Industrial big data-based equipment full life cycle intelligent management system

Abstract

The present application relates to the technical field of industrial big data, and in particular to a device full life cycle intelligent management system based on industrial big data, comprising a multi-source heterogeneous data acquisition and fusion module, a data value perception storage module, a device health state evaluation and adaptive prediction module, a full life cycle business decision support module, a dynamic data archiving and migration module, and a man-machine interaction module.The present application comprehensively integrates device full life cycle data through multi-source heterogeneous data fusion and multi-head attention mechanism, and improves decision comprehensiveness; adopts dynamic shunting and bimodal compression based on Shannon entropy, distinguishes between key transient and stable normal data, and improves storage efficiency; combines an adaptive prediction model based on online learning, responds to device aging and working condition drift, and improves fault prediction accuracy; dynamically archives fusion data access frequency and device health state, avoids key data error archiving, and guarantees diagnosis real-time performance.

Description

Technical Field

[0001] This invention relates to the field of industrial big data technology, and more specifically, to an intelligent management system for the entire lifecycle of equipment based on industrial big data. Background Technology

[0002] With the deepening of industrial big data and intelligent manufacturing, modern industrial equipment is increasingly developing towards larger scale, greater complexity, and greater automation, and the proportion of equipment assets in enterprise production and operation continues to rise. The entire lifecycle management of equipment, from procurement, installation, operation, maintenance to disposal, directly impacts an enterprise's production efficiency, product quality, operation and maintenance costs, and safety levels. In recent years, industrial big data technology has flourished, providing a new technological path for intelligent equipment management through the collection and analysis of multi-source heterogeneous data such as equipment operating status, maintenance records, and environmental parameters.

[0003] In recent years, industrial internet and big data technologies have made some progress in the field of equipment management. For example, Chinese invention patent application CN120124836A proposes a method for full lifecycle management of equipment based on an industrial internet platform, which realizes data collection, predictive maintenance, and remote control through the equipment layer, edge layer, PaaS layer, and SaaS layer. However, this method has the following drawbacks: First, it adopts a uniform processing strategy for all equipment operation data, failing to distinguish the differences in data value density under different operating conditions (such as steady-state operation and transient faults), resulting in massive amounts of normal data occupying a large amount of storage resources, while critical transient data may be overwhelmed due to insufficient sampling; Second, its predictive maintenance model relies on fixed machine learning modules and lacks an adaptive update mechanism for data distribution drift, making it difficult to cope with model failure caused by equipment aging or changes in operating conditions; Third, this method does not involve core business links such as procurement decisions, spare parts inventory optimization, and asset ledger management in the entire equipment lifecycle, resulting in limited decision support capabilities.

[0004] Another Chinese invention patent application, CN121765405A, proposes a method and system for industrial data lifecycle management. It distinguishes between critical transient windows and stable normal windows using Shannon entropy calculation, and employs differential coding and DCT sparsification for separate processing to construct a hybrid storage structure. It then implements hierarchical archiving based on access frequency. While this method is innovative in data storage optimization, it still has significant shortcomings: First, its core focus is on the compression and hierarchical storage of time-series data, without addressing business decision support in equipment lifecycle management (such as maintenance planning, spare parts demand forecasting, and scrap assessment), limiting the technical solution to the data management level. Second, this method only processes numerical sensor data and lacks the ability to integrate and analyze unstructured maintenance records, fault description texts, equipment procurement contracts, and other multi-source heterogeneous data. Third, its hot and cold data migration strategy is based solely on query access frequency, without considering changes in the equipment's own health status or predictive maintenance needs, which may lead to the incorrect archiving of equipment data about to fail, affecting the real-time nature of fault diagnosis.

[0005] In summary, existing technical solutions still have significant shortcomings in areas such as data fusion, value-aware storage, intelligent decision support, and multi-service collaboration for equipment lifecycle management. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent equipment lifecycle management system based on industrial big data, in order to solve the problems of existing equipment lifecycle management systems mentioned in the background art, such as the lack of differentiation of data value density, lack of adaptive prediction models, insufficient support for lifecycle business decisions, poor ability to integrate multi-source heterogeneous data, and disconnect between cold and hot migration and equipment health status.

[0007] To achieve the above objectives, the present invention aims to provide an intelligent management system for the entire lifecycle of equipment based on industrial big data, including a multi-source heterogeneous data acquisition and fusion module, a data value perception and storage module, an equipment health status assessment and adaptive prediction module, a lifecycle business decision support module, a dynamic data archiving and migration module, and a human-computer interaction module.

[0008] The multi-source heterogeneous data acquisition and fusion module is used to collect multi-source heterogeneous data throughout the entire lifecycle of the device, including structured data, semi-structured data, unstructured data, and real-time sensor data. This module uses a pre-trained BERT model to extract textual semantic feature vectors and employs a multi-head attention mechanism to weightedly fuse the multi-source features, generating a unified feature vector for the entire device lifecycle. .

[0009] The data value perception and storage module, connected to the multi-source heterogeneous data acquisition and fusion module, is used to perform importance scoring based on information entropy for time series data of real-time sensing data. This module divides the data sequence into time windows of preset length and calculates the Shannon entropy for each window. The elbow rule is used to dynamically determine the shunt threshold. ,Will The window is marked as a critical transient window, otherwise it is marked as a stationary normal window. Differential pulse code modulation lossless compression is used for the critical transient window, and discrete cosine transform and non-uniform quantization lossy compression is used for the stationary normal window to construct a hybrid storage data frame.

[0010] The equipment health status assessment and adaptive prediction module, connected to the data value perception and storage module, is used to construct an equipment health assessment model based on the fused full lifecycle feature vectors and predict the remaining lifespan using support vector regression. Random survival forest is used to predict future time windows. Internal failure probability This module also includes an online learning and update unit that periodically collects new data to calculate prediction errors. When the error exceeds a preset threshold, it triggers model retraining to achieve adaptive model updates.

