value is high but still within the allowable range of the dynamic threshold, the system will not generate a false alarm. This mechanism enables condition assessment to accurately adapt to changes in operating conditions, greatly improving the accuracy and reliability of early warnings.

[0056] Preferably, the dynamic threshold analysis model further includes a long-term evolution trend term, and step S32 includes: S321: Obtain the evolution sequence of characteristic information of industrial equipment during its historical healthy operating cycle; S322: Extract the mean drift trajectory that changes with service duration from the feature information evolution sequence as a long-term evolution trend term; S323: The baseline threshold is corrected by using environmental compensation factors and long-term evolution trend terms to obtain the target dynamic threshold, so that the target dynamic threshold matches the current service aging level of industrial equipment.

[0057] Besides being affected by external environment and operating conditions, the performance and condition of industrial equipment also undergo irreversible and slow changes over time, a process known as service aging or performance degradation. Even under identical operating conditions, the normal operating characteristic baseline of a brand-new piece of equipment and one that has been running for five years may differ. For example, due to long-term mechanical wear, the normal vibration level of a bearing will slowly and continuously increase. If this long-term evolutionary trend is ignored and the baseline thresholds established early in the equipment's lifespan are still used, the system may frequently generate false alarms due to normal wear as the equipment ages.

[0058] To enable the model to adapt to the entire lifecycle of the device, this application further introduces a long-term evolution trend term into the dynamic threshold model. This trend term is specifically used to quantify and compensate for characteristic baseline drift caused by the aging of the device itself.

[0059] The first step in achieving this functionality is to acquire the evolution sequence of the device's characteristic information over its long history of healthy operation. Edge processing nodes store and record the characteristic information extracted during each healthy operation, along with a timestamp. This creates a vast historical database that comprehensively records the device's characteristic changes from its initial operation to its gradual decline.

[0060] Next, the condition assessment unit will periodically or as needed analyze this historical feature information evolution sequence. By applying time series analysis techniques, such as moving averages, linear regression, or more complex machine learning models, the system can extract the overall trend of the mean value of the feature information changing with service duration (i.e., operating time). This trend curve, or its mathematical expression, is defined as the long-term evolution trend term. For example, the model might find that the baseline value of a certain vibration feature steadily increases by an average of 0.01 units per month. This drift over time is a quantitative description of the equipment aging process.

[0061] When calculating the target dynamic threshold, this long-term evolution trend term, along with the aforementioned environmental compensation factor, is applied to the baseline threshold. The revised formula becomes more refined, for example, Target Dynamic Threshold = (Baseline Threshold + Long-Term Evolution Trend Term) × Environmental Compensation Factor. The long-term evolution trend term here is calculated based on the equipment's current total service life. For example, if the equipment has been operating for 30 months, then the value of the trend term would be 30 × 0.01 = 0.3.

[0062] Through this dual correction, the final target dynamic threshold not only adapts to the device's current external operating environment but also matches the device's current internal aging level. This threshold truly represents the device's health boundary at this moment, under these specific circumstances. Comparing real-time feature information with such a highly customized threshold allows for the most accurate differentiation between slow changes caused by normal aging and abnormal jumps caused by sudden failures, thus maintaining extremely high early warning accuracy throughout the device's entire lifecycle.

[0063] After the state assessment unit outputs an anomaly warning result, in order to form a complete closed-loop control system of perception-analysis-decision-execution, the method also includes: S4: Based on the severity of the abnormality corresponding to the abnormality warning result, send the corresponding control command to the control terminal of the industrial equipment. The control command is used to trigger the industrial equipment to perform shutdown protection, frequency reduction operation, or switch to standby mode.

[0064] Simply issuing alarm messages is insufficient in many scenarios. The system needs to be able to automatically and differentiate intervention measures based on the urgency and severity of the alarm in order to maximize equipment safety and ensure production continuity.

