An industrial middle platform and edge computing-based device monitoring and prediction method and system

Abstract

The application discloses a kind of equipment monitoring prediction method and system based on industry middle platform and edge computing, the system uses monitoring acquisition unit and real-time collection equipment vibration, temperature and current data, is analyzed by innovative three-layer prediction architecture: edge layer is deployed lightweight threshold rule model and carries out real-time anomaly detection, middle platform layer is deployed XGBoost model and executes trend analysis, cloud layer is deployed time convolution network model and realizes depth time series analysis, and corresponding hierarchical early warning mechanism is established.The application solves the technical problems, such as that early warning mechanism is not intelligent enough in traditional equipment monitoring, prediction model is single, data preprocessing is not perfect enough and the like.Through the cooperative cooperation of three-layer prediction architecture, all-around monitoring is realized, system response delay is reduced, and early warning accuracy is improved;Through the combination of hierarchical early warning mechanism and parameter correlation analysis, system reliability is enhanced, and equipment operation and maintenance efficiency is improved;Edge computing mode is used, and network bandwidth demand is effectively reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial equipment monitoring, and in particular to an equipment monitoring and prediction method and system based on industrial middle platform and edge computing. BACKGROUND

[0002] With the rapid development of Industry 4.0 and intelligent manufacturing, equipment health management and predictive maintenance have become a key link to ensure the stable operation of industrial production. Unplanned downtime of industrial equipment causes huge economic losses to manufacturing industry every year, while the use of predictive maintenance can significantly reduce equipment downtime and greatly improve equipment utilization. The monitoring and prediction system of industrial equipment can effectively prevent equipment failure, reduce unplanned downtime and improve equipment utilization by collecting real-time equipment operation data and combining artificial intelligence and big data analysis technology. At present, there are many technical solutions in the field of industrial equipment monitoring and prediction, but in actual application, there are still many challenges:

[0003] 1) From the perspective of prediction model architecture, traditional single model prediction schemes often use a unified algorithm framework to handle all prediction tasks, which is difficult to balance the needs of real-time and prediction accuracy. For example, motor bearing fault warning requires extremely fast response speed, while equipment life prediction requires comprehensive historical data for deep analysis. Using the same prediction model cannot meet the needs of these two types of requirements at the same time, resulting in low efficiency of system resource utilization and unsatisfactory prediction results.

[0004] 2) From the perspective of system deployment, pure cloud solutions have strong computing power, but there is a serious problem of system response delay. For example, a large steel enterprise produces a large amount of real-time data from its rolling mill equipment. If all the data is uploaded to the cloud for processing, it will cause a large network delay. This delay not only causes a huge pressure on network bandwidth, but also affects the stability of the monitoring system due to network fluctuations, which cannot meet the strict requirements of real-time performance in industrial field.

[0005] 3) From the perspective of monitoring mechanism, the existing simple threshold monitoring scheme lacks the ability to analyze the correlation between multiple parameters of the equipment. For example, in industrial pump systems, vibration, temperature, current and other parameters often have complex coupling relationships. Independent monitoring of a single parameter cannot detect complex fault modes such as vibration abnormalities caused by bearing overheating, which can easily cause false alarms or missed alarms, affecting the accuracy of the warning. Multi-parameter correlation analysis can significantly improve the accuracy of fault warning.

[0006] 4) From the perspective of early warning mechanism, the prior art generally lacks systematic hierarchical early warning design. In actual industrial scenarios, equipment abnormalities can be divided into different levels such as early warning, performance degradation, fault warning, etc., and different maintenance strategies need to be taken for each level. However, the existing binary judgment mechanism (normal / abnormal) cannot provide such fine-grained early warning information, resulting in unreasonable allocation of maintenance resources. For example, for slight performance degradation, it may only need to be handled at the next planned maintenance, while serious failure requires immediate shutdown for repair.

[0007] In addition, with the development of edge computing technology, its application in industrial scenarios is becoming mature, and edge computing can significantly reduce data processing delay and greatly reduce bandwidth occupancy.

[0008] Therefore, there is an urgent need for a device monitoring and prediction system that can comprehensively utilize the advantages of edge computing and industrial middle platform, has multi-level prediction capability and intelligent early warning mechanism. The system should be able to realize the coordinated cooperation of three-level prediction architecture, the correlation analysis of multiple parameters, and the intelligent grading of early warning mechanism, so as to comprehensively improve the monitoring and prediction effect and operation and maintenance efficiency of industrial equipment. SUMMARY

[0009] The problem to be solved by the present application is how to provide a device monitoring and prediction method and system based on industrial middle platform and edge computing, to solve the technical problems of large response delay, single early warning mechanism, insufficient parameter correlation analysis, etc. in traditional device monitoring and prediction.

