Dynamic data integration and predictive maintenance method based on cloud-edge-end architecture

Abstract

The invention belongs to the technical field of equipment maintenance, and relates to a dynamic data integration and predictive maintenance method based on a cloud-edge-end architecture, and the method comprises the steps: collecting the operation data of equipment at an end layer, carrying out the denoising and standardized preprocessing, carrying out the preliminary cleaning and abnormal value detection of the preprocessed data at an edge layer, and carrying out the abnormal value detection; transmitting the cleaned data to the cloud; at the cloud end, an equipment health database, maintenance activity data and a maintenance resource database are integrated through a data fusion technology, and comprehensive data support is provided for subsequent analysis; and finally, performing prediction analysis on the integrated data by using the trained deep learning model, thereby accurately predicting the occurrence probability and the fault type of the equipment fault, and formulating a corresponding maintenance strategy. According to the method, the accuracy of equipment fault prediction is improved, the timeliness and predictability of maintenance activities are realized, the influence of faults on production is greatly reduced, the production efficiency is improved, and the service life of equipment is prolonged.

Description

Technical Field

[0001] This application belongs to the technical field of equipment maintenance. More specifically, it relates to a dynamic data integration and predictive maintenance method based on a cloud-edge-end architecture. Background Art

[0002] With the rapid development of intelligent manufacturing, the number and types of equipment in the enterprise production process are increasing day by day, and the operation data of the equipment also shows an explosive growth. The data generated during the operation of these devices contains rich information, which is of great significance for the maintenance and optimization of the devices. However, how to efficiently process and analyze these large-scale device data to achieve dynamic data integration and predictive maintenance of the devices has become a major challenge faced by the current manufacturing industry.

[0003] Currently, for equipment maintenance problems, the existing technologies mainly include regular maintenance and failure-based maintenance. Regular maintenance is to repair and replace parts according to a fixed cycle based on the running time and wear degree of the equipment. The disadvantage of this method is that it cannot accurately predict the specific time of equipment failure, which may lead to over-maintenance or untimely maintenance. Failure-based maintenance is to repair the equipment when a failure occurs. Although this method saves maintenance costs, equipment failures may have a serious impact on production. Summary of the Invention

[0004] The present invention provides a dynamic data integration and predictive maintenance method based on a cloud-edge-end architecture, aiming to solve the technical problem that equipment failures caused by current regular maintenance and failure-based maintenance methods will have a serious impact on production.

[0005] A dynamic data integration and predictive maintenance method based on a cloud-edge-end architecture, characterized by including the following steps:

[0006] Step 1: The edge layer collects the operation data of the device and preprocesses the collected data, including denoising and normalization processing;

[0007] Step 2: The edge layer receives the data preprocessed by the edge layer, and performs preliminary data cleaning and outlier detection on the received data, and sends the data after data cleaning and outlier detection to the cloud;

[0008] Step 3: The cloud stores the received device operation data in the device health database, and integrates the data in the device health database, maintenance activity data, and maintenance resource database through data fusion technology;

[0009] Step 4: Use the trained deep learning model and the integrated data for prediction to obtain the occurrence probability of equipment failure, failure type, and maintenance strategy.

[0010] The dynamic data integration and predictive maintenance method based on the cloud-edge-end architecture proposed by the present invention effectively solves the technical problem that equipment failures caused by regular maintenance and failure-based maintenance methods seriously affect production through the following steps: First, at the edge layer, the operation data of the equipment is collected and preprocessed by denoising and standardization, ensuring the quality and consistency of the data; Second, the edge layer performs preliminary cleaning and outlier detection on the preprocessed data, further improving the reliability of the data and transmitting the cleaned data to the cloud; At the cloud, the equipment health database, maintenance activity data, and maintenance resource database are integrated through data fusion technology, providing comprehensive data support for subsequent analysis; Finally, the integrated data is predicted and analyzed using the trained deep learning model to accurately predict the occurrence probability and type of equipment failures and formulate corresponding maintenance strategies. This method not only improves the accuracy of equipment failure prediction but also realizes the timeliness and predictability of maintenance activities, greatly reducing the impact of failures on production and improving production efficiency and equipment service life.