[0011] The full lifecycle business decision support module connects with the equipment health status assessment and adaptive prediction module to score equipment health. Remaining lifespan and failure probability By combining multi-dimensional constraints such as spare parts inventory, maintenance resources, and production plans, a full lifecycle business decision-making solution is generated, including maintenance plans, spare parts procurement plans, overhaul plans, and scrap assessments. The maintenance plan optimization employs a multi-objective function. ,in, The total cost of preventative maintenance or troubleshooting for equipment. The length of time that the equipment cannot operate normally due to maintenance or malfunction. These are weighting coefficients used to balance the importance of different objectives in a multi-objective optimization function.

[0012] The dynamic data archiving and migration module, connected to the data value-aware storage module and the equipment health status assessment and adaptive prediction module, dynamically adjusts the storage hierarchy of data segments in the hybrid storage structure based on data access frequency, equipment health status, and predicted maintenance needs. This module also calculates a popularity value based on query frequency. It obtains the current device health score and remaining lifespan. If the remaining lifespan or health score is below the threshold, it prohibits migrating the corresponding data segment to cold storage, ensuring that critical data always resides in the high-speed storage layer. Furthermore, the background simplified data segments in normal health status are migrated from solid-state drives to mechanical hard drives or tape libraries to achieve hierarchical archiving of hot and cold data.

[0013] The human-computer interaction module, connected to the above modules, is used to visually display the device's full lifecycle status, health trends, prediction results, business decision-making solutions, and storage resource usage, and supports users to adjust decision parameters online.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements the above-described system when executing the computer program.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0016] 1. In this intelligent management system for the entire lifecycle of equipment based on industrial big data, a multi-source heterogeneous data acquisition and fusion module is set up, and a multi-head attention mechanism is used to fuse structured, semi-structured, unstructured and real-time sensor data. This solves the problem of single data dimension and inability to effectively fuse multi-source heterogeneous data in the existing technology, and significantly improves the comprehensiveness and accuracy of the decision-making basis for the entire lifecycle.

[0017] 2. In this intelligent equipment lifecycle management system based on industrial big data, by setting up a data value perception storage module, and adopting dynamic diversion decision based on Shannon entropy and dual-modal compression coding, the critical transient window and the stable normal window are distinguished. This solves the problem of data value density not being distinguished and massive normal data occupying storage resources in the existing technology, and greatly improves storage efficiency while ensuring that key fault information is intact.

[0018] 3. In this intelligent equipment lifecycle management system based on industrial big data, by setting up an equipment health status assessment and adaptive prediction module, and combining it with an online learning mechanism to adaptively update the prediction model, the problem of the lack of adaptability of the prediction model to data distribution drift in the existing technology is solved. It can cope with equipment aging and changes in operating conditions, and significantly improve the accuracy and robustness of fault prediction.

[0019] 4. In this intelligent equipment lifecycle management system based on industrial big data, by setting up a lifecycle business decision support module, the system deeply integrates equipment health status with business processes such as maintenance plans, spare parts procurement, and scrap assessment, which solves the problem of lacking lifecycle business decision support in existing technologies and realizes intelligent closed-loop management from data to decision.

[0020] 5. In this intelligent equipment lifecycle management system based on industrial big data, by setting up a dynamic data archiving and migration module, data access popularity and equipment health status are integrated, avoiding the incorrect archiving of critical data of equipment about to fail. This solves the problem of the disconnect between cold and hot migration strategies and equipment health status in existing technologies, and ensures the real-time nature of fault diagnosis and data availability. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall principle of the intelligent management system for the entire life cycle of equipment based on industrial big data of the present invention.

[0022] Figure 2 This is a schematic diagram of the principle of the multi-source heterogeneous data acquisition and fusion module of the present invention.

[0023] Figure 3 This is a schematic diagram of the data value perception storage module of the present invention.

[0024] Figure 4 This is a schematic diagram of the data flow of the dynamic data archiving and migration module of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] In one specific embodiment, such as Figure 1 As shown, the intelligent equipment lifecycle management system based on industrial big data includes a multi-source heterogeneous data acquisition and fusion module, a data value perception and storage module, an equipment health status assessment and adaptive prediction module, a lifecycle business decision support module, a dynamic data archiving and migration module, and a human-machine interaction module.

[0027] The multi-source heterogeneous data acquisition and fusion module is used to collect multi-source heterogeneous data throughout the entire life cycle of the device, and to perform feature extraction and cross-modal fusion on the multi-source heterogeneous data to generate a unified feature vector for the entire life cycle of the device.

[0028] Through the multi-source heterogeneous data acquisition and fusion module, equipment ledgers, purchase contracts, maintenance records, operation logs, fault texts, and real-time sensor data such as vibration and temperature can be uniformly integrated, breaking the limitations of data silos in traditional systems and significantly improving the comprehensiveness and accuracy of decision-making basis throughout the entire life cycle.

[0029] like Figure 2As shown, the multi-source heterogeneous data acquisition and fusion module includes a structured data extraction unit, a semi-structured data parsing unit, an unstructured data processing unit, a real-time sensor data access unit, and a cross-modal fusion unit.

[0030] The structured data extraction unit is used to collect structured data of equipment, including ledgers, purchase contracts, maintenance records, and spare parts inventory. Specifically, it first collects structured data of equipment from an enterprise resource planning system, manufacturing execution system, or database management system. This structured data includes equipment ledgers, purchase contracts, maintenance records, and spare parts inventory. The collected structured data undergoes format validation and outlier detection, removing data with obvious errors or exceeding reasonable limits (such as negative inventory, future timestamps, etc.). Then, the validated structured data is organized according to a unified data format to generate corresponding structured feature vectors. The output is sent to the cross-modal fusion unit.

[0031] The semi-structured data parsing unit is used to parse semi-structured data from equipment, including operation logs and maintenance work orders. Specifically: First, semi-structured data, including equipment operation logs and maintenance work orders, is obtained from the equipment control system or maintenance management system. Then, the operation logs are parsed, extracting key fields (such as timestamps, alarm codes, operation events, and operating status) and converting them into key-value pair format. Next, the maintenance work orders are parsed, extracting information such as work order number, equipment ID, fault description, repair measures, repair duration, and repair personnel, and converting them into structured table format. Finally, the processed semi-structured data is converted into a unified feature representation, generating corresponding semi-structured feature vectors. The output is sent to the cross-modal fusion unit.

