Equipment fault prediction and early warning method and system for data mining of Internet of Things

Abstract

The invention relates to the technical field of Internet of Things data mining, and discloses an equipment fault prediction and early warning method and system for Internet of Things data mining, and the method comprises the steps: constructing an equipment association map through equipment communication behavior analysis, extracting a multi-dimensional feature matrix through multi-scale decomposition and self-adaptive calibration, and obtaining an equipment association map; a graph convolutional network is utilized to fuse associated information between devices to obtain deep feature representation, a meta-learning framework and a virtual sample generation method are introduced to solve the problem of small sample fault recognition, an online learning and active learning mechanism is established to realize continuous optimization of a model, a multi-task learning framework is constructed to collaboratively optimize health degree evaluation and fault prediction, and fault prediction is realized. And adopting reinforcement learning to optimize an early warning strategy and generating intelligent disposal suggestions based on the knowledge graph. According to the method, high-precision fault identification can be realized under the condition that fault samples are scarce, long-term stable prediction performance is maintained through model adaptive evolution, and a personalized early warning decision and an executable maintenance scheme are provided.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) data mining technology, and more specifically, to a method and system for predicting and warning of device failures through IoT data mining. Background Technology

[0002] With the development of Industrial Internet of Things (IIoT) technology, a large number of industrial devices are connected to the network, generating massive amounts of multimodal sensor data. Equipment fault prediction and early warning have become key technologies for ensuring production continuity and reducing maintenance costs. Existing equipment fault prediction methods are mainly based on data analysis of single devices, identifying faults by extracting the device's own operating characteristics. However, in actual industrial scenarios, there are process correlations and physical proximity relationships between devices. The failure of a single device may affect the operating status of related devices. Single-device analysis methods cannot capture the associated fault modes and fault propagation paths between devices.

[0003] Existing methods face difficulties in dealing with the scarcity of fault samples. Industrial equipment has a low failure frequency, and samples of certain rare fault types are extremely scarce. Traditional supervised learning methods are prone to overfitting under small sample conditions, resulting in low accuracy in identifying rare faults. Although some methods attempt to expand the sample using data augmentation techniques, the generated virtual samples often lack physical mechanism constraints, deviate from the real fault characteristics, and have limited effectiveness in model training.

[0004] Furthermore, existing fault prediction models typically employ static parameters after deployment, failing to adapt to dynamic changes in equipment status and fault modes. As equipment is used and maintained, its degradation characteristics and fault modes evolve, gradually diminishing the predictive performance of static models. Existing early warning methods often use fixed thresholds to trigger warnings, failing to comprehensively consider factors such as equipment importance, the urgency of production tasks, and the availability of maintenance resources. This results in high false alarm or false negative rates, impacting the practicality of early warning systems. Summary of the Invention

[0005] This invention provides a method and system for device fault prediction and early warning through IoT data mining, which solves the technical problems in related technologies, such as the inability of single device analysis to capture the associated fault modes between devices, the low accuracy of rare fault identification under small sample conditions, the inability of models to adapt to dynamic changes in device status, and the lack of comprehensive consideration in early warning decisions, resulting in high false alarm or false alarm rates.

[0006] This invention provides a method for device fault prediction and early warning through IoT data mining, comprising the following steps: S1, based on dynamic topology discovery and multi-source data standardization access through device communication behavior analysis, obtains device association map and multimodal data stream; S2, acquire device association map and multimodal data stream, and use sensor data feature extraction and quality enhancement methods to obtain multidimensional feature matrix and data quality assessment results; S3: Obtain the device association map, multi-dimensional feature matrix and data quality assessment results, and use the association feature learning and incremental feature update method to obtain deep feature representation; S4. Obtain deep feature representations and use small sample fault identification and virtual sample generation methods to obtain a fault identification model; S5. Based on the fault identification model and deep feature representation, an adaptive evolution fault prediction model is obtained by using continuous model optimization and performance monitoring methods. S6. Based on the deep feature representation and adaptive evolution fault prediction model, a health assessment and fault prediction collaborative optimization method is adopted to obtain the quantitative results of equipment health status and the prediction of remaining life. S7 acquires quantitative results of equipment health status and prediction of remaining lifespan, and uses an intelligent handling suggestion generation method to obtain early warning decisions and executable maintenance plans.