[0011] Preferably, the data integration in step 3 includes data integration for the occurrence probability of equipment failures, data integration for equipment type prediction, and data integration for maintenance strategy prediction.

[0012] Preferably, the steps of the data integration for the occurrence probability of equipment failures are as follows:

[0013] Extract temperature features, humidity features, load features, health change trend features, and operation duration features related to equipment failures based on the equipment health database;

[0014] Among them, the temperature feature extraction is as follows:

[0015]

[0016] In the formula: T avg (t) represents the average ambient temperature at time t; N temp represents the number of temperature data points recorded within time t; T i (t) represents the value of the i-th temperature data point;

[0017] The humidity feature extraction is as follows:

[0018]

[0019] In the formula: H avg (t) represents the average ambient humidity at time t; N humidity represents the number of humidity data points recorded within event t; H i (t) represents the value of the i-th humidity data point;

[0020] The load feature extraction is as follows:

[0021]

[0022] Where: L avg (t) represents the average load at time t; N load represents the number of load data points recorded within time t; L i (t) represents the value of the i-th load data point;

[0023] The health change trend features are extracted as follows:

[0024] HI(t) = w 1 ·HI 1 (T) + w 2 ·HI 2 (t) + … + w n ·H n (t);

[0025] Where: HI 1 (T), HI 2 (t), …, HI n (t) represent the health indicators after processing the data of each sensor of the device at time t; w 1 , w 2 , …, w n represent the weights of the data of each sensor; HI(t) represents the health index of the device at time t;

[0026] ΔHI(t) = HI(t) - HI(t - 1);

[0027] Where: ΔHI(t) represents the change in the health index of the device;

[0028] The operating duration features related to device failures are extracted as follows:

[0029]

[0030] Where: T run (t) represents the cumulative operating duration of the device from the start of operation to the current time point t; Δt i represents the time within the i-th operating cycle of the device;

[0031] Based on the maintenance activity database, failure history features, maintenance duration features, and device maintenance strategy features are extracted;

[0032] The failure history features are extracted as follows:

[0033]

[0034] Where: F freq(t) represents the failure occurrence frequency of the device within time t; N failures (t) represents the number of failures that occurred to the device within time t; T total (t) represents the total operating time of the device within time t;

[0035] ΔT fail (t) = t - T last_failure ;

[0036] In the formula: ΔT fail (t) represents the time interval from the last failure occurrence of the device to the current moment; T last_failure represents the time of the last failure occurrence of the device; t represents the current time;

[0037] The extraction of the maintenance duration feature is as follows:

[0038]

[0039] In the formula: D repair (t) represents the average maintenance time of the device within time t; N repair (t) represents the number of maintenance times of the device within time t; T repair,i represents the time taken for the i-th maintenance;

[0040] The extraction of the device maintenance strategy feature is as follows:

[0041]

[0042] In the formula: M freq (t) represents the maintenance frequency within time t; N maintenaces (t) represents the number of maintenance times within time t; T total (t) represents the total operating time of the device within time t;

[0043] Based on the above-extracted features, feature fusion is performed to form a comprehensive feature vector, including temperature features, humidity features, load features, health change trend features, operating duration features related to device failures, fault history features, maintenance duration features, and device maintenance strategy features.

[0044] Preferably, the steps for data integration for device type prediction are as follows:

[0045] Extract health change trend features, operating duration features related to device failures, temperature features, humidity features, and device load based on the device health database;

[0046] Extract fault type history features, maintenance type features, maintenance duration features, fault mode change trend features, and fault mode evolution trend features from the maintenance activity database;

[0047] Fuse the extracted features to form a comprehensive feature vector, including health change trend features, operation duration features related to equipment failures, temperature features, humidity features, equipment load, historical features of failure types, maintenance type features, maintenance duration features, failure mode change trend features, and failure mode evolution trend features.