[0065] When the status assessment unit determines that the device is in an abnormal state, the output warning result is not just a simple yes / no signal, but also includes a quantitative assessment of the severity of the abnormality. This severity can be determined based on the extent to which the characteristic information exceeds a dynamic threshold. For example, exceeding the threshold by less than 10% is defined as a mild abnormality, exceeding it by 10% to 50% is defined as a moderate abnormality, and exceeding it by more than 50% is defined as a severe abnormality.

[0066] The edge processing node integrates a decision logic module. This module selects and sends corresponding control commands from a preset control strategy library based on the severity of the received anomaly. These commands are sent directly to the control terminal of the industrial equipment, such as a programmable logic controller (PLC), via an industrial fieldbus (such as Profinet or EtherCAT).

[0067] Specifically, the control strategy is hierarchical: For severe anomalies, this usually indicates an impending or already occurring serious fault that could cause permanent equipment damage or a safety accident. In this case, the decision module will immediately send a highest-priority shutdown protection command. Upon receiving this command, the PLC will immediately execute the emergency shutdown procedure, cutting off the power source to safely stop the equipment in the shortest possible time and prevent the fault from escalating.

[0068] For moderate anomalies, this may indicate a significant equipment malfunction, but there is still some buffer time. To reduce equipment load and slow the progression of the fault while maintaining basic production, the decision module will send a frequency reduction operation command. For example, for a CNC machine tool, this command may reduce its spindle speed and feed rate to 50% of their rated values. This allows some non-critical machining tasks to continue even when the machine is malfunctioning, while giving maintenance personnel time to prepare and intervene.

[0069] For minor anomalies, which are often very early signs of malfunctions or recoverable temporary issues, the decision module may send a command to switch to standby mode to avoid production interruption. For example, if a main conveyor belt on a production line experiences a slight vibration anomaly, the system can instruct the system to switch to an adjacent standby conveyor belt to continue production, while simultaneously issuing a detailed diagnostic report and maintenance recommendations for the main conveyor belt.

[0070] Through this automated, hierarchical response mechanism based on the severity of anomalies, edge computing nodes transform from simple monitors into intelligent agents with preliminary autonomous control capabilities, enabling refined and preventative management of industrial equipment.

[0071] In some large-scale production scenarios, to improve the scalability and robustness of the entire monitoring system, multiple edge processing nodes can be used to form a decentralized distributed cluster. This method also includes: S5: Real-time monitoring of resource utilization parameters of each edge processing node; S6: Dynamically load balance computing tasks across multiple edge processing nodes based on resource utilization parameters to eliminate single-point processing bottlenecks in centralized architectures.

[0072] When the number of devices requiring monitoring within a factory is very large, relying on a single edge processing node may encounter computing power bottlenecks. Moreover, if that node experiences a hardware or software failure, the monitoring tasks of all connected devices will be interrupted, creating a single point of failure. To solve this problem, multiple edge processing nodes can be organized into a distributed cluster.

[0073] In this architecture, multiple edge processing nodes are interconnected via a high-speed local area network, forming a decentralized, peer-to-peer collaborative computing resource pool. Each node has the ability to independently process tasks, while also being aware of the presence and status of other nodes in the cluster.

[0074] To enable collaborative operation of the cluster, the system executes a continuous monitoring and scheduling process. First, each edge processing node monitors its own resource utilization parameters in real time, including the CPU load percentage, memory usage, network interface throughput, and (if equipped) GPU utilization. The nodes periodically broadcast this status information to all other nodes in the cluster.

[0075] Each node in the cluster runs a distributed task scheduling algorithm. When a new data analysis task (e.g., real-time analysis of vibration data from a newly connected device) needs to be assigned, the scheduling algorithm does not simply follow the principle of physical proximity, but rather considers the overall load of the cluster. It checks the real-time resource utilization parameters of all nodes and finds the node with the most idle resources and the most abundant computing power.