[0010] To solve the above technical problems, the technical solution adopted by the present application is:

[0011] A device monitoring and prediction method based on industrial middle platform and edge computing, characterized in that it comprises the following steps: S1. Collecting running data including vibration data, temperature data and current data, setting sampling time sequence according to device running characteristics, and collecting data according to preset sampling frequency and preliminarily storing in edge controller;

[0012] S2. Distributing the running data to a three-level prediction model after preprocessing, the preprocessing including data cleaning, data repair and exception handling of the running data, the three-level prediction model including a lightweight threshold rule model deployed in the edge layer, an XGBoost model deployed in the middle platform layer and a time convolution network model deployed in the cloud layer; wherein the middle platform layer is an industrial middle platform;

[0013] S3. Utilize the three-layer prediction model to perform equipment state analysis and prediction, wherein the lightweight threshold rule model of the edge layer performs anomaly detection based on the operation data and the parameter benchmark interval, the XGBoost model of the middle layer performs feature analysis and trend prediction on the operation data, and performs anomaly judgment based on the parameter benchmark interval, and the time convolution network model of the cloud layer performs deep time series analysis and long-term prediction on the operation data, and performs risk assessment based on the parameter benchmark interval;

[0014] The parameter benchmark interval includes a vibration parameter benchmark interval, a temperature parameter benchmark interval, and a current parameter benchmark interval. The vibration parameter benchmark interval is a range of mean plus or minus standard deviation calculated based on the vibration data. The temperature parameter benchmark interval is a range of mean plus or minus standard deviation calculated based on the temperature data. The current parameter benchmark interval is a range of mean plus or minus standard deviation calculated based on the current data.

[0015] S4. Establish a hierarchical early warning mechanism based on the prediction results of the three-layer prediction model. When the edge layer detects an anomaly in the parameter benchmark interval, a first-level warning is triggered. When the middle layer detects an abnormal trend in the operation data, a second-level warning is triggered. When the cloud layer predicts a running risk, a third-level warning is triggered.

[0016] S5. Perform response processing according to the warning results of the hierarchical early warning mechanism. When the first-level warning is triggered, send a device stop running instruction to the device control system and push an emergency alarm information to the operation and maintenance personnel. When the second-level warning is triggered, generate a device inspection task and adjust the device maintenance plan. When the third-level warning is triggered, the cloud layer generates a device operation and maintenance analysis report based on the prediction data.

[0017] Further, the lightweight threshold rule model of the S3 edge layer includes parameter benchmark interval calculation and anomaly detection based on the operation data.

[0018] A sliding window based on time series is used to dynamically monitor the operation data. The sliding window moves on the time axis according to a preset step size and calculates the mean and standard deviation within the sliding window. When it is detected that a parameter exceeds the parameter benchmark interval, parameter correlation analysis is performed.

[0019] The parameter correlation analysis includes counting the frequency of the vibration data, the temperature data, and the current data exceeding the parameter benchmark interval within the sliding window, and generating a parameter anomaly correlation matrix. When a parameter exceeds the parameter benchmark interval, it is checked whether other parameters in the parameter anomaly correlation matrix also appear abnormal. If it is confirmed that the parameter is abnormal, the warning mechanism in S4 is triggered.

[0020] In a time period in which the parameter anomaly correlation matrix does not confirm an anomaly, periodically add the statistical value of the sliding window to historical data to recalculate the parameter reference interval.

[0021] Further, the XGBoost model of the S3 middle layer performs feature analysis and trend prediction on the running data, specifically including:

[0022] A feature matrix is constructed for the vibration data, the temperature data and the current data in the running data respectively, wherein a frequency spectrum distribution matrix of the vibration data is obtained by performing Fourier transform on the vibration data, a statistical feature matrix of the temperature data is constructed by calculating the mean and standard deviation of the temperature data, and a load feature matrix of the current data is constructed by calculating the effective value of the current data;

[0023] A comprehensive feature matrix of the running data is formed by aligning the time stamps of the frequency spectrum distribution matrix, the statistical feature matrix and the load feature matrix, and the XGBoost model determines the split gain of each dimension feature in the comprehensive feature matrix by iterative calculation to generate a feature weight coefficient;

[0024] The XGBoost model generates a device state prediction value based on the comprehensive feature matrix and the feature weight coefficient, and triggers the early warning mechanism in the S4 when the device state prediction value exceeds the parameter reference interval and meets a preset accuracy condition;

[0025] The XGBoost model sets a performance evaluation period, and statistics the prediction accuracy in the evaluation period, and performs incremental training using newly added running data to update the feature weight coefficient when the prediction accuracy is lower than a preset threshold.

[0026] Further, the time convolution network model of the S3 cloud layer performs deep time series analysis and long-term prediction on the running data, specifically including:

[0027] The vibration data, the temperature data and the current data in the running data are constructed into a time series data matrix, and training samples are extracted from the time series data matrix according to a preset sampling interval;

[0028] The time convolution network model includes a causal convolution structure and a dilated convolution structure, wherein the causal convolution structure is used to extract time series features, and the dilated convolution structure is used to expand the receptive field range; the training samples are used to construct a running state feature matrix through the causal convolution structure, and the running state feature matrix is mapped to a deep feature space using the dilated convolution structure to form a state representation vector of the running data;

[0029] The time convolution network model generates a predicted value of the operation data based on a state feature vector of the operation data, and when the predicted value continuously deviates from a parameter benchmark interval within a preset time range, triggers a warning mechanism in the S4;

[0030] The time convolution network model sets a model update triggering condition, wherein when the number of newly added training samples reaches a preset sample threshold, the time convolution network model is retrained using historical data in the time series data matrix and the newly added training samples, and the state feature vector is updated.