[0048] Preferably, the steps for data integration for maintenance strategy prediction are as follows:

[0049] Extract the maintenance duration, maintenance type, maintenance cost, maintenance personnel load, spare part availability, equipment health status, and maintenance strategy type for each maintenance based on maintenance activity data and the maintenance resource database;

[0050] Perform feature fusion based on the above - extracted features to form a comprehensive feature vector, including maintenance duration, maintenance type, maintenance cost, maintenance personnel load, spare part availability, equipment health status, and maintenance strategy type.

[0051] Preferably, the specific structure of the deep - learning model is as follows:

[0052] Input layer: It includes three input units. The first input unit is used to input the input features for failure occurrence probability prediction; the second input unit is used to input the input features for failure type prediction; the third input unit is used to input the input features for maintenance strategy prediction;

[0053] Feature extraction layer: It includes a failure occurrence probability prediction feature extraction layer, an equipment type prediction feature extraction layer, and a maintenance strategy prediction feature extraction layer;

[0054] The failure occurrence probability feature extraction layer converts the input features into a high - dimensional feature representation through a fully - connected layer;

[0055] The failure type prediction feature extraction layer generates hidden - layer features related to the equipment type task based on the input features;

[0056] The maintenance strategy prediction feature extraction layer extracts features from the input data to obtain a preliminary feature representation of the maintenance strategy;

[0057] Shared feature layer: Through a shared feature fusion layer, fuse the output results of the failure occurrence probability feature extraction layer and the failure type prediction feature extraction layer;

[0058] Failure occurrence probability prediction: Based on the output of the shared feature layer, process it through a fully - connected layer to output the probability of failure occurrence;

[0059] Failure type prediction: Based on the output of the shared feature layer, process it through a fully - connected layer to output the predicted type of failure;

[0060] Fusion layer: Fuse the prediction results of the fault type and the probability of fault occurrence with the output results of the maintenance strategy prediction feature extraction layer to form a complete feature vector;

[0061] Maintenance strategy prediction layer: Input the output of the fusion layer into the maintenance strategy prediction layer, and perform feature extraction through a deep neural network composed of fully connected layers;

[0062] Output layer: Output the type of maintenance strategy and the maintenance time based on the output results of the maintenance strategy prediction layer.

[0063] Preferably, the loss function of the deep learning model is as follows:

[0064] L = λ 1 L prob + λ 2 L type + λ 3 L startegy ;

[0065] In the formula: L prob represents the prediction loss of the probability of fault occurrence, and the mean square error is adopted; L type is the prediction loss of the fault type and is the cross-entropy loss; L strategy represents the prediction loss of the maintenance strategy and is the cross-entropy loss; λ 1 , λ 2 , λ 3 all represent the weighting coefficients.

[0066] The beneficial effects of the present invention include:

[0067] The dynamic data integration and predictive maintenance method based on the cloud-edge-end architecture proposed by the present invention effectively solves the technical problem that the equipment failure caused by the regular maintenance and the fault-based maintenance method seriously affects the production through the following steps: First, at the edge layer, the operation data of the equipment is collected and preprocessed by denoising and standardization to ensure the quality and consistency of the data; Second, the edge layer performs preliminary cleaning and outlier detection on the preprocessed data to further improve the reliability of the data and transmits the cleaned data to the cloud; At the cloud, the equipment health database, maintenance activity data, and maintenance resource database are integrated through data fusion technology to provide comprehensive data support for subsequent analysis; Finally, the integrated data is predicted and analyzed by using the trained deep learning model, so as to accurately predict the probability of equipment failure, the type of fault, and formulate corresponding maintenance strategies. This method not only improves the accuracy of equipment failure prediction, but also realizes the timeliness and predictability of maintenance activities, greatly reduces the impact of faults on production, and improves production efficiency and equipment service life. Description of the Drawings

[0068] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the accompanying drawings required for use in the embodiments or the description of the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0069] Figure 1 It is the overall step block diagram provided by the embodiment of the present invention. Specific embodiments

[0070] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application more clearly understood, the following further details the present application in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0071] See Figure 1 As shown, a further description is made of the optimal embodiment of the present invention;

[0072] A dynamic data integration and predictive maintenance method based on a cloud-edge-end architecture, characterized by including the following steps:

[0073] Step 1: The edge layer collects the operation data of the device and preprocesses the collected data, including denoising and normalization processing; among them, Z-score normalization is used for normalization; low-pass filters are used for denoising: for high-frequency noise, low-pass filters (such as moving average filters, Butterworth filters, etc.) are used to remove the noise with a frequency higher than the working signal frequency.