[0076] Then, the scheduling algorithm dynamically assigns this new computational task to the selected, least busy node. This means that even if a device is physically closest to node A, if node A's CPU load is as high as 90% while node B's load is only 30%, then the data flow from that device will be routed to node B for processing. This dynamic load balancing of computational tasks based on real-time resource utilization ensures that the computational pressure is evenly distributed throughout the cluster, preventing any single node from becoming a performance bottleneck due to task overload.

[0077] Furthermore, this decentralized cluster architecture inherently possesses high availability. If any node in the cluster goes offline due to a failure, the remaining nodes will immediately detect it. The computing tasks originally running on the failed node will be automatically and smoothly migrated to other healthy nodes by the cluster's self-healing mechanism, thus ensuring the continuity and uninterrupted nature of the monitoring service and completely eliminating the vulnerability of traditional single-point architectures.

[0078] Please refer to Figure 2 , Figure 3 This application provides an industrial equipment operation data analysis system based on edge computing, the system comprising: Data acquisition unit 201 is used to acquire raw operating data of industrial equipment; The data preprocessing unit 202 is used to perform localized processing on the raw operating data and extract feature information reflecting the operating status of industrial equipment; The status assessment unit 203 is used to call the dynamic threshold analysis model to assess the health status of industrial equipment in real time based on feature information and output abnormal warning results.

[0079] This system serves as the physical carrier for implementing the aforementioned methods. The data acquisition unit 201 is the system's sensing layer, consisting of sensors (such as accelerometers, thermocouples, and pressure transmitters) deployed in various parts of the industrial equipment, as well as hardware interface modules responsible for signal conditioning, analog-to-digital conversion, and protocol conversion. These components work together to transform the physical operating status of the equipment into a digital raw data stream.

[0080] The data preprocessing unit 202 and the state evaluation unit are typically run as software modules on the processors of edge processing nodes deployed in the field. The edge processing node is the core of the system and is a rugged industrial-grade computer. The data preprocessing unit 202 is responsible for executing the aforementioned algorithms such as denoising, normalization, and time-frequency domain transformation, compressing and refining massive amounts of raw data into a small amount of high-information-density feature information.

[0081] The status assessment unit 203 implements a dynamic threshold analysis model. Internally, it maintains a mapping database of baseline thresholds, environmental conditions, and compensation factors, as well as a historical database of long-term equipment aging trends. Based on the input feature information and various compensation factors, this unit performs real-time health status assessments and generates early warnings.

[0082] These three units are tightly integrated into the edge processing node, forming a highly efficient and collaborative working whole. The data flow within the system starts from the data acquisition unit 201, passes through the data preprocessing unit 202, and finally reaches the status assessment unit 203. The entire process is completed locally in a closed loop, ensuring rapid response and accurate judgment of changes in the status of industrial equipment.

[0083] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

Claims

1. A method for analyzing industrial equipment operation data based on edge computing, used for status monitoring of industrial equipment, characterized in that, The method is applied to an edge processing node, which is deployed near the network edge of the industrial equipment. The edge processing node is communicatively connected to a data acquisition unit and integrates a data preprocessing unit and a status assessment unit. The method includes the following steps: S1: Obtain the raw operating data of the industrial equipment through the data acquisition unit; S2: The data preprocessing unit performs localization processing on the raw operating data and extracts feature information reflecting the operating status of the industrial equipment; S3: The dynamic threshold analysis model is invoked through the status assessment unit to assess the health status of the industrial equipment in real time based on the feature information and output the abnormal warning result.

2. The method for analyzing industrial equipment operation data based on edge computing according to claim 1, characterized in that, The data acquisition unit includes multiple distributed acquisition modules located in different parts of the industrial equipment. Step S1 includes: S11: The multi-source heterogeneous data generated by the industrial equipment is collected in parallel by the multiple distributed acquisition modules; S12: Perform protocol parsing and format standardization on the multi-source heterogeneous data to generate the original running data in a unified format.