[0031] The application also discloses an equipment monitoring and prediction system based on industrial middle platform and edge computing, which comprises a monitoring and collecting unit, an edge computing unit, an industrial middle platform unit and a cloud processing unit.

[0032] The edge computing unit is in communication connection with the monitoring and collecting unit through a field bus, and is provided with a data preprocessing module and an edge analysis module.

[0033] The industrial middle platform unit is in data connection with the edge computing unit, and is provided with a data analysis module, a trend prediction module and a task management module.

[0034] The cloud processing unit is in data connection with the edge computing unit and the industrial middle platform unit, and is provided with a deep learning module and a warning processing module.

[0035] The application has the following advantages:

[0036] (1) The application adopts an innovative combination of a three-layer prediction architecture, realizes all-round coverage from real-time monitoring to long-term prediction through the synergistic cooperation of the lightweight threshold rule model of the edge layer, the XGBoost model of the middle layer and the time convolution network model of the cloud layer, effectively reduces the system response delay and improves the prediction accuracy.

[0037] (2) The application designs a systematic hierarchical early warning mechanism, closely combines the first-level warning, the second-level warning and the third-level warning with the three-layer architecture, forms a complete early warning response system, and realizes comprehensive early warning from emergency exception handling to long-term risk prediction.

[0038] (3) The application adopts a sliding window for dynamic monitoring, constructs a parameter anomaly correlation matrix, realizes cross-validation of multi-dimensional parameters such as vibration, temperature and current, and improves the accuracy and reliability of anomaly detection.

[0039] (4) The application selects the most suitable algorithm model at each level, the lightweight rule model of the edge layer meets the real-time requirement, the XGBoost model of the middle layer is suitable for medium-scale data analysis, and the time convolution network of the cloud layer is suitable for deep time series analysis, realizing reasonable allocation of computing resources and optimization of system performance. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 It is a device monitoring and prediction method flowchart based on industrial middle platform and edge computing.

[0041] Figure 2 It is a three-layer prediction model architecture diagram.

[0042] Figure 3 It is a device monitoring and prediction system architecture diagram based on industrial middle platform and edge computing. DETAILED DESCRIPTION

[0043] As shown in Figure 1 The application embodiment discloses a device monitoring and prediction method based on industrial middle platform and edge computing, which realizes all-round coverage from real-time monitoring to long-term prediction through a three-layer prediction architecture. The application deploys a lightweight threshold rule model in the edge layer for real-time anomaly detection, deploys an XGBoost model in the middle layer for trend prediction, and deploys a time convolution network model in the cloud layer for deep time series analysis, forming a complete prediction analysis system. Combined with a hierarchical early warning mechanism, the application can timely discover device anomalies and perform corresponding level processing, and the specific implementation steps are as follows:

[0044] S1: Data acquisition

[0045] In the embodiment, the monitoring acquisition unit comprises a sensor assembly and a data acquisition assembly. The sensor assembly comprises a vibration sensor, a temperature sensor and a current sensor for acquiring operation data including vibration data, temperature data and current data. The vibration sensor is arranged on the main bearing housing of the equipment, the temperature sensor is arranged on the heat generating area of the equipment, and the current sensor is arranged on the power supply circuit of the equipment.

[0046] The installation positions of the sensors are reasonably arranged according to the operation characteristics of the equipment to ensure that accurate operation data is obtained. During the data acquisition process, the system sets corresponding sampling time sequences according to the characteristics of different types of sensors to ensure the continuity and integrity of the data. Preferably, the sampling frequency of the vibration sensor is set to 1 kHz, the sampling frequency of the temperature sensor is set to 1 Hz, and the sampling frequency of the current sensor is set to 10 Hz. The data acquisition assembly is electrically connected with the sensor assembly, and data is acquired at the preset sampling frequency and preliminarily stored in the edge controller.

[0047] S2: Data preprocessing and distribution

[0048] In the embodiment, the acquired operation data is preprocessed and distributed to a three-layer prediction model. The data preprocessing includes data cleaning, data repair and exception handling on the operation data. The three-layer prediction model comprises a lightweight threshold rule model deployed in the edge layer, an XGBoost model deployed in the middle layer and a time convolution network model deployed in the cloud layer.

[0049] Specifically, the data cleaning process first detects outliers in the acquired vibration data, temperature data and current data, and removes data points that obviously deviate from the normal range. Preferably, the 3σ criterion is used for outlier judgment, that is, data points exceeding the range of plus or minus three times the standard deviation of the average value are marked as outliers. For missing data points, the system uses a data repair method for processing. Preferably, for short-time data missing, a linear interpolation method is used for repair, and for long-time data missing, statistical characteristic values of historical data are used for filling.