[0074] Step 2: The edge layer receives the data preprocessed by the edge layer, and performs preliminary data cleaning and outlier detection on the received data, and sends the data after data cleaning and outlier detection to the cloud;

[0075] Missing data processing: The edge layer needs to detect and process the missing data from the edge layer. Common processing methods include:

[0076] Interpolation method: For time series data, methods such as linear interpolation and spline interpolation can be used to fill in the missing values.

[0077] Mean / median filling: For some simple sensor data, the missing values can be filled with the mean or median of the historical data.

[0078] Duplicate data detection: The edge layer needs to check whether there are duplicate data records, remove the duplicates, and ensure the uniqueness of the data.

[0079] Redundant data removal: For some unnecessary redundant data, compression can be performed to reduce the burden of data transmission.

[0080] Outlier detection based on statistical methods: Use statistical methods (such as the mean-standard deviation method, Z-score detection, etc.) to detect outliers in the data. If some data points are far from the average value (for example, more than 3 standard deviations), they can be considered outliers.

[0081] Step 3: The cloud stores the received device operation data in the device health database and integrates the data in the device health database, maintenance activity data, and maintenance resource database through data fusion technology;

[0082] The data integration in Step 3 includes data integration for the probability of device failure, data integration for device type prediction, and data integration for maintenance strategy prediction.

[0083] The steps for the data integration for the probability of device failure are as follows:

[0084] Extract temperature features, humidity features, load features, health change trend features, and operation duration features related to device failures based on the device health database;

[0085] Among them, the temperature feature extraction is as follows:

[0086]

[0087] In the formula: T avg (t) represents the average ambient temperature at time t; N temp represents the number of temperature data points recorded within time t; T i (t) represents the value of the i-th temperature data point;

[0088] The humidity feature extraction is as follows:

[0089]

[0090] In the formula: H avg (t) represents the average ambient humidity at time t; N humidity represents the number of humidity data points recorded within event t; H i (t) represents the value of the i-th humidity data point;

[0091] The load feature extraction is as follows:

[0092]

[0093] In the formula: L avg (t) represents the average load at time t; N loadRepresents the number of load data points recorded within time t; L i (t) represents the value of the i-th load data point;

[0094] The health change trend feature is extracted as follows:

[0095] HI(t) = w 1 ·HI 1 (T) + w 2 ·HI 2 (t) + … + w n ·HI n (t);

[0096] In the formula: HI 1 (T), HI 2 (t), …, HI n (t) represents the health index of the device after processing the data of each sensor at time t; w 1 , w 2 , …, w n represents the weight of each sensor data; HI(t) represents the health index of the device at time t;

[0097] ΔHI(t) = HI(t) - HI(t - 1);

[0098] In the formula: ΔHI(t) represents the change in the health index of the device;

[0099] The operation duration feature related to equipment failure is extracted as follows:

[0100]

[0101] In the formula: T run (t) represents the cumulative operation duration of the device from the start of operation to the current time point t; Δt i represents the time within the i-th operation cycle of the device;

[0102] Extract the fault history feature, maintenance duration feature, and equipment maintenance strategy feature based on the said maintenance activity database;

[0103] The fault history feature is extracted as follows:

[0104]

[0105] In the formula: F freq (t) represents the fault occurrence frequency of the device within time t; N failures (t) represents the number of faults that occurred to the device within time t; T total (t) represents the total operation time of the device within time t;

[0106] ΔT failΔ(t) = t - T last_failure ;

[0107] where: ΔT fail Δ(t) represents the time interval from the last equipment failure to the current moment; T last_failure represents the time of the last equipment failure; t represents the current time;

[0108] The extraction of the maintenance duration feature is as follows:

[0109]