[0032] The unstructured data processing unit is used to extract semantic feature vectors from unstructured data, including fault description text, maintenance reports, and equipment manuals, using a pre-trained BERT model. Specifically: First, unstructured data, including fault description text, maintenance reports, and equipment manual PDFs, is retrieved from a document management system or historical database. Then, the text data is preprocessed, including removing special characters, word segmentation, and stop word removal. For PDF files, text extraction is performed first (e.g., using OCR or a PDF parsing library). Next, the pre-trained BERT model is used to extract semantic features from the preprocessed text, converting each text paragraph or the entire document into a fixed-dimensional semantic feature vector. Finally, all semantic feature vectors are pooled (e.g., average pooling) to obtain the overall unstructured data semantic feature vector. The output is sent to the cross-modal fusion unit.

[0033] By using the BERT model to extract deep semantic features from unstructured text, implicit knowledge in maintenance reports and fault descriptions can be automatically mined, providing richer information for equipment health assessment than manual annotation.

[0034] The real-time sensor data access unit is used to access real-time sensor data, including vibration, temperature, current, and pressure. Specifically: First, it accesses the real-time sensor data deployed on the device via an industrial Ethernet, wireless sensor network, or fieldbus interface. This real-time sensor data includes vibration (accelerometer), temperature (thermocouple / resistance temperature detector), current (current transformer), and pressure (pressure transmitter). Then, it performs time synchronization on the real-time sensor data to ensure all sensor data have a unified time reference (e.g., using NTP protocol or hardware synchronization signal). Next, it performs filtering and noise reduction on the real-time sensor data, using moving average filtering, median filtering, or low-pass filtering to remove high-frequency noise and abnormal pulses. Finally, it extracts features from the processed real-time sensor data according to sensor type (e.g., extracting time-domain features: mean, variance, peak value, root mean square; frequency-domain features: FFT spectral energy distribution) to generate corresponding real-time sensor feature vectors. The output is sent to the cross-modal fusion unit.

[0035] The real-time sensor data access unit ensures low-latency transmission and accurate synchronization of high-frequency sampled data, enabling subsequent fault prediction to be based on the actual dynamic response of the device, rather than the fuzzy trend after downsampling.

[0036] The cross-modal fusion unit is used to perform weighted fusion of the multi-source heterogeneous data using a multi-head attention mechanism to generate a unified device lifecycle feature vector. Specifically: First, it receives structured feature vectors from the structured data extraction unit. Semi-structured feature vectors from semi-structured data parsing units Unstructured semantic feature vectors from unstructured data processing units and real-time sensing feature vectors from the real-time sensing data access unit Then, a multi-head attention mechanism is used to weight and fuse the four types of feature vectors to calculate the association weights between different modal features: a query vector, key vector, and value vector are generated for each modal feature, and the attention weights between modalities are calculated using scaled dot product attention. Next, based on the calculated attention weights, the modal features are weighted and summed to generate a unified device lifecycle feature vector. Finally, the fused device lifecycle feature vector is output to the data value perception and storage module and the device health status assessment and adaptive prediction module as input for subsequent analysis and decision-making.

[0037] The cross-modal fusion of the multi-head attention mechanism can automatically learn the correlation between vibration and maintenance records, and the coupling between temperature and load rate, so that equipment health status assessment no longer depends on a single data source, significantly improving the robustness and generalization ability of the model.

[0038] like Figure 3 As shown, the data value perception storage module is connected to the multi-source heterogeneous data acquisition and fusion module. It is used to perform an importance score based on information entropy on the time series of real-time sensing data, and divide the data window into a critical transient window and a stable normal window according to the score. The critical transient window is stored using lossless differential encoding, while the stable normal window is stored using lossy compression. By distinguishing between critical transient data and stable normal data, this module can compress the storage amount of stable background data to 5%-10% of the original while ensuring zero loss of fault symptoms, which greatly reduces the long-term storage cost and constructs a hybrid storage structure.

[0039] The data value perception storage module includes a time window segmentation unit, an importance scoring unit, a flow decision unit, and a hybrid encoding unit.

[0040] The time window segmentation unit is used to segment the real-time sensor data sequence into time windows of a preset length. Specifically: First, continuous sensor time series data (such as vibration, temperature, current, pressure, etc.) is acquired from the real-time sensor data access unit. Then, according to the preset time window length (e.g., 1024 sampling points), the original time series data is sequentially segmented into multiple non-overlapping time windows. The preset time window length is determined based on the statistical average duration of typical transient events (such as fault impacts, load surges) in historical data, typically taken as 1.2 times the duration distribution, to ensure that a single window can completely encompass an independent transient event. Finally, the data from each segmented time window is sequentially output to the importance scoring unit.

[0041] Importance scoring units are used to calculate the Shannon entropy of the data sequence within each time window. Specifically: First, the system receives the original data sequence within each time window output by the time window segmentation unit. Then, it performs numerical discretization on the original data within the window, mapping continuous analog signal values ​​to discrete symbols in a finite set of states (using the symbol aggregation approximation SAX method). Next, it calculates the probability distribution of each discrete symbol appearing within the current time window, assuming the number of discretized states is... , No. The probability of each state occurring within the window is: Then, the information entropy value of the window is calculated using the Shannon entropy formula. The calculation formula is:

[0042]

[0043] Finally, the calculated Shannon entropy is... This score serves as the importance score for that time window, and is stored in the temporal importance score sequence in chronological order. The above steps are repeated until all time windows have been processed, at which point the complete temporal importance score sequence is output to the flow decision unit.