[0050] The preprocessed data is distributed to three-layer prediction models according to different time scales and computing requirements. Among them, the edge layer receives real-time data stream for rapid anomaly detection; the middle layer receives down-sampled medium time scale data for trend analysis; and the cloud layer receives long time series data for deep time series analysis. Preferably, the original sampling data is processed as follows: the data collected by the vibration sensor at a sampling frequency of 1 kHz is averaged every 1000 data points at the middle layer, converted to 1 Hz data; the data collected by the current sensor at a sampling frequency of 10 Hz is averaged every 10 data points at the middle layer, converted to 1 Hz data; the data collected by the temperature sensor at a sampling frequency of 1 Hz remains unchanged. Through this down-sampling processing, the middle layer obtains unified 1 Hz medium time scale data for trend analysis. Preferably, for the long time series data of the cloud layer, the system further down-samples the above 1 Hz data: a set of statistical values including mean, standard deviation, maximum value, and minimum value are calculated every 3600 data points (i.e. 1 hour of data), thereby obtaining long time series data in units of hours, facilitating long-term trend analysis. Through this preprocessing and distribution mechanism, the application realizes the guarantee of data quality and the rational allocation of computing resources, providing a reliable data foundation for subsequent multi-level prediction analysis.

[0051] S3: Three-layer prediction model analysis and prediction

[0052] Reference Figure 2 In this embodiment, the system uses a three-layer prediction model for equipment state analysis and prediction. Specifically, different prediction models are deployed at the edge layer, the middle layer, and the cloud layer to realize multi-level state monitoring and prediction analysis.

[0053] The lightweight threshold rule model of the edge layer performs anomaly detection based on the operating data and the parameter reference interval. The parameter reference interval includes a vibration parameter reference interval, a temperature parameter reference interval, and a current parameter reference interval. The vibration parameter reference interval is a range of mean plus or minus standard deviation calculated based on the vibration data, the temperature parameter reference interval is a range of mean plus or minus standard deviation calculated based on the temperature data, and the current parameter reference interval is a range of mean plus or minus standard deviation calculated based on the current data. Preferably, the parameter reference interval uses a ±1σ (one standard deviation) range as the normal operation interval, for example: when the vibration mean is 2.5 mm / s and the standard deviation is 0.3 mm / s, the vibration parameter reference interval is [2.2 mm / s, 2.8 mm / s]; when the temperature mean is 45℃ and the standard deviation is 2℃, the temperature parameter reference interval is [43℃, 47℃]; and when the current mean is 80A and the standard deviation is 5A, the current parameter reference interval is [75A, 85A].

[0054] Specifically, the edge layer adopts a time series-based sliding window to dynamically monitor the operation data. The sliding window moves along the time axis by a preset step size and calculates the mean and standard deviation within the sliding window. When a parameter is detected to be outside the parameter reference interval, parameter correlation analysis is performed. The parameter correlation analysis includes counting the frequency of the vibration data, the temperature data, and the current data exceeding the parameter reference interval within the sliding window to generate a parameter anomaly correlation matrix. When a parameter is outside the parameter reference interval, the system checks whether other parameters in the parameter anomaly correlation matrix are also abnormal at the same time. If an abnormality is confirmed, the early warning mechanism is triggered. Preferably, taking the operation process of an industrial pump as an example: when the sliding window detects that the bearing vibration value reaches 2.9 mm / s, which is outside the reference interval [2.2 mm / s, 2.8 mm / s], the system finds through the parameter anomaly correlation matrix that the bearing temperature also rises to 48°C, which is outside the reference interval [43°C, 47°C], and the motor current rises to 87A, which is outside the reference interval [75A, 85A]. Since three parameters are simultaneously abnormal and have a correlation (bearing failure usually exhibits a coupled effect of increased vibration, rising temperature, and rising current), the system confirms that it is a real abnormality and triggers the early warning mechanism. In contrast, when only the vibration value is detected to be temporarily abnormal while the temperature and current remain within the normal range, the system determines that it may be a temporary disturbance and does not immediately trigger the early warning mechanism.

[0055] Preferably, the time length of the sliding window is 10 minutes, and the preset step length is 1 minute, which ensures the balance between the continuity of data sampling and the efficiency of calculation. In each sliding window, the system records the maximum value, minimum value, mean value and standard deviation of each parameter, which is used to evaluate the stability of parameter change. Preferably, the parameter anomaly correlation matrix adopts a three-dimensional matrix structure, wherein the first dimension represents the parameter type (vibration, temperature, current), the second dimension represents the time window number, and the third dimension represents the abnormal state mark (0 represents normal, 1 represents abnormal). Preferably, the parameter anomaly correlation matrix adopts a three-dimensional matrix structure, with a dimension of 3xNx2, wherein: the first dimension size is 3, corresponding to vibration, temperature and current three parameters respectively; the second dimension size is N, indicating the data of the last N time windows, for example, N = 10 indicates the data of the last 10 time windows; the third dimension size is 2, which records whether the parameter value is out of the reference interval (0 represents normal, 1 represents abnormal). For example, the matrix element [0][5][1] = 1 indicates that the vibration parameter (first dimension index 0) in the 6th time window (second dimension index 5) has an abnormal mark bit (third dimension index 1) value of 1, i.e. abnormal; similarly, [1][5][1] = 1 indicates that the temperature parameter (first dimension index 1) in the 6th time window (second dimension index 5) has an abnormal mark bit (third dimension index 1) value of 1, i.e. abnormal. When the system detects that [0][5][1] and [1][5][1] are both 1, it means that the vibration and temperature parameters have both appeared abnormal in the same time window, and in this case the system can quickly confirm the correlation of parameter anomaly.