[0110] where: D repair Δ(t) represents the average maintenance time of the equipment within time t; N reqair N(t) represents the number of maintenance times of the equipment within time t; T repair,i represents the time taken for the i-th maintenance;

[0111] The extraction of the equipment maintenance strategy feature is as follows:

[0112]

[0113] where: M freq M(t) represents the maintenance frequency within time t; N maintenaces N(t) represents the number of maintenance times within time t; T total T(t) represents the total running time of the equipment within time t;

[0114] Based on the above-extracted features, feature fusion is performed to form a comprehensive feature vector, including temperature features, humidity features, load features, health change trend features, running duration features related to equipment failures, fault history features, maintenance duration features, and equipment maintenance strategy features.

[0115] The steps for data integration for equipment type prediction are as follows:

[0116] Extract health change trend features, running duration features related to equipment failures, temperature features, humidity features, and equipment load from the equipment health database; among them, the extraction of health change trend features, running duration features related to equipment failures, and equipment load features is the same as the features extracted for equipment failure probability prediction. Of course, since feature fusion will be performed in the deep learning model, we can not extract the above three features in this step, and only need to extract temperature features and humidity features. The extraction of temperature features and humidity features is to extract the temperature value and humidity value at each time point;

[0117] Extract maintenance type features, maintenance duration features, and fault mode change trend features from the maintenance activity database;

[0118] Feature extraction of maintenance type:

[0119]

[0120] Where: R type_freq (t) represents the maintenance frequency of a certain type of fault of the device within time t; N repairs_type (t) represents the number of maintenance times of a certain type of fault that occurs to the device within time t; T total (t) represents the total operating time of the device;

[0121] Among them, the maintenance duration feature is the same as the feature extracted for the device fault probability prediction; of course, since feature fusion will be performed in the deep learning model, we can not extract it in this step;

[0122] The extraction of the fault mode change trend feature is as follows:

[0123] The fault mode of the device will change with time and operating status. By analyzing the change of the fault mode of the device in different time periods, that is, by mapping the fault mode to a numerical value, for example, severe is 5 and minor is 2; based on this, the change trend of the fault mode is determined:

[0124] ΔF mode (t) = F mode (t) - F mode (t - 1);

[0125] Where: ΔF mode (t) represents the change amount of the device fault mode; F mode (t) represents the fault mode of the device at time t; F mode (t - 1) represents the fault mode of the device at time t - 1;

[0126] Fuse the extracted features to form a comprehensive feature vector, including the health change trend feature, the operating duration feature related to device faults, the temperature feature, the humidity feature, the device load, the maintenance type feature, the maintenance duration feature, and the fault mode change trend feature.

[0127] The steps of data integration for maintenance strategy prediction are as follows:

[0128] Extract the maintenance duration, maintenance type, maintenance personnel load, spare part availability, device health status, and maintenance strategy type of each maintenance based on the maintenance activity data and the maintenance resource database;

[0129] Maintenance duration Repair_Duration(t):

[0130] RePair_Duration(t) = T end -Tstart ;

[0131] Where: T end The end time of maintenance; T start represents the start time of maintenance;

[0132] Maintenance type: According to historical maintenance records, the occurrence frequencies of different maintenance types are counted, such as preventive maintenance (PM), corrective maintenance (CM), emergency maintenance (EM), etc.:

[0133]

[0134] Where: Repair_Type_Frequency(t) represents the occurrence frequency of a certain maintenance type; N Repair_Type (t) represents the number of a certain type of maintenance performed on the equipment within time t; N Total represents the total number of maintenance of the equipment;

[0135] Maintenance personnel load: The current workload of each maintenance personnel is an important factor affecting the maintenance strategy. By calculating the number of tasks to be repaired by the current maintenance personnel, their ability to handle current equipment failures can be evaluated.

[0136]

[0137] Where: N pending_tasks (t) represents the number of tasks to be processed by the maintenance personnel at the current time t; N total_tasks represents the total number of tasks of the maintenance personnel;

[0138] Spare part availability: The availability of maintenance spare parts affects the selection of maintenance strategies. Determine whether the currently required spare parts are sufficient through inventory data.