[0044] The flow decision unit is used to determine the Shannon entropy. With adaptive shunt threshold If a comparison is made, If the condition is met, it is marked as a critical transient window; otherwise, it is marked as a stationary normal window. Specifically: First, the temporal importance score sequence output by the importance scoring unit is received. Then, the elbow rule is used to dynamically determine the adaptive shunting threshold. :

[0045] Sort the time-series importance score sequences in descending order to obtain an ordered score sequence;

[0046] Construct a rating distribution curve with the sorted rank as the x-axis and the corresponding rating value as the y-axis;

[0047] Connect the start and end points of the curve, and calculate the vertical distance from each data point on the curve to the connecting line;

[0048] The score corresponding to the maximum vertical distance is taken as the adaptive triage threshold. .

[0049] Next, iterate through all time windows and calculate the Shannon entropy for each window. With adaptive shunt threshold Comparison:

[0050] like If so, mark the window as a critical transient window (containing fault symptoms or important events).

[0051] like If so, then mark the window as a stable normal window (steady-state operating background data).

[0052] Finally, the labeled window classification results, along with the original window data, are output to the hybrid encoding unit.

[0053] Adaptive diversion based on Shannon entropy can automatically identify the data value boundary under different operating conditions without the need for manual threshold setting, thus avoiding the omission of key information or excessive retention of normal data due to inappropriate experience values.

[0054] The hybrid coding unit uses differential pulse code modulation for lossless compression of critical transient windows and discrete cosine transform and non-uniform quantization for lossy compression of stationary windows, generating hybrid storage data frames. Specifically: First, it receives the window classification results and corresponding original window data output from the stream decision unit. Then, for data marked as critical transient windows:

[0055] Extract the raw data sequence within the window Keep the first data point As the baseline value for the entire quantity, Represents the total number of data points in the original data sequence;

[0056] First-order differential pulse code modulation (DPCM) is used to calculate the difference between adjacent data points: ;

[0057] Based on a pre-set global mapping dictionary (a Huffman coding table generated from historical data statistics), the difference value sequence is matched to the corresponding binary codeword;

[0058] The binary encoding of the base value and the differential codeword are concatenated in sequence to generate a lossless compressed binary bit stream, i.e., a high-fidelity data segment.

[0059] Then, for the data marked as a stationary normal window:

[0060] Extract the original data sequence within the window and use a one-dimensional discrete cosine transform (DCT) to convert the time-domain signal into a frequency-domain coefficient matrix;

[0061] Based on the preset compression intensity (target energy retention rate, such as 95%), a high-frequency component mask matrix is ​​generated, and the high-frequency coefficients are set to zero to obtain a sparse transformation coefficient matrix.

[0062] A non-uniform quantizer is used to perform interval mapping and encoding on the retained low-frequency coefficients to generate a lossy compressed binary bit stream, i.e., the background simplified data segment.

[0063] Next, the generated high-fidelity data segment and the background simplified data segment are physically concatenated according to a preset frame structure (header information + background segment + high-fidelity segment + checksum) to generate a hybrid storage data frame. Finally, the hybrid storage data frame is written to the first-level storage medium (such as a solid-state drive), and a metadata index containing timestamps, importance tags, and physical addresses is created for subsequent retrieval and archiving.

[0064] The dual-modal compression strategy enables 100% lossless restoration of critical transient data for fault diagnosis, while the compression ratio of stable background data can reach more than 20 times, greatly saving storage space while ensuring analysis accuracy.

[0065] The equipment health status assessment and adaptive prediction module is connected to the data value perception and storage module. It is used to construct an equipment health assessment model based on the fused equipment life cycle feature vector, and to adaptively update the prediction model using an online learning mechanism, outputting equipment health score, remaining life prediction and potential failure probability.

[0066] This module quantifies the health status of equipment into a continuous score between 0 and 1, and provides the remaining lifespan and failure probability, providing a quantifiable scientific basis for maintenance planning and spare parts procurement, thus changing the traditional "periodic maintenance" model.

[0067] The equipment health status assessment and adaptive prediction module includes a health assessment unit, a remaining life prediction unit, a failure probability prediction unit, and an online learning and updating unit.

[0068] Health assessment units are used to assess health based on full life-cycle feature vectors. Construct a device health scoring function. Specifically: First, receive the device's full lifecycle feature vector output by the cross-modal fusion unit. Then, the weight vector is trained based on historical data. and bias terms Construct a device health scoring function:

[0069]

[0070] in, The device health score is given, with an output range of [0,1]. A higher score indicates a better device health status. This is the Sigmoid function, with an output range of [0,1]. It is a weight vector The transpose of .

[0071] Finally, the calculated device health score will be used. Output to the full lifecycle business decision support module and the dynamic data archiving and migration module.

[0072] For example: Health assessment models use a logistic regression form of the Sigmoid function.

[0073]

[0074] Assuming the feature vector of the entire life cycle of the device The weight vector includes eight key features (such as root mean square vibration, mean temperature, peak current, pressure variance, equipment age, maintenance frequency, fault text semantic score, and load rate). and bias terms The example data is shown in the table below:

[0075] Feature number Feature Name weight value illustrate 1 Root mean square of vibration 0.28 The greater the vibration intensity, the lower the health status (positive correlation). 2 Average temperature 0.18 Excessively high temperature indicates an abnormality 3 Peak current 0.15 Current surge reflects sudden load changes 4 Pressure variance 0.12 Pressure fluctuations reflect system stability 5 Equipment service life 0.10 The longer the service period, the lower the health level. 6 Number of repairs 0.08 Frequent repairs indicate low reliability. 7 Error text semantic score 0.05 Risk score extracted from maintenance report 8 load rate 0.04 Prolonged high load reduces health

[0076] Weight vector ,satisfy .

[0077] Bias term .

[0078] The remaining lifetime prediction unit is used to predict the remaining lifetime using a support vector regression or random forest regression model, with historical feature sequences as input. Specifically: First, historical feature sequences (composed of full lifecycle feature vectors from multiple time windows) and corresponding actual remaining lifespan labels of the devices are extracted from the historical database. Then, a support vector regression or random forest regression model is used, with the historical feature sequences as input and the actual remaining lifespan as output, to train the remaining lifespan prediction model. Next, the device's full lifecycle feature vector at the current moment is obtained in real time. And the device's full lifecycle feature vector at the current moment. Input the trained remaining lifetime prediction model to predict the remaining lifetime of the equipment. (Unit: hours or operating cycles). Finally, the predicted remaining lifespan will be... Output to the full lifecycle business decision support module and the dynamic data archiving and migration module.