[0056] During the period when the parameter anomaly correlation matrix does not confirm the anomaly, the statistical values of the sliding window are periodically added to the historical data to recalculate the parameter reference interval. Preferably, the system performs parameter reference interval update calculation every 24 hours, using the historical data of the last 7 days as the calculation basis to ensure the timeliness and stability of the reference interval. Preferably, when performing parameter correlation analysis, the system sets the minimum abnormal duration threshold to 30 seconds, i.e. only when the parameter anomaly duration exceeds the threshold will it be recorded in the parameter anomaly correlation matrix, in order to filter out false positives caused by transient fluctuations.

[0057] At the middle layer, the XGBoost model performs feature analysis and trend prediction on the operation data. Specifically, first, a feature matrix is constructed for each of the vibration data, the temperature data, and the current data in the operation data. The frequency spectrum distribution matrix of the vibration data is obtained by Fourier transform of the vibration data, the statistical feature matrix of the temperature data is constructed by calculating the mean and standard deviation of the temperature data, and the load feature matrix of the current data is constructed by calculating the effective value of the current data. Preferably, the frequency spectrum distribution matrix of the vibration data contains the amplitude of the main frequency component, and the matrix dimension is MxF, where M is the number of time windows and F is the number of frequency components. The statistical feature matrix of the temperature data has a dimension of Mx2, containing two statistics of mean and standard deviation for each time window. The load feature matrix of the current data has a dimension of Mxl, recording the effective value in each time window. Through this feature extraction method, the system can capture the frequency characteristics of vibration, the statistical characteristics of temperature, and the load characteristics of current, respectively.

[0058] The timestamps of the frequency spectrum distribution matrix, the statistical feature matrix, and the load feature matrix are aligned to form a comprehensive feature matrix of the operation data. The XGBoost model determines the split gain of each dimension feature in the comprehensive feature matrix by iterative calculation to generate feature weight coefficients. Preferably, the time alignment of the comprehensive feature matrix uses 1 minute as the basic time unit, downsamples data with a sampling frequency higher than 1 minute, and supplements data with a sampling frequency lower than 1 minute using linear interpolation method to ensure the time consistency of the feature data.

[0059] The XGBoost model generates a device state prediction value based on the comprehensive feature matrix and the feature weight coefficients. When the device state prediction value exceeds the parameter reference interval and meets the preset accuracy condition, the warning mechanism is triggered. Preferably, taking the fan vibration parameter prediction as an example: the XGBoost model obtains a prediction value of 2.9 mm / s based on the comprehensive feature matrix analysis, which exceeds the vibration parameter reference interval [2.2 mm / s, 2.8 mm / s]; at the same time, the model's accuracy evaluation of this prediction result is 92%, which is higher than the preset accuracy threshold of 85%, so the system triggers a secondary warning. If the prediction accuracy is lower than the preset threshold, even if the prediction value exceeds the parameter reference interval, the system will not trigger a warning to avoid false positives.

[0060] Preferably, the main parameter configuration of the XGBoost model is as follows: the maximum tree depth is set to 6, the number of weak learners is set to 100, the learning rate is set to 0.1, the minimum split loss threshold is set to 1, the L1 regularization term coefficient is set to 0.1, and the L2 regularization term coefficient is set to 0.2.

[0061] The XGBoost model sets a performance evaluation period, and the prediction accuracy is calculated during the evaluation period. When the prediction accuracy is lower than the preset threshold, the incremental training is performed using the newly added operation data to update the feature weight coefficient. Preferably, the performance evaluation period is set to every 8 hours, and the preset accuracy threshold is set to 85%. During incremental training, the historical data of the last 30 days is used as the training set, and the latest 1000 samples are retained for model validation each time. During the incremental training process, the learning rate is dynamically adjusted to 0.5 times the base learning rate to avoid overfitting of the model. Preferably, to improve the reliability of the prediction, the system sets a confidence evaluation mechanism for the prediction result. Specifically, by calculating the variance of the predicted value, the prediction result is divided into high, medium and low confidence levels, corresponding to prediction results with confidence greater than 90%, 70%-90% and less than 70% respectively. When the confidence of the prediction is less than 70%, the system will mark a low confidence prompt in the warning information.

[0062] In the cloud layer, the time convolution network model performs deep time series analysis and long-term prediction on the operation data. Specifically, the vibration data, temperature data and current data in the operation data are constructed into a time series data matrix, and training samples are extracted from the time series data matrix according to a preset sampling interval. Preferably, the construction of the time series data matrix uses a sliding window method, with a window length of 168 hours (7 days) and a step length of 1 hour, ensuring the capture of periodic characteristics of equipment operation. For different types of data, time down-sampling is performed according to their sampling frequencies, and they are all converted to hour-level data. Preferably, the extraction of training samples uses a stratified sampling strategy, and the data is divided into 80% training set and 20% validation set. During the sample extraction process, samples of various operating conditions are ensured to be included in the training set to improve the generalization ability of the model.