[0139]

[0140] Where: S available (t) represents the number of available spare parts in the inventory at the current time t; S required (t) represents the number of spare parts required for the current maintenance task;

[0141] Equipment health status: The health status of the equipment (such as health index, failure type, etc.) is a key factor in predictive maintenance strategies. The health status is closely related to the maintenance strategy. Usually, equipment with a low health index requires more frequent preventive maintenance or corrective maintenance.

[0142]

[0143] Where: H current (t) represents the health index of the equipment at time t; H maxRepresents the maximum health index of the device;

[0144] Maintenance strategy type: Based on each maintenance, extract the strategy type of each maintenance, such as preventive, corrective, or emergency repair; and encode the type, for example, preventive is 0, corrective is 1, and emergency is 2.

[0145] Perform feature fusion based on the above-extracted features to form a comprehensive feature vector, including repair duration, repair type, repair personnel load, spare part availability, device health status, and maintenance strategy type.

[0146] Step 4: Use the trained deep learning model and the integrated data for prediction to obtain the occurrence probability of device failures, failure types, and maintenance strategies.

[0147] The specific structure of the deep learning model is as follows:

[0148] Input layer: Includes three input units. The first input unit is used to input the input features for predicting the occurrence probability of failures; the second input unit is used to input the input features for predicting the failure type; the third input unit is used to input the input features for predicting the maintenance strategy;

[0149] Feature extraction layer: Includes a failure occurrence probability prediction feature extraction layer, a device type prediction feature extraction layer, and a maintenance strategy prediction feature extraction layer;

[0150] The failure occurrence probability feature extraction layer converts the input features into a high-dimensional feature representation through a fully connected layer;

[0151] Layer structure: Convert the input features into a high-dimensional representation through a fully connected layer (Dense Layer); after one or more fully connected layers, convert it into a high-dimensional feature representation, using the ReLU activation function.

[0152] The failure type prediction feature extraction layer generates hidden layer features related to the device type task based on the input features;

[0153] Layer structure: Generate hidden layer features related to the device type based on the input features, process them using a fully connected layer, and use the ReLU activation function;

[0154] The maintenance strategy prediction feature extraction layer extracts features from the input data to obtain a preliminary feature representation of the maintenance strategy;

[0155] Layer structure: The input data is processed through multiple fully connected layers to obtain a preliminary feature representation:

[0156] Shared Feature Layer: The output results of the fault occurrence probability feature extraction layer and the fault type prediction feature extraction layer are fused through a shared feature fusion layer;

[0157] The output results of the fault occurrence probability feature extraction layer and the fault type prediction feature extraction layer are fused through a shared feature fusion layer. This helps to share useful information between different tasks.

[0158] Layer Design: The outputs of the fault occurrence probability feature extraction layer and the fault type prediction feature extraction layer are concatenated; and further processed through multiple fully connected layers.

[0159] Fault Occurrence Probability Prediction: Based on the output of the shared feature layer, the probability of fault occurrence is output through a fully connected layer; the Sigmoid activation function is used to limit the output value to the interval [0,1].

[0160] Fault Type Prediction: Based on the output of the shared feature layer, it is processed through a fully connected layer to output the predicted type of the fault; the Softmax activation function is used for multi-class classification to output the probability of each fault type.

[0161] Fusion Layer: The prediction results of the fault type and the fault occurrence probability are fused with the output result of the maintenance strategy prediction feature extraction layer to form a complete feature vector; the previous output is concatenated with the features of the maintenance strategy;

[0162] Maintenance Strategy Prediction Layer: The output of the fusion layer is input into the maintenance strategy prediction layer, and feature extraction is performed through a deep neural network composed of fully connected layers;

[0163] Output Layer: Based on the output result of the maintenance strategy prediction layer, the type of the maintenance strategy and the maintenance time are output.

[0164] The loss function of the deep learning model is as follows:

[0165] L = λ 1 L prob + λ 2 L type + λ 3 L strategy ;

[0166] In the formula: L prob represents the fault occurrence probability prediction loss, and the mean square error is adopted; L type is the fault type prediction loss, which is the cross-entropy loss; L strategy represents the maintenance strategy prediction loss, which is the cross-entropy loss; λ 1 , λ 2 , λ 3 all represent weighting coefficients.