[0079] The failure probability prediction unit uses a random survival forest model to output the device's future time window. Internal failure probability Specifically: First, historical feature sequences of the equipment and corresponding fault occurrence times and fault status labels (0 = normal, 1 = fault) are extracted from the historical database. Then, a random survival forest algorithm is used, with the historical feature sequences as input and the fault status and fault occurrence time as output, to establish a cumulative risk function and a survival function. Next, the equipment's full lifecycle feature vector is obtained in real time. And set the length of the future time window. (e.g., 7 days, 30 days, etc.) Using a trained fault probability prediction model, the output of the device's future fault prediction results is determined. Probability of failure occurring within a given time period Finally, the predicted failure probability will be... Output to the full lifecycle business decision support module.

[0080] The trained fault probability prediction model (random survival forest) is shown in the table below:

[0081] Parameter name Example value illustrate Model type Random Survival Forest Used to predict the probability of device failure within a future time window. Number of trees 100 Number of decision trees in ensemble learning Maximum depth 5 The maximum depth of each decision tree to prevent overfitting. Minimum number of samples per node 10 Minimum number of samples required for an internal node to continue splitting Split Criteria division Node splitting is based on maximizing survival differences. Feature sampling number 3 The number of features randomly selected in each split Predicted output future The probability of failure within a given time period, with a value ranging from [0,1].

[0082] For example, based on historical data of a certain type of CNC machine tool (including the operating characteristic sequence and failure time label of 500 machines), the C-index (consistency index) of the trained random survival forest model on the test set is 0.82, indicating that the model has good ranking ability.

[0083] The online learning update unit is used to periodically collect new data, calculate prediction errors, and trigger model retraining when the error exceeds a preset threshold, thus achieving adaptive model updates. Specifically: First, new equipment operation data, maintenance records, and fault information are collected periodically (e.g., weekly or monthly) to construct a new training sample set. Then, the prediction error of the current prediction model (health assessment model, remaining life prediction model, and failure probability prediction model) is calculated using the new sample set. Next, a preset model update threshold is set (e.g., mean absolute error of health score > 0.05, remaining life prediction error > 10%, AUC decrease in failure probability prediction > 5%, where AUC is the area under the ROC curve of the failure probability prediction model, with AUC values ​​between 0 and 1; the closer the value is to 1, the stronger the model's ability to distinguish between faulty and non-faulty samples, and the more reliable the prediction results; the closer the value is to 0.5, the closer the model's prediction ability is to random guessing). The calculated prediction error is compared with the preset threshold. If the prediction error of any model exceeds the preset threshold, model retraining is triggered: the new samples are merged with historical samples (or a sliding window is used to retain the most recent N samples), and the model is retrained. The retrained model parameters (such as weight vectors) Bias terms The model updates each prediction unit with support vectors from support vector regression and tree structures from random survival forests, replacing the original model. Finally, a model update log is recorded, including update time, triggering reason, training sample size, and performance comparison before and after the update, for system monitoring and auditing.

[0084] The online learning and updating mechanism enables the prediction model to continuously adapt to long-term drift such as equipment aging and changes in operating conditions, avoiding the problem of a sharp drop in accuracy after several months of operation of traditional fixed models, and significantly improving the long-term reliability of predictions.

[0085] The full lifecycle business decision support module is connected to the equipment health status assessment and adaptive prediction module. It is used to generate a full lifecycle business decision scheme, including maintenance plan, spare parts procurement plan, overhaul plan and scrap assessment, based on the equipment health score, remaining life prediction and potential failure probability, combined with the constraints of spare parts inventory, maintenance resources and production plan.

[0086] This module directly transforms the quantitative results of health status into executable maintenance, procurement, and disposal plans, connecting the entire chain of "data → status → decision → execution" and realizing an intelligent closed loop for equipment management.

[0087] The full lifecycle business decision support module includes a maintenance plan optimization unit, a spare parts procurement optimization unit, and a scrap assessment unit.

[0088] The maintenance plan optimization unit is used to score equipment health. and remaining lifespan Taking into account maintenance resource constraints, a preventative maintenance plan is generated using a multi-objective optimization function. Specifically: First, the system receives the equipment health score output from the equipment health status assessment and adaptive prediction module. and remaining lifespan Then, obtain current maintenance resource constraints, including the number of available maintenance personnel, the inventory of maintenance tools and spare parts, and the planned downtime window. Next, construct a multi-objective optimization function that comprehensively considers maintenance costs, downtime, and equipment health.

[0089]

[0090] in, The total cost of preventative maintenance or troubleshooting for equipment. The length of time that the equipment cannot operate normally due to maintenance or malfunction. These are weighting coefficients used to balance the importance of different objectives in a multi-objective optimization function, and they satisfy... It can be adjusted according to the company's strategy, for example =0.4、 =0.3、 =0.3. Then, based on the remaining lifespan... Set a time window for preventative maintenance (e.g., in) Maintenance is scheduled 80%-90% of the time, in conjunction with equipment health scores. Determine whether early intervention is needed (e.g.) (If the value is less than 0.3, maintenance should be arranged immediately). Simultaneously, a multi-objective optimization problem is solved (using weighted sum or Pareto optimization methods), outputting a preventative maintenance plan, including suggested maintenance times, maintenance content (such as bearing replacement, lubrication, calibration, etc.), a list of required spare parts, and estimated downtime. Finally, the generated maintenance plan is output to the human-machine interface module for equipment management personnel to review and execute.

[0091] Multi-objective optimization can automatically balance maintenance costs, production losses, and equipment health, avoiding the waste caused by "over-maintenance" and the risks of "under-maintenance".