[0063] The time convolution network model comprises a causal convolution structure and an inflation convolution structure, wherein the causal convolution structure is used to extract time sequence features, and the inflation convolution structure is used to expand the receptive field range. The training sample is subjected to a running state feature matrix construction operation through the causal convolution structure, and the running state feature matrix is mapped to a deep feature space by using the inflation convolution structure to form a state representation vector of the running data. Preferably, the specific configuration of the causal convolution structure is as follows: the first layer convolution kernel size is 3, and the output channel number is 32; the second layer convolution kernel size is 3, and the output channel number is 64; the third layer convolution kernel size is 3, and the output channel number is 128; a batch normalization layer and a ReLU activation function are connected after each layer of convolution to avoid the problem of gradient disappearance. Preferably, the inflation convolution structure adopts an incremental inflation rate setting: the first layer inflation rate is 1, and the receptive field is 3; the second layer inflation rate is 2, and the receptive field is 7; the third layer inflation rate is 4, and the receptive field is 15. Through the incremental inflation rate setting, the model can effectively capture the dependence relationship of different time scales.

[0064] The time convolution network model generates a predicted value of the running data based on the state representation vector of the running data, and triggers a warning mechanism when the predicted value continuously deviates from the parameter reference interval within a preset time range. Preferably, the generation of the predicted value adopts a uniform time interval, and the device state in the future 7 days is predicted, and a predicted point is generated every 6 hours to evaluate the long-term running risk of the device. When the predicted value continuously deviates from the parameter reference interval at a plurality of continuous time points, a three-level warning is triggered.

[0065] The time convolution network model sets a model update trigger condition, wherein when the number of newly added training samples reaches a preset sample threshold, the time convolution network model is retrained by using the historical data in the time sequence data matrix and the newly added training samples, and the state representation vector is updated. Preferably, the trigger condition of the model update is specifically as follows: the number of newly added samples reaches 20% of the original training set, the proportion of samples with a prediction error exceeding a preset threshold exceeds 15%, the prediction accuracy of 7 consecutive days is lower than 80%, and any one of the above conditions is met to trigger the model update. Preferably, the following strategies are adopted in the model update process: the historical data in the last 90 days are retained as a basic training set; a transfer learning method is used to fine-tune only the last two layers of network parameters; an early stopping strategy is adopted in the training process, and the training is stopped when the loss of the verification set does not improve for 5 consecutive epochs; and the updated model needs to pass backtest verification, and only when the accuracy is higher than that of the original model, the model is deployed online.

[0066] S4: hierarchical warning mechanism establishment

[0067] In this embodiment, a hierarchical warning mechanism is established based on the prediction results of the three-layer prediction model, and a full-range warning from real-time abnormalities to long-term risks is realized.

[0068] Specifically, the first-level early warning is triggered when the edge layer detects abnormality of the parameter reference interval. The first-level early warning is an emergency abnormality warning for real-time running state of the equipment. When the lightweight threshold rule model of the edge layer detects that the vibration data, temperature data or current data exceeds the respective parameter reference interval, and the abnormality is confirmed through the parameter abnormality correlation matrix, the system immediately triggers the first-level early warning.

[0069] The second-level early warning is triggered when the middle platform layer detects abnormal trend of the running data. The second-level early warning is a predictive warning for running trend of the equipment. The second-level early warning is triggered when the equipment state value predicted by the XGBoost model of the middle platform layer exceeds the parameter reference interval and meets the preset accuracy condition.

[0070] The third-level early warning is triggered when the cloud layer predicts running risk. The third-level early warning is a predictive warning for long-term running risk of the equipment. The third-level early warning is triggered when the equipment state predicted by the time convolution network model of the cloud layer continuously deviates from the parameter reference interval within a preset time range.

[0071] S5: Early warning response processing

[0072] In the embodiment, the system responds to the early warning result of the hierarchical early warning mechanism. The response processing is specifically implemented through the following technical means:

[0073] When the first-level early warning is triggered, the edge analysis module of the edge computing unit performs the following response processing: sending a device stop running instruction to the device control system through an industrial communication protocol, and pushing an emergency alarm information to the operation and maintenance personnel. Preferably, the edge analysis module writes a stop instruction to the programmable logic controller (PLC) of the device by using the Modbus TCP protocol, and simultaneously pushes the alarm information to the mobile terminal of the operation and maintenance personnel in real time through WebSocket. The alarm information is encapsulated in JSON format, and includes fields such as device ID, abnormal parameter, current value of the parameter, trigger time, etc.

[0074] When the second-level early warning is triggered, the task management module of the industrial middle platform unit performs the following response processing: generating a device inspection task and adjusting the device maintenance plan. Preferably, the task management module creates an inspection task record in the maintenance management database, which includes fields such as work order number, device number, inspection item, planned time, etc., and updates the related records of the maintenance plan.

[0075] When the third-level early warning is triggered, the early warning processing module of the cloud processing unit performs the following response processing: generating a device operation and maintenance analysis report based on the predicted data. Preferably, the early warning processing module generates a device performance trend analysis and a fault risk assessment based on the predicted data, and sends the analysis report to the designated management personnel and technical personnel through email, so as to realize timely delivery of the risk early warning information.

[0076] As Figure 3 shown, the application also discloses an industrial middle platform and edge computing-based equipment monitoring and prediction system, which comprises a monitoring and collecting unit, an edge computing unit, an industrial middle platform unit and a cloud processing unit, and each unit is interconnected through the industrial Ethernet. The system realizes all the functions of the aforementioned equipment monitoring and prediction method, and the specific implementation of each unit is described in detail below.