[0167] The dynamic data integration and predictive maintenance method based on the cloud-edge-end architecture proposed by the present invention effectively solves the technical problem that the equipment failure caused by the regular maintenance and the failure-based maintenance method seriously affects the production through the following steps: First, at the edge layer, the operation data of the equipment is collected and preprocessed by denoising and standardization to ensure the quality and consistency of the data; Second, the edge layer conducts preliminary cleaning and outlier detection on the preprocessed data to further improve the reliability of the data and transmits the cleaned data to the cloud; At the cloud, the equipment health database, maintenance activity data, and maintenance resource database are integrated through data fusion technology to provide comprehensive data support for subsequent analysis; Finally, the integrated data is predicted and analyzed using the trained deep learning model to accurately predict the occurrence probability and type of equipment failure and formulate corresponding maintenance strategies. This method not only improves the accuracy of equipment failure prediction but also realizes the timeliness and predictability of maintenance activities, greatly reducing the impact of failures on production and improving production efficiency and equipment service life.

[0168] The above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A dynamic data integration and predictive maintenance method based on cloud-edge-end architecture, characterized in that: The following steps are involved: Step 1: The terminal layer collects the operation data of the equipment and pre-processes the collected data, including denoising and standardization; Step 2: The edge layer receives the data preprocessed by the end layer, performs preliminary data cleaning and outlier detection on the received data, and sends the cleaned and outlier detected data to the cloud; Step 3: The cloud stores the received equipment operation data into the equipment health database, and integrates the data of the equipment health database, maintenance activity data, and maintenance resource database through data fusion technology; Step 4: Use the trained deep learning model and integrated data to make predictions to obtain the probability of equipment failure, failure type, and maintenance strategy.

2. The dynamic data integration and predictive maintenance method based on cloud-edge-end architecture according to claim 1, characterized in that: The data integration in step 3 includes data integration for equipment failure probability, data integration for equipment type prediction, and data integration for maintenance strategy prediction.

3. The dynamic data integration and predictive maintenance method based on cloud-edge-end architecture according to claim 2 is characterized in that: The steps of data integration for equipment failure probability are as follows: Extracting temperature features, humidity features, load features, health change trend features, and operating time features related to equipment failures based on the equipment health database; The temperature feature extraction is as follows: Where: T avg (t) represents the average ambient temperature at time t; N temp represents the number of temperature data points recorded within time t; T i (t) represents the value of the i-th temperature data point; Humidity feature extraction is as follows: Where: H avg (t) represents the average ambient humidity at time t; N humidity represents the number of humidity data points recorded in event t; H i (t) represents the value of the i-th humidity data point; The load characteristics are extracted as follows: Where: L avg (t) represents the average load at time t; N load Indicates the number of load data points recorded within time t; L i (t) represents the value of the i-th load data point; The health change trend features are extracted as follows: HI(t)=w1·HI1(T)+w2·HI2(t)+…+w n ·HI n (t); Where: HI1(T), HI2(t), …, HI n (t) represents the health indicator of the device after processing the sensor data at time t; w1, w2, ..., w n Represents the weight of each sensor data; HI(t) represents the health index of the device at time t; ΔHI(t)=HI(t)-HI(t-1); Where: ΔHI(t) represents the change in the equipment health index; The running time features related to equipment failure are extracted as follows: Where: T run (t) represents the cumulative operating time of the equipment from the time it was put into operation to the current time point t; Δt i Indicates the time of the device in the i-th operation cycle; Extracting fault history features, maintenance duration features, and equipment maintenance strategy features based on the maintenance activity database; The fault history feature extraction is as follows: Where: F freq (t) represents the frequency of equipment failure within time t; N failures (t) represents the number of equipment failures within time t; T total (t) represents the total operating time of the equipment within time t; ΔT fail (t)=t-T last_failure ; Where: ΔT fail (t) represents the time interval from the last equipment failure to the current time; T last_failure Indicates the time when the last fault of the device occurred; t indicates the current time; The maintenance duration feature is extracted as follows: Where: D repair (t) represents the average maintenance time of the equipment within time t; N repair (t) represents the number of times the equipment is repaired within time t; T repair,i represents the time spent on the i-th maintenance; The equipment maintenance strategy features are extracted as follows: Where: M freq (t) represents the maintenance frequency within time t; N maintenaces (t) represents the number of maintenance times within time t; T total (t) represents the total operating time of the equipment within time t; Based on the above extracted features, feature fusion is performed to form a comprehensive feature vector, including temperature features, humidity features, load features, health change trend features, equipment failure-related operating time features, fault history features, maintenance duration features, and equipment maintenance strategy features.