[0092] The spare parts procurement optimization unit is used to optimize the procurement based on the probability of failure. To determine the optimal spare parts inventory level and spare parts lead time, dynamic programming is used. Specifically: First, the equipment receives the output from the failure probability prediction unit within the future time window. Internal failure probability Then, obtain the current inventory level of spare parts, the lead time for spare parts procurement, the unit price of spare parts, and the cost of stockouts. Next, use dynamic programming to determine the optimal spare parts inventory level, with the objective of minimizing total costs (including holding costs, procurement costs, and stockout costs).

[0093] Define a state variable: current inventory level. ;

[0094] Decision variable: Order quantity ;

[0095] Cost function: ,in, This is a total cost function, representing the total cost at the current inventory level. Next, order The total cost of spare parts The cost of ordering spare parts for the unit (including purchase price, transportation, inspection and other expenses). The holding costs of spare parts for the unit (including warehousing, capital occupation, insurance, depreciation and other expenses). The cost of spare parts shortages per unit (production losses, emergency procurement premiums, downtime losses, etc. caused by spare parts shortages). For the predicted demand, This refers to the out-of-stock quantity, which is the total inventory after ordering. Unable to meet demand The part. If If so, the shortage quantity is 0.

[0096] In addition, forecast demand Indicates the next procurement lead time The quantity of spare parts the equipment may consume within the time frame from order placement to delivery. This demand is determined by the equipment's future... The probability of a failure occurring within a given timeframe and the number of spare parts required for each failure are determined by this factor. The specific calculation formula is as follows:

[0097]

[0098] in, The quantity of the same equipment or a group of equipment of the same type. For the first Taiwan equipment in the future time window The failure probability within (output by the failure probability prediction unit). For the first The number of spare parts required when a device fails (usually 1, but some components may require multiple).

[0099] The optimal ordering strategy is solved recursively.

[0100] Then, combine the equipment health score and remaining lifespan Adjust the safety stock level. Simultaneously, output spare parts procurement suggestions, including spare parts name, suggested quantity, suggested procurement time, and estimated cost. These suggestions are then output to the human-computer interaction module and can be linked with the enterprise resource planning system to generate purchase orders.

[0101] Among them, when lower or When the safety stock is low, appropriately increase it; for example, set the base safety stock coefficient to 1.0 (the safety stock level under normal circumstances). Based on... and The classification is used to determine the adjustment coefficient, as shown in the table below:

[0102] Device health rating Remaining lifespan (Hour) Adjustment coefficient illustrate ≥0.7 ≥500 1.0 The equipment is in good condition, and a basic safety stock is maintained. 0.5-0.7 200-500 1.2 The equipment is slightly deteriorated; additional spare parts should be added as needed. 0.3-0.5 100-200 1.5 Equipment with moderate risk; significantly increase safety stock. ＜0.3 ＜100 2.0 Equipment on the verge of failure, safety stock doubled.

[0103] like and They belong to different price ranges, so a higher adjustment factor is applied (i.e., a more conservative inventory strategy).

[0104] Adjusted safety stock = base safety stock × adjustment factor.

[0105] By combining dynamic programming with failure probability, spare parts inventory can be transformed from "empirical estimation" to "probabilistic prediction," which can reduce the capital tied up in spare parts inventory by an average of 20%-30% while ensuring supply security.

[0106] The end-of-life assessment unit comprehensively considers the remaining lifespan of the equipment, maintenance costs, residual value, and the cost of purchasing new equipment to calculate the economic life of the equipment and generate end-of-life recommendations. Specifically, it first receives the remaining lifespan output from the equipment health status assessment and adaptive prediction module. Then, obtain the equipment's cumulative operating years, historical maintenance cost data, current residual value estimate, new equipment purchase cost, and expected performance improvement. Next, calculate the equipment's economic life. The minimum average annual cost method is adopted:

[0107] Calculate the average annual total cost of equipment year by year ;

[0108] Select to make The smallest year As economic lifespan.

[0109] Then predict the remaining lifespan. With economic life Comparison:

[0110] like ≤ If the value is ×0.2, it is recommended to immediately initiate the scrap assessment process;

[0111] like > If the value is ×0.5, it is recommended to continue running and monitor continuously.

[0112] If it falls between these two extremes, it is recommended to develop a major overhaul or upgrade plan.

[0113] Simultaneously, a scrap assessment report is output, including a suggested scrapping time window, estimated residual value, recommendations for purchasing alternative equipment, and a return on investment analysis. The scrap assessment results are then output to the human-computer interaction module for management decision-making.

[0114] Economic life-based end-of-life assessments provide a quantitative financial basis for equipment replacement, avoiding asset losses caused by "premature scrapping" or "excessive service life" due to intuition-based decisions.

[0115] like Figure 4 As shown, the dynamic data archiving and migration module is connected to the data value perception storage module and the equipment health status assessment and adaptive prediction module. It is used to dynamically adjust the storage level of data segments in the hybrid storage structure according to the data access popularity, equipment health status and predicted maintenance needs, so as to realize the hierarchical archiving of hot and cold data.

[0116] The dynamic data archiving and migration module adds health status constraints to the traditional hot data archiving, ensuring that all data from devices about to fail is always retained in the high-speed storage layer, completely solving the problem of critical data response delay caused by cold and hot migration.

[0117] The dynamic data archiving and migration module includes an access popularity monitoring unit, a health status association unit, and a hierarchical migration unit.

[0118] The access frequency monitoring unit is used to statistically analyze the query frequency and access latency of each hybrid storage data frame and calculate the popularity value. Specifically: First, it statistically analyzes the query frequency and access latency of each hybrid storage data frame, including:

[0119] Preview visit count : Only read the number of queries for the simplified background data segment;

[0120] Deep access count : The number of queries to read high-fidelity data segments (or complete data frames).

[0121] Then obtain the system statistics for the maximum access frequency. (Maximum number of accesses per frame within the current period) and I / O time capacity limit (Maximum I / O processing capacity of a single storage node). The heat value is then calculated using the following formula. :

[0122]

[0123] in, I / O overhead (calculated from access patterns and the amount of data read). Let be the weighting coefficient, satisfying ( , ).

[0124] Finally, the calculated heat value Output to the health status association unit and the hierarchical migration unit.