[0077] In this embodiment, the monitoring and collecting unit comprises the sensor assembly and the data collecting assembly. The sensor assembly comprises the vibration sensor, the temperature sensor and the current sensor, as described above, the vibration sensor is arranged on the main bearing shell of the equipment, the temperature sensor is arranged on the heat generating area of the equipment, and the current sensor is arranged on the power supply line of the equipment. The data collecting assembly is electrically connected with the sensor assembly and is used to collect the equipment operation data according to the preset sampling frequency. Preferably, the data collecting assembly adopts a modular design and comprises the multi-channel analog signal collecting card and the data preprocessing unit, and the data preprocessing unit is equipped with an industrial-grade storage card for local data caching.

[0078] The edge computing unit is in communication connection with the monitoring and collecting unit through the field bus and is provided with the data preprocessing module and the edge analysis module. The data preprocessing module is used to clean and repair the equipment operation data, and the edge analysis module is internally provided with the lightweight threshold rule model and is used to perform real-time anomaly detection and send the equipment stop operation instruction to the equipment control system and push the emergency alarm information to the operation and maintenance personnel when the first-level early warning is triggered. Preferably, the edge computing unit adopts an industrial-grade edge computing gateway, and the edge analysis module realizes real-time pushing of early warning information through the publish-subscribe mode.

[0079] The industrial middle platform unit is in data connection with the edge computing unit and is provided with the data analysis module, the trend prediction module and the task management module. The data analysis module is used to construct the feature matrix and perform feature analysis at the middle platform layer, the trend prediction module is internally provided with the XGBoost prediction model and is used to perform trend prediction analysis at the middle platform layer, and the task management module is used to generate the equipment inspection task and adjust the maintenance plan when the second-level early warning is triggered. Preferably, the industrial middle platform unit adopts an enterprise-level server configuration and realizes acceleration of feature extraction and model calculation through the parallel computing framework.

[0080] The cloud processing unit is in data connection with the edge computing unit and the industrial middle platform unit, and is provided with the deep learning module and the early warning processing module. The deep learning module is internally provided with the time convolution network model, which is used for long-term prediction, and the early warning processing module is used for generating the equipment operation and maintenance analysis report based on the prediction data when the three-level early warning is triggered. Preferably, the cloud processing unit adopts the containerized deployment scheme and is equipped with a deep learning acceleration unit for model training and inference. The message queue mechanism is used for data transmission between the units of the system, and the encryption protocol is used to ensure the safety of data transmission. The overall system is configured and managed through the Web management interface, supports centralized collection and analysis of logs, and realizes real-time monitoring of system performance.

[0081] Through the implementation of the above technical solutions, the system realizes all-around monitoring and predictive maintenance of industrial equipment. The units of the system cooperate with each other, the edge layer ensures real-time response, the middle platform layer realizes trend analysis, and the cloud layer completes deep prediction, forming a complete monitoring and prediction system, effectively improving the equipment operation and maintenance efficiency and reducing the risk of equipment failure. The hierarchical architecture design of the system fully utilizes the advantages of edge computing, industrial middle platform and cloud computing, realizes the reasonable allocation of computing resources and the optimization of data processing efficiency.