[0007] In a preferred embodiment, S1 includes: Real-time communication data streams of each device in the device group are collected, and the communication source address, destination address, interaction frequency and data flow direction features are extracted to construct a device communication behavior matrix. The association strength value between devices is calculated based on the communication behavior matrix. When the association strength value exceeds the preset association threshold, a connection edge is established in the device association graph. Establish a device metadata registry and develop data parsing adapter modules for different communication protocols to convert the parsed data into a unified internal format.

[0008] In a preferred embodiment, S2 includes: Wavelet transform is used to decompose the sensor time-series data in the multimodal data stream into multiple layers to obtain approximation coefficients and detail coefficients at different scales. Construct a sliding time window, calculate statistical characteristics within each time window, and perform a fast Fourier transform on the data to obtain the frequency domain amplitude spectrum; A sensor baseline model is established. When the deviation between the measured value and the baseline exceeds the preset drift threshold and lasts for a preset time, the sensor is determined to have drifted and adaptive calibration is performed.

[0009] In a preferred embodiment, S3 includes: Each device is treated as a node in the graph, and the feature vector of the node is composed of a multi-dimensional feature matrix. The associated edges in the device association graph are used as the edges of the graph. The design uses a graph convolutional network to learn the associated features of device groups. A neighbor sampling strategy is adopted to randomly sample neighbor nodes of the target node for feature aggregation. An attention mechanism is introduced to calculate the attention score of the neighbor nodes and aggregate the features of the neighbor nodes in a weighted manner. A time attention module is introduced to calculate the time attention scores of the current moment and each historical moment, and to aggregate the features of each historical moment in a weighted manner.

[0010] In a preferred embodiment, S4 includes: A meta-learning framework is introduced to train the fault identification model. The identification tasks of different equipment types, different operating conditions, and different fault types are used as the training task set. For each training task, a support set and a query set are constructed. Design a prototype network to compute the prototype of each category for samples in the support set, compute the distance between the feature vector of the sample to be predicted in the query set and the prototype of each category, and classify the query sample into the category corresponding to the nearest prototype. A metric learning method is introduced, and a triplet loss function is used for training. The loss function makes the distance between the anchor point and the positive sample less than the distance between the anchor point and the negative sample plus a preset boundary value. A memory enhancement network is established, which stores the feature representations of historical fault samples through an external memory module. When performing fault prediction, the feature vector of the current sample is used as the query, and the information of the retrieved historical samples is weighted and aggregated as auxiliary information. A virtual fault sample generator is constructed. A generative adversarial network is used to learn the distribution characteristics of fault samples. The generator generates feature vectors of virtual samples, and the discriminator outputs the probability that a sample is a real sample or a virtual sample.

[0011] In a preferred embodiment, S5 includes: An online learning mechanism is established. When a device malfunctions and is labeled, the malfunction sample is used as new training data. An incremental learning algorithm is used to update the model parameters. An elastic weight consolidation method is used to add a regularization term to the loss function to limit the variation of important parameters. The design of the sample importance assessment mechanism comprehensively evaluates the information content of the sample by considering its uncertainty, representativeness, and novelty, and calculates the importance score of the sample; a forgetting mechanism is introduced to assign a time decay weight to each historical sample. An active learning strategy is adopted, and when the uncertainty of a predicted sample exceeds a preset threshold, the sample is submitted to the operation and maintenance personnel for annotation. Establish a model performance monitoring mechanism to track the model's performance metrics in each fault category in real time, and trigger model retraining or parameter adjustment when the performance metrics fall below the threshold.

[0012] In a preferred embodiment, S6 includes: A multi-task learning framework is constructed, and a shared feature extraction network is designed to take deep feature representations as input. The output of the shared network is connected to the fault classification branch and the health regression branch. Design a joint loss function that includes fault classification loss, health regression loss, and consistency constraint loss; establish a multi-dimensional evaluation system for equipment health, including performance, reliability, safety, and economic dimensions, and use a weight learning method to automatically determine the fusion weights of each dimension index; A health degradation trend model was established, and the degradation curve was fitted by regression method and extrapolated to predict the health value at future time points; the Weibull distribution model was used to predict the remaining service life of the equipment.

[0013] In a preferred embodiment, the early warning decision in S7 includes: The optimization of the early warning strategy is modeled as a reinforcement learning problem. The state space is defined to include the fault prediction probability, health score and degradation rate, equipment importance level, urgency of the current production task, and availability of maintenance resources. The action space is defined to include not issuing an early warning, issuing a warning of concern level, issuing a warning level warning, and issuing an emergency level warning. The reward function is designed to comprehensively consider the accuracy, timeliness, and economy of the early warning. A deep Q-network algorithm is used to train the early warning decision agent. The neural network takes the state vector as input and outputs the Q value of each action.

[0014] In a preferred embodiment, the maintenance scheme executable in S7 includes: Construct a knowledge graph of equipment failures, where nodes represent knowledge entities including equipment type entities, failure type entities, failure cause entities, handling measures entities, and spare parts entities, and edges represent the relationships between entities; The design incorporates a knowledge graph-based reasoning mechanism. Warning information is used as the input to the knowledge graph, and path search is performed within the knowledge graph to find related response measures. The found response measures are scored and ranked based on historical success rate, implementation difficulty, required time, and required cost. A case-based reasoning method is introduced to establish a historical failure case database. The similarity between the current failure and historical cases is calculated, and the response plan is extracted from the historical case with the highest similarity. The candidate solutions are comprehensively evaluated, and a weighted summation method is used to calculate the comprehensive score of each solution. The solution with the highest comprehensive score is selected as the recommended main solution.

[0015] In a preferred embodiment, an IoT data mining device fault prediction and early warning system is used to perform the steps of the above-described IoT data mining device fault prediction and early warning method, including: The dynamic topology discovery and data access module is used for dynamic topology discovery and standardized access of multi-source data based on device communication behavior analysis, to obtain device association maps and multimodal data streams; The feature extraction and quality enhancement module is used to acquire device association maps and multimodal data streams. It employs sensor data feature extraction and quality enhancement methods to obtain multidimensional feature matrices and data quality assessment results. The association feature learning module is used to obtain device association maps, multi-dimensional feature matrices and data quality assessment results. It uses association feature learning and incremental feature update methods to obtain deep feature representations. The fault identification module is used to obtain deep feature representations and adopts a small sample fault identification and virtual sample generation method to obtain a fault identification model. The model optimization module is used to obtain an adaptive evolution fault prediction model based on the fault identification model and deep feature representation, using continuous model optimization and performance monitoring methods. The health assessment module is used to obtain the quantitative results of equipment health status and the prediction of remaining life based on the deep feature representation and the adaptive evolution of the fault prediction model, using a collaborative optimization method of health assessment and fault prediction. The early warning and decision-making module is used to obtain quantitative results of equipment health status and prediction of remaining lifespan. It adopts an intelligent handling suggestion generation method to obtain early warning decisions and executable maintenance plans.

[0016] The beneficial effects of this invention are as follows: By fusing inter-device correlation information through graph convolutional networks, it is possible to identify cross-device related faults and fault propagation paths. Combined with meta-learning framework and virtual sample generation method, it can achieve high-precision identification of rare faults when fault samples are scarce, thereby improving the accuracy and recall of fault prediction. Continuous model optimization is achieved through online and active learning mechanisms, enabling the model to adapt to dynamic changes in equipment status and fault modes. Reinforcement learning is used to optimize early warning strategies, comprehensively considering equipment status, business needs, and resource availability to balance the accuracy, timeliness, and economy of early warnings, reduce false alarm and false negative rates, and generate executable maintenance plans based on knowledge graphs and case reasoning to improve the efficiency of fault handling. Attached Figure Description

[0017] Figure 1 This is a flowchart of a device fault prediction and early warning method for Internet of Things data mining according to the present invention; Figure 2 This is a block diagram of an IoT data mining device fault prediction and early warning system according to the present invention; Figure 3 This is a flowchart of a device fault prediction and early warning method for Internet of Things data mining according to the present invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a device fault prediction and early warning method using Internet of Things (IoT) data mining, such as... Figures 1 to 3 As shown, it includes the following steps: S1, based on dynamic topology discovery and multi-source data standardization access through device communication behavior analysis, obtains device association map and multimodal data stream; S11 collects multimodal sensor data and communication data from each device in the equipment group. For the CNC machine tool equipment group, the collected sensor data includes: temperature data, including spindle temperature, bearing temperature, coolant temperature, ambient temperature, etc.; vibration data, including spindle vibration, feed axis vibration, machine tool base vibration, etc., using accelerometers to measure three-axis vibration signals; current data, including spindle motor current, feed axis motor current, coolant pump current, etc.; position data, including the actual position of each axis, encoder feedback position, etc.; pressure data, including hydraulic system pressure, pneumatic system pressure, etc.; sound data, collected through acoustic sensors to collect equipment operating noise; tool status data, including tool wear, tool vibration, cutting force, etc. Simultaneously, it collects equipment operating status data, including equipment power-on / off status, working mode, machining program number, spindle speed setpoint, feed rate setpoint, etc. It also collects communication data between devices, including data exchange messages, control commands, and status feedback. All data is timestamped to record the precise moment of data acquisition.

[0020] S12: Collect real-time communication data streams and device identification information from each device in the device group. By monitoring data interaction messages between devices, extract communication source addresses, destination addresses, message types, interaction frequencies, and data flow characteristics to construct a device communication behavior matrix. Based on the communication behavior matrix, calculate the communication frequency and data traffic between any two devices. Use a normalization method to map the communication frequency and data traffic to the interval between 0 and 1 to obtain the association strength value between devices. When the association strength value exceeds a preset association threshold, it is determined that there is a process association or physical proximity relationship between the two devices, and a connection edge is established in the device association graph. By continuously monitoring the communication data stream, real-time detection of device access and removal events is achieved. When a new device is detected to be accessed, it is automatically added to the device association graph and the association strength with other devices is calculated. When a device is detected to be removed, the corresponding node and its associated edge are deleted from the graph, realizing dynamic updates of the device topology.

[0021] S13. Establish a device metadata registry to record the device type, manufacturer information, sensor configuration list, data communication protocol, sampling frequency parameters, and data format specifications for each device. For devices from different manufacturers and using different communication protocols, develop corresponding data parsing adapter modules. Each adapter module is responsible for parsing data packets of a specific protocol and extracting sensor measurement values ​​and timestamp information. Design a unified internal data format standard, including device identification fields, sensor type fields, measurement value fields, timestamp fields, and data quality identifier fields. Each adapter module converts the parsed data into the unified internal format and adds a data quality identifier, which is used to mark whether the data is abnormal, missing, or if the sensor is faulty.

[0022] S14 employs a semantic mapping-based approach to process similar sensor data from different devices. A sensor semantic dictionary is established, defining standard sensor type names and measurement units. For example, spindle temperature sensors from different devices are uniformly mapped to the standard semantic type "spindle temperature." For sensor data with inconsistent measurement units, unit conversion is performed, such as converting Celsius to Kelvin temperature or minutes to radians per second. Through semantic mapping and unit conversion, heterogeneous sensor data from different devices are mapped to a unified semantic space, facilitating subsequent data fusion and feature extraction.

[0023] S15 implements a real-time buffering and batch processing mechanism for the data stream. For sensor data with high sampling frequencies, such as vibration sensors sampling at the kilohertz level, an online aggregation method is used to calculate the statistical characteristic values ​​of the data within a fixed time window, including the mean, maximum, minimum, and standard deviation. The aggregated characteristic values ​​are then used as representative data for that time window. For sensor data with low sampling frequencies, such as temperature sensors sampling at the second or minute level, a buffering mechanism is used to temporarily store the data in a buffer, waiting for the timestamps of other sensor data to be aligned before processing. Through real-time buffering and batch processing mechanisms, the real-time performance and integrity of data processing are balanced, ensuring that sensor data with different sampling frequencies can be processed collaboratively.

[0024] Through the above sub-steps, a real-time updated device association map and a unified format multimodal data stream are obtained. The device association map reflects the dynamic topology of the device group and the relationships between devices, while the multimodal data stream contains various sensor measurement data from each device.

[0025] S2, receive the device association map and multimodal data stream, and use sensor data feature extraction and quality enhancement methods to obtain multidimensional feature matrix and data quality assessment results; S21, perform multi-scale decomposition on the time-series data of each sensor in the multimodal data stream obtained in step S1. Using wavelet transform, select appropriate wavelet basis functions, such as Daubechies wavelet or Morlet wavelet, to perform multi-level decomposition of the time-series data, obtaining approximation coefficients and detail coefficients at different scales. Approximation coefficients reflect the low-frequency trend components of the data, while detail coefficients reflect the high-frequency fluctuation components. Coefficients at different scales correspond to different operating states and fault modes of the equipment; for example, low-frequency components reflect the performance degradation trend of the equipment, while high-frequency components reflect vibration anomalies and impact events. Calculate the energy characteristic for the coefficients at each scale; the energy characteristic is defined as the sum of squares of the coefficients, reflecting the intensity of the component at that scale.

[0026] S22 extracts statistical features from sensor data in the time domain. A sliding time window is constructed, with the window size adaptively determined based on the device's operating cycle. For example, for rotating equipment, the window size is set to the length of several rotation cycles. Within each time window, statistical features such as mean, variance, skewness, kurtosis, crest factor, impulse factor, and margin factor are calculated. The mean reflects the central tendency of the data, variance reflects the dispersion of the data, skewness reflects the symmetry of the data distribution, and kurtosis reflects the sharpness of the data distribution. The crest factor is defined as the ratio of the peak value to the root mean square value, the impulse factor is defined as the ratio of the peak value to the absolute mean value, and the margin factor is defined as the ratio of the peak value to the root mean square amplitude. These factors are sensitive to shocks and abnormal vibrations of the equipment.

[0027] S23. Extract the spectral features of the sensor data in the frequency domain. Perform a Fast Fourier Transform on the data within each time window to obtain the frequency domain amplitude spectrum. Identify the peak frequency components in the amplitude spectrum; the peak frequencies correspond to the characteristic frequencies of the device, such as rotational frequency, gear meshing frequency, and bearing fault characteristic frequencies. Calculate the amplitude and phase of each characteristic frequency; the amplitude reflects the intensity of the frequency component, and the phase reflects its temporal relationship. Calculate global frequency domain features such as the centroid frequency, mean square frequency, and spectral entropy; the centroid frequency reflects the concentration of spectral energy, the mean square frequency reflects the dispersion of the spectrum, and the spectral entropy reflects the complexity and uncertainty of the spectrum.

[0028] S24. Establish a sensor baseline model for detecting and calibrating sensor drift. Sensor drift refers to a systematic shift in sensor measurements over time, leading to inaccurate measurements. A moving average method is used to calculate the long-term trend of the sensor output, which is then used as the sensor baseline. The deviation between the current measurement and the baseline is calculated. When the deviation exceeds a preset drift threshold and persists for a certain period, sensor drift is determined. For drifting sensors, an adaptive calibration method is used to correct subsequent measurements based on the deviation. The corrected measurement is equal to the original measurement minus the deviation. The sensor's drift history is recorded. When frequent drift or excessive drift occurs, a sensor maintenance reminder is generated, suggesting inspection or replacement of the sensor.

[0029] S25 employs a statistical process control-based anomaly detection method to identify outliers in the data. For each extracted feature, the mean and standard deviation are calculated based on historical normal operation data, establishing upper and lower control limits. The upper control limit equals the mean plus three times the standard deviation, and the lower control limit equals the mean minus three times the standard deviation. When a feature value exceeds these limits, it is marked as an outlier. Cluster analysis is performed on the detected outliers, using density clustering to group outliers that are close in time into anomalous events. The duration and impact of these anomalous events are analyzed; short-lived, isolated outliers may be noise interference, while persistent anomalous events may indicate genuine equipment malfunctions. These genuine anomalous events are then marked as important features for subsequent fault prediction.

[0030] Through the above sub-steps, a multi-dimensional feature matrix and data quality assessment results are obtained. The multi-dimensional feature matrix includes various types of features such as time-domain features, frequency-domain features, and multi-scale features, comprehensively depicting the operating status of the equipment. The data quality assessment results include sensor drift indicators, outlier indicators, and missing data indicators.

[0031] In some embodiments, since the feature distribution varies significantly across different equipment types and operating conditions, an adaptive feature extraction method based on operating condition identification can be employed to improve the targeting and effectiveness of feature extraction. Specifically, the operating conditions of the equipment are automatically identified, and the equipment's operating data is divided into different operating condition categories using clustering methods, such as no-load, light-load, heavy-load, acceleration, and deceleration conditions. For each operating condition category, feature extraction parameters and anomaly detection thresholds are established; for example, the normal range of vibration amplitude under heavy-load conditions is higher than that under light-load conditions. During real-time feature extraction, the current operating condition category is identified, and then the parameters corresponding to the operating condition are used for feature extraction and anomaly detection, achieving adaptive feature extraction based on operating conditions.

[0032] S3 receives the device association map, multi-dimensional feature matrix and data quality assessment results, and uses the association feature learning and incremental feature update method to obtain deep feature representation; S31. Based on the device association graph obtained in step S1, the multi-dimensional feature matrix obtained in step S2, and the data quality assessment results, a graph structure representation of the device group is constructed. Each device is treated as a node in the graph, and the feature vector of the node is composed of the multi-dimensional features of that device. According to the data quality assessment results obtained in step S2, features containing anomaly markers or sensor drift markers are labeled and given special attention in subsequent feature learning, as these abnormal features are often related to faults. The association edges in the device association graph are treated as edges of the graph, and the weight of the edge is determined by the association strength value calculated in step S1. For large-scale device groups with a large number of devices, a hierarchical graph representation method is used to reduce computational complexity. Based on the physical location, process flow, or functional grouping of the devices, the device group is divided into multiple subgraphs, each containing several closely related devices. Dense connections are established within the subgraphs, and sparse connections are established between subgraphs by selecting representative nodes. The representative node can be the node with the highest connectivity or the node with the most critical function in the subgraph.

[0033] S32, a lightweight graph convolutional network is designed for device cluster association feature learning. The basic operation of a graph convolutional network is the aggregation and updating of node features. For a target node, the feature information of its neighboring nodes is aggregated, and then transformed by combining it with its own features to obtain the updated feature representation. To reduce computational complexity, a neighbor sampling strategy is adopted. For each target node, instead of aggregating all neighboring nodes, a fixed number of neighboring nodes are randomly sampled for aggregation. The number of samples is adaptively determined based on the size of the device cluster and computational resources, minimizing the number of samples while ensuring the feature learning effect. A multi-layer graph convolutional structure is designed. The first layer aggregates the features of one-hop neighbors, and the second layer aggregates the features of two-hop neighbors. By stacking multiple layers, the receptive field is expanded, enabling each node to obtain information from neighbors at greater distances.

[0034] S33 introduces an attention mechanism into the graph convolutional network, enabling the model to automatically learn the importance weights of different neighboring nodes. For the target node and its sampled neighboring nodes, attention scores are calculated, reflecting the degree of influence of neighboring nodes on the target node. The attention scores are calculated using a learnable attention function, taking the target node features and neighboring node features as input, and then processing them through neural network layers. The attention scores of all neighboring nodes are normalized, and a softmax function is used to convert the attention scores into weight coefficients, with the sum of these weight coefficients equal to 1. During feature aggregation, the features of neighboring nodes are multiplied by their corresponding weight coefficients and then summed to obtain a weighted aggregated feature representation. Through this attention mechanism, the model can automatically identify the most valuable associated devices for fault prediction, enhancing the targeting of feature learning.

[0035] S34 introduces a time attention module to perform weighted aggregation of features from historical time windows. The device's historical operational data is divided into multiple time windows, each corresponding to a feature vector at a given moment. For fault prediction at the current moment, not only the features of the current moment but also features from historical moments are needed to capture the fault's evolution. Time attention scores are calculated for the current moment and each historical moment, reflecting the importance of historical moments to the current prediction. Generally, historical moments closer to the current moment are more important, but for some progressive faults, feature changes from earlier moments may be more predictive. A learnable time attention function is used to automatically learn time weights. The features of each historical moment are multiplied by the time weights and then summed to obtain a time-weighted feature representation.

[0036] S35 implements an incremental feature update mechanism to improve the efficiency of feature computation. A feature cache is established to store the feature representation of each device within a historical time window. When performing feature computation at a new time, it checks whether the device topology has changed. If the topology has not changed and the features of the device's neighboring nodes have not changed significantly, the feature representation from the previous time step is directly read from the cache, and only the newly added features at the current time step are incrementally updated. The incremental update uses a sliding window approach, removing features from the earliest time step, adding features from the current time step, and recalculating the time-weighted aggregation. If the topology has changed or the features of neighboring nodes have changed significantly, the complete graph convolution