[0125] The health status association unit is used to obtain the current device health score. and remaining lifespan And preset health threshold and remaining lifetime threshold ,like or If so, migrating the corresponding data segment to cold storage is prohibited. Specifically: Obtain the current device's health score. and remaining lifespan (From the device health status assessment and adaptive prediction module). Then, a health threshold is preset. (e.g., 0.3) and remaining lifetime threshold (For example, 100 hours). Then, conditional checks are performed:

[0126] like or If the device is deemed to be in a high-risk state, the corresponding data segment should not be migrated to cold storage.

[0127] Otherwise, normal migration is allowed based on the heat value. At the same time, the judgment result (migration allowed or migration prohibited) is output to the hierarchical migration unit.

[0128] The health status association unit acts as a "safety valve," automatically preventing data migration when the device is nearing a fault. This ensures that the high-fidelity data required for diagnosis is always stored in a high-speed storage layer, guaranteeing the real-time nature of fault analysis.

[0129] The graded migration unit is used to preset the cold storage temperature threshold. and will satisfy Furthermore, the background simplified data segment, which is in normal health condition, is migrated from the first-level storage medium to the second-level storage medium, while the metadata index is updated. Specifically: First, the access popularity value output by the access popularity monitoring unit is received. The migration permission flag is output by the unit associated with the health status. A preset cold storage heat threshold is also included. (For example, 0.2). Then determine whether the migration conditions are met:

[0130] like If the health status associated unit allows migration, then the background simplified data segment of the data frame is determined to be a "cold data block";

[0131] Otherwise, the migration will not be performed.

[0132] For data frames that meet the migration conditions:

[0133] Read the background simplified data segment from the primary storage medium (such as a solid-state drive);

[0134] Write the background simplified data segment to the secondary storage medium (such as a hard disk drive or tape library) in a sequential write manner to maximize throughput performance;

[0135] Update metadata index: Modify the original physical address mapping to point to a new address in the second-level storage medium, while retaining the high-fidelity data segment and metadata index in the first-level storage medium;

[0136] Send a TRIM or Deallocate command to the first-level storage medium to release the space occupied by the original background data segment;

[0137] Finally, a migration log is recorded, including information such as migration time, data frame ID, source address, destination address, and data segment size, for system auditing.

[0138] The human-computer interaction module connects to the above modules and is used to visually display the device's full lifecycle status, health trends, prediction results, business decision-making solutions, and storage resource usage, and supports users to adjust decision parameters online.

[0139] In a specific application, this system is implemented using a CNC machine tool from a manufacturing company as an example. The multi-source heterogeneous data acquisition and fusion module collects data from the Enterprise Resource Planning (ERP) system, Manufacturing Execution System (MES), sensor network, and document library: structured data includes machine tool ledgers, historical maintenance records, and spare parts inventory; semi-structured data includes machine tool operation logs (including alarm codes and operation records); unstructured data includes fault repair reports and equipment manual PDFs; and real-time sensor data includes spindle vibration, cutting temperature, spindle current, and coolant pressure. The BERT model is used to extract text semantic vectors, which are then fused using a multi-head attention mechanism to obtain a full lifecycle feature vector. .

[0140] The data value perception and storage module divides the vibration sensor's time series data into windows of 1024 sampling points and calculates the Shannon entropy for each window. Historical data shows that the entropy value is approximately 0.2-0.5 during normal cutting, while it surges to over 1.8 when the tool chipps. The system automatically calculates the shunting threshold using the elbow rule. High-entropy windows are marked as critical transient windows and lossless compression is performed using differential coding; low-entropy windows are marked as stationary normal windows and lossy compression using DCT and nonlinear quantization (compression ratio approximately 20:1). Hybrid storage data frames are written to NVMe solid-state drives.

[0141] The equipment health status assessment and adaptive prediction module trains a support vector regression model based on historical feature sequences to predict remaining lifespan in real time. When the model's prediction error exceeds 10% for three consecutive times, the online learning unit automatically triggers retraining, loading data from the last 30 days to update the model parameters. The system outputs the current machine tool health score. (1 point) Predict remaining lifespan Hourly probability of failure in the next 7 days .

[0142] Based on the above results and spare parts inventory (2 spindle bearings in stock, with a lead time of 5 days), the full lifecycle business decision support module generates a maintenance plan: It recommends preventative maintenance after 280 hours, replacing the spindle bearing; simultaneously, it triggers a spare parts procurement suggestion, increasing the safety stock by 1 piece. The scrap assessment module calculates the economic lifespan. The current machine tool has been in operation for 8 years, with a predicted remaining lifespan of 320 hours, an annual maintenance cost of approximately 42,000 yuan, and a new equipment purchase cost of 600,000 yuan. It recommends continuing operation for another 2 years before assessing scrapping.

[0143] The dynamic data archiving and migration module monitors data access activity: In the past 30 days, the simplified background data segment of this machine tool was previewed 120 times and deeply accessed only 3 times, with a popularity value of [missing value]. The machine tool's health level was below the cold archiving threshold of 0.2, and its remaining lifespan was 320 hours, which was greater than 100 hours (the migration prohibition threshold). Therefore, the background data segment was migrated from the solid-state drive (SSD) to the hard disk drive (HDD) for archiving, while the high-fidelity data segment and metadata index remained on the SSD. Three months later, the machine tool's health level dropped to 0.28, with a remaining lifespan of 50 hours. The system automatically prohibited the migration operation to ensure that all data remained accessible at high speed, providing a guarantee for emergency fault diagnosis.

[0144] The human-machine interaction module displays the machine tool's entire lifecycle trajectory, health trend curve, remaining life prediction, spare parts inventory warning, and storage resource usage heat map on a command screen. Equipment managers can adjust prediction model parameters and maintenance plan thresholds through the interface.

[0145] The system in this embodiment realizes data-driven intelligent management of the entire life cycle of equipment, effectively solving the problems proposed in the background technology, such as the lack of differentiation of data value density, lack of model adaptation, insufficient support for business decision-making, poor fusion of multi-source heterogeneous data, and disconnect between cold and hot migration and health status.