Claims

1. An industrial middle platform and edge computing-based device monitoring prediction method, characterized in that, The method comprises the following steps: S1. Collecting operation data including vibration data, temperature data and current data, setting sampling timing according to device operation characteristics, collecting data at a preset sampling frequency and preliminarily storing in an edge controller; S2. Distributing the operation data to a three-layer prediction model after preprocessing, the preprocessing including data cleaning, data repair and exception handling on the operation data, the three-layer prediction model including a lightweight threshold rule model deployed in an edge layer, an XGBoost model deployed in a middle layer and a time convolution network model deployed in a cloud layer; wherein the middle layer is an industrial middle layer; S3. Using the three-layer prediction model to analyze and predict the device state, wherein the lightweight threshold rule model of the edge layer performs abnormal detection based on the operation data and a parameter reference interval, the XGBoost model of the middle layer performs feature analysis and trend prediction on the operation data, and performs abnormal judgment based on the parameter reference interval, and the time convolution network model of the cloud layer performs deep time series analysis and long-term prediction on the operation data, and performs risk assessment based on the parameter reference interval; Wherein, the parameter reference interval includes a vibration parameter reference interval, a temperature parameter reference interval and a current parameter reference interval, the vibration parameter reference interval is a range of mean plus or minus standard deviation calculated based on the vibration data, the temperature parameter reference interval is a range of mean plus or minus standard deviation calculated based on the temperature data, and the current parameter reference interval is a range of mean plus or minus standard deviation calculated based on the current data; S4. Establishing a hierarchical early warning mechanism based on the prediction results of the three-layer prediction model, triggering a first-level early warning when the edge layer detects an abnormal parameter reference interval, triggering a second-level early warning when the middle layer detects an abnormal trend of the operation data, and triggering a third-level early warning when the cloud layer predicts a running risk; S5. Responding to the early warning results of the hierarchical early warning mechanism, sending a device stop running instruction to a device control system and pushing an emergency alarm information to an operation and maintenance personnel when the first-level early warning is triggered, generating a device inspection task and adjusting a device maintenance plan when the second-level early warning is triggered, and generating a device operation and maintenance analysis report based on prediction data when the third-level early warning is triggered; The lightweight threshold rule model of the edge layer in S3 performs parameter reference interval calculation and abnormal detection based on the operation data, specifically including: Using a time series-based sliding window to dynamically monitor the operation data, wherein the sliding window moves on the time axis according to a preset step size and calculates the mean and standard deviation within the sliding window, and when it is detected that a parameter exceeds the parameter reference interval, parameter correlation analysis is performed; The parameter correlation analysis includes counting the frequency of the vibration data, the temperature data and the current data exceeding the parameter reference interval in the sliding window, generating a parameter anomaly correlation matrix, when a parameter exceeds the parameter reference interval, checking whether other parameters in the parameter anomaly correlation matrix are abnormal at the same time, and triggering the warning mechanism in S4 if the anomaly is confirmed; During the period when the parameter anomaly correlation matrix does not confirm the anomaly, the statistical value of the sliding window is added to the historical data to recalculate the parameter reference interval; The XGBoost model of the S3 middle layer performs feature analysis and trend prediction on the running data, specifically including: A feature matrix is constructed for the vibration data, the temperature data and the current data in the running data, respectively, wherein the frequency spectrum distribution matrix of the vibration data is obtained by performing Fourier transform on the vibration data, the statistical feature matrix of the temperature data is constructed by calculating the mean and standard deviation of the temperature data, and the load feature matrix of the current data is constructed by calculating the effective value of the current data; The comprehensive feature matrix of the running data is formed by aligning the time stamps of the frequency spectrum distribution matrix, the statistical feature matrix and the load feature matrix, and the XGBoost model determines the split gain of each dimension feature in the comprehensive feature matrix by iterative calculation to generate feature weight coefficients; The XGBoost model generates a device state prediction value based on the comprehensive feature matrix and the feature weight coefficients, and triggers the warning mechanism in S4 when the device state prediction value exceeds the parameter reference interval and meets the preset accuracy condition; The XGBoost model sets a performance evaluation period, and the prediction accuracy is counted during the evaluation period, and when the prediction accuracy is lower than the preset threshold, the incremental training is performed using the newly added running data to update the feature weight coefficients.

2. The device monitoring and prediction method based on industrial middle platform and edge computing according to claim 1, characterized in that, The time convolution network model of the S3 cloud layer performs deep time series analysis and long-term prediction on the running data, specifically including: The vibration data, the temperature data and the current data in the running data are constructed into a time series data matrix, and training samples are extracted from the time series data matrix according to a preset sampling interval; The time convolution network model includes a causal convolution structure and a dilated convolution structure, wherein the causal convolution structure is used to extract time series features, and the dilated convolution structure is used to expand the receptive field range; the training samples are used to construct a running state feature matrix through the causal convolution structure, and the running state feature matrix is mapped to a deep feature space using the dilated convolution structure to form a state representation vector of the running data; The time convolution network model generates a prediction value of the running data based on the state representation vector of the running data, and triggers the warning mechanism in S4 when the prediction value continuously deviates from the parameter reference interval within a preset time range. The time convolution network model sets a model update trigger condition, wherein when the number of newly added training samples reaches a preset sample threshold, the time convolution network model is retrained using historical data in the time series data matrix and the newly added training samples, and the state representation vector is updated.

3. An industrial middle platform and edge computing based device monitoring and prediction system, characterized in that, The device monitoring prediction system is used to implement the prediction method of claim 1, and the device monitoring prediction system comprises a monitoring acquisition unit, an edge computing unit, an industrial middle platform unit and a cloud processing unit. The monitoring acquisition unit comprises a sensor assembly and a data acquisition assembly, the sensor assembly comprises a vibration sensor, a temperature sensor and a current sensor, the vibration sensor is arranged on a main bearing shell of the device, the temperature sensor is arranged on a heat generating area of the device, and the current sensor is arranged on a power supply line of the device, the data acquisition assembly is electrically connected with the sensor assembly, and is used to collect device operation data according to a preset sampling frequency. The edge computing unit is in communication connection with the monitoring acquisition unit through a field bus, and is provided with a data preprocessing module and an edge analysis module, the data preprocessing module is used to clean and repair the device operation data, the edge analysis module is internally provided with a lightweight threshold rule model, is used to perform real-time anomaly detection, and sends a device stop operation instruction to a device control system and pushes emergency alarm information to an operation and maintenance personnel when a first level early warning is triggered; The industrial middle platform unit is in data connection with the edge computing unit, and is provided with a data analysis module, a trend prediction module and a task management module, the data analysis module is used to construct a feature matrix and perform feature analysis at the middle platform layer, the trend prediction module is internally provided with an XGBoost prediction model, is used to perform trend prediction analysis at the middle platform layer, and the task management module is used to generate a device inspection task and adjust a maintenance plan when a second level early warning is triggered; The cloud processing unit is in data connection with the edge computing unit and the industrial middle platform unit, and is provided with a deep learning module and an early warning processing module, the deep learning module is internally provided with a time convolution network model, is used to perform long-term prediction, and the early warning processing module is used to generate a device operation and maintenance analysis report based on prediction data when a third level early warning is triggered.

Images