3. The method for analyzing industrial equipment operation data based on edge computing according to claim 1, characterized in that, Multiple distributed acquisition modules include at least two redundant acquisition units deployed within the same physically associated area of ​​the industrial equipment. Step S11 includes: S111: Obtain redundant observation data collected by each of the redundant acquisition units for the same physical quantity; S112: Calculate the spatiotemporal correlation index between the redundant observation data, and evaluate the data confidence of each redundant acquisition unit based on the spatiotemporal correlation index; S113: The redundant observation data are weighted and fused according to the data confidence level to generate the original operating data. When the data confidence level of any redundant acquisition unit is lower than a preset confidence threshold, the weight corresponding to the redundant acquisition unit is reset to zero, and a sensor fault self-check command is triggered.

4. The method for analyzing industrial equipment operation data based on edge computing according to claim 1, characterized in that, Step S2 includes: S21: The original operating data is denoised and normalized to obtain standardized operating data; S22: Perform time-frequency domain feature transformation on the standardized operating data to extract the energy distribution characteristics of the industrial equipment within a preset frequency band, which are used as the feature information.

5. The method for analyzing industrial equipment operation data based on edge computing according to claim 4, characterized in that, The edge processing node is deployed within the local area network of the production site where the industrial equipment is located. Before step S21, the following steps are included: S211: Establish a real-time data exchange path within the edge processing node; S212: The raw operating data is directly transmitted from the data acquisition unit to the data preprocessing unit through the real-time data exchange path to reduce the response delay of data during wide area network transmission.

6. The method for analyzing industrial equipment operation data based on edge computing according to claim 1, characterized in that, The dynamic threshold analysis model includes a baseline threshold and an environmental compensation factor. Step S3 includes: S31: Obtain the current environmental operating condition parameters of the industrial equipment, wherein the environmental operating condition parameters include at least the ambient temperature and the load fluctuation rate; S32: Determine the corresponding environmental compensation factor based on the environmental operating parameters, and use the environmental compensation factor to correct the benchmark threshold to obtain the target dynamic threshold; S33: Compare the feature information with the target dynamic threshold. When the feature information exceeds the target dynamic threshold, determine that the industrial equipment is in an abnormal state.

7. The method for analyzing industrial equipment operation data based on edge computing according to claim 6, characterized in that, The dynamic threshold analysis model also includes a long-term evolution trend term, and step S32 includes: S321: Obtain the feature information evolution sequence of the industrial equipment during its historical healthy operating cycle; S322: Extract the mean drift trajectory that changes with service duration from the feature information evolution sequence, and use it as the long-term evolution trend term; S323: The benchmark threshold is modified by using the environmental compensation factor and the long-term evolution trend term to obtain the target dynamic threshold, so that the target dynamic threshold matches the current service aging degree of the industrial equipment.

8. The method for analyzing industrial equipment operation data based on edge computing according to claim 1, characterized in that, Step S3 is followed by: S4: Based on the severity of the abnormality corresponding to the abnormality warning result, send a corresponding control command to the control terminal of the industrial equipment. The control command is used to trigger the industrial equipment to perform shutdown protection, frequency reduction operation, or switch to standby mode.

9. The method for analyzing industrial equipment operation data based on edge computing according to claim 1, characterized in that, The edge processing nodes are multiple and constitute a decentralized distributed cluster, and the method further includes: S5: Real-time monitoring of resource utilization parameters of each edge processing node; S6: Dynamic load balancing of computing tasks is performed among multiple edge processing nodes based on the resource utilization parameters to eliminate single-point processing bottlenecks in centralized architecture.

10. An industrial equipment operation data analysis system based on edge computing, characterized in that, The system includes: The data acquisition unit is used to acquire the raw operating data of the industrial equipment; The data preprocessing unit is used to perform localization processing on the raw operating data and extract feature information reflecting the operating status of the industrial equipment. The status assessment unit is used to invoke the dynamic threshold analysis model to assess the health status of the industrial equipment in real time based on the feature information and output abnormal warning results.

Images