4. The dynamic data integration and predictive maintenance method based on cloud-edge-end architecture according to claim 2, characterized in that: The steps of data integration for device type prediction are as follows: Extract health change trend characteristics, equipment failure-related operation time characteristics, temperature characteristics, humidity characteristics, and equipment load based on the equipment health database; Extract the fault type history characteristics, maintenance type characteristics, maintenance duration characteristics, fault mode change trend characteristics and fault mode evolution trend characteristics from the maintenance activity database; The extracted features are fused to form a comprehensive feature vector, including health change trend features, equipment failure-related operating time features, temperature features, humidity features, equipment load, fault type history features, maintenance type features, maintenance duration features, failure mode change trend features, and failure mode evolution trend features.

5. The dynamic data integration and predictive maintenance method based on cloud-edge-end architecture according to claim 2, characterized in that: The steps of data integration for maintenance strategy prediction are as follows: Extract the maintenance duration, maintenance type, maintenance cost, maintenance personnel load, spare parts availability, equipment health status and maintenance strategy type for each maintenance based on maintenance activity data and maintenance resource database; Based on the above extracted features, feature fusion is performed to form a comprehensive feature vector, including maintenance time, maintenance type, maintenance cost, maintenance personnel load, spare parts availability, equipment health status and maintenance strategy type.

6. The dynamic data integration and predictive maintenance method based on cloud-edge-end architecture according to claim 1, characterized in that: The specific structure of the deep learning model is as follows: Input layer: includes three input units, wherein the first input unit is used to input the input features of fault probability prediction; the second input unit is used to input the input features of fault type prediction; the third input unit is used to input the input features of maintenance strategy prediction; Feature extraction layer: including the fault probability prediction feature extraction layer, the equipment type prediction feature extraction layer and the maintenance strategy prediction feature extraction layer; The fault occurrence probability feature extraction layer converts the input features into high-dimensional feature representations through a fully connected layer; The fault type prediction feature extraction layer generates hidden layer features related to the equipment type task based on the input features; The maintenance strategy prediction feature extraction layer extracts features from the input data to obtain a preliminary feature representation of the maintenance strategy; Shared feature layer: The output results of the fault probability feature extraction layer and the fault type prediction feature extraction layer are fused through a shared feature fusion layer; Fault probability prediction: Based on the output of the shared feature layer, the probability of fault occurrence is output through the fully connected layer; Fault type prediction: Based on the output of the shared feature layer, the fully connected layer processes the output and outputs the predicted fault type. Fusion layer: Fusion the prediction results of fault type and fault probability with the output results of the maintenance strategy prediction feature extraction layer to form a complete feature vector; Maintenance strategy prediction layer: The output of the fusion layer is input into the maintenance strategy prediction layer, and feature extraction is performed through a deep neural network composed of fully connected layers; Output layer: Outputs the type of maintenance strategy and maintenance time based on the output results of the maintenance strategy prediction layer.

7. The method for dynamic data integration and predictive maintenance based on cloud-edge-end architecture according to claim 6, characterized in that: The loss function of the deep learning model is as follows: L=λ1L prob +λ2L type +λ3L strategy ; Where: L prob represents the prediction loss of the probability of failure, using mean square error; L type The prediction loss for the fault type is the cross entropy loss; L strategy represents the maintenance strategy prediction loss, which is the cross entropy loss; λ1, λ2, and λ3 all represent weighting coefficients.

Images