[0146] Embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the system described above.

[0147] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent management system for the entire lifecycle of equipment based on industrial big data, characterized in that: include: The multi-source heterogeneous data acquisition and fusion module is used to acquire multi-source heterogeneous data throughout the entire life cycle of the device, and to perform feature extraction and cross-modal fusion on the multi-source heterogeneous data to generate a unified feature vector for the entire life cycle of the device. The data value perception and storage module is connected to the multi-source heterogeneous data acquisition and fusion module. It is used to perform an importance score based on information entropy on the time series of real-time sensing data, and divide the data window into a critical transient window and a stable normal window according to the score. The critical transient window is stored using lossless differential encoding, and the stable normal window is stored using lossy compression, thus constructing a hybrid storage structure. The equipment health status assessment and adaptive prediction module is connected to the data value perception and storage module. It is used to construct an equipment health assessment model based on the fused equipment life cycle feature vector, and to adaptively update the prediction model using an online learning mechanism, outputting equipment health score, remaining life prediction and potential failure probability. The full lifecycle business decision support module is connected to the equipment health status assessment and adaptive prediction module. It is used to generate a full lifecycle business decision scheme, including maintenance plan, spare parts procurement plan, overhaul plan and scrap assessment, based on the equipment health score, remaining life prediction and potential failure probability, combined with the constraints of spare parts inventory, maintenance resources and production plan. The dynamic data archiving and migration module is connected to the data value perception storage module and the equipment health status assessment and adaptive prediction module. It is used to dynamically adjust the storage level of data segments in the hybrid storage structure according to the data access popularity, equipment health status and predicted maintenance needs, so as to realize hierarchical archiving of hot and cold data.

2. The intelligent equipment lifecycle management system based on industrial big data according to claim 1, characterized in that, The multi-source heterogeneous data includes structured data, semi-structured data, unstructured data, and real-time sensor data. The multi-source heterogeneous data acquisition and fusion module includes: The structured data extraction unit is used to collect structured data of the equipment, including ledgers, purchase contracts, maintenance records, and spare parts inventory. A semi-structured data parsing unit is used to parse semi-structured data of the equipment, including operation logs and maintenance work orders; An unstructured data processing unit is used to extract semantic feature vectors from unstructured data using a pre-trained BERT model. The unstructured data includes fault description text, maintenance reports, and equipment manuals. A real-time sensor data access unit is used to access real-time sensor data, including vibration, temperature, current, and pressure. The cross-modal fusion unit is used to perform weighted fusion of the multi-source heterogeneous data using a multi-head attention mechanism to generate a unified device lifecycle feature vector.

3. The intelligent equipment lifecycle management system based on industrial big data according to claim 1, characterized in that, The data value perception storage module includes: The time window segmentation unit is used to segment the real-time sensor data sequence into time windows of a preset length. Importance scoring unit, used to calculate the Shannon entropy of the data sequence within each time window; The flow decision unit is used to compare the Shannon entropy with the adaptive shunting threshold. If the Shannon entropy is greater than the adaptive shunting threshold, it is marked as a critical transient window; otherwise, it is marked as a stationary normal window. The hybrid coding unit is used to perform lossless compression of critical transient windows using differential pulse code modulation, and lossy compression of stationary normal windows using discrete cosine transform and non-uniform quantization, to generate hybrid storage data frames.

4. The intelligent equipment lifecycle management system based on industrial big data according to claim 3, characterized in that, The adaptive shunting threshold is dynamically determined using the elbow rule, specifically: The temporal importance score sequence is sorted in descending order, the vertical distance from each point to the line connecting the start and end points is calculated, and the score value corresponding to the maximum distance is taken as the adaptive diversion threshold.

5. The intelligent equipment lifecycle management system based on industrial big data according to claim 1, characterized in that, The equipment health status assessment and adaptive prediction module includes: The health assessment unit is used to construct a device health scoring function based on feature vectors throughout the entire life cycle. The remaining lifetime prediction unit is used to predict the remaining lifetime using a support vector regression or random forest regression model, with historical feature sequences as input. The failure probability prediction unit is used to output the failure probability of the device within a future time window using a random survival forest model. The online learning update unit is used to periodically collect new data, calculate prediction errors, and trigger model retraining when the error exceeds a preset threshold, thereby achieving adaptive updates of the model.

6. The intelligent equipment lifecycle management system based on industrial big data according to claim 1, characterized in that, The full lifecycle business decision support module includes: The maintenance plan optimization unit is used to generate a preventive maintenance plan based on the equipment health score and remaining life, combined with maintenance resource constraints, using a multi-objective optimization function. The spare parts procurement optimization unit is used to determine the optimal spare parts inventory level based on the failure probability and spare parts lead time using dynamic programming. The end-of-life assessment unit is used to calculate the economic life of equipment by taking into account the remaining life of the equipment, maintenance costs, residual value, and the cost of purchasing new equipment, and to generate end-of-life recommendations.

7. The intelligent equipment lifecycle management system based on industrial big data according to claim 1, characterized in that, The dynamic data archiving and migration module includes: The access popularity monitoring unit is used to count the query frequency and access latency of each hybrid storage data frame and calculate the popularity value; The health status association unit is used to obtain the current device health score and remaining lifespan, and preset the health threshold and remaining lifespan threshold. If the remaining lifespan is less than the remaining lifespan threshold or the device health score is less than the health threshold, the corresponding data segment is prohibited from being migrated to cold storage. The hierarchical migration unit is used to preset the cold storage heat threshold and migrate the background simplified data segments that meet the requirements of having a heat value less than the cold storage heat threshold and being in normal health status from the first-level storage medium to the second-level storage medium, while updating the metadata index.

8. The intelligent equipment lifecycle management system based on industrial big data according to claim 1, characterized in that, The system also includes a human-computer interaction module, which is connected to each module and is used to display the device's full life cycle status, health trend, prediction results, business decision-making schemes and storage resource usage in a visual manner, and supports users to adjust decision parameters online.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the intelligent equipment lifecycle management system based on industrial big data as described in any one of claims 1-8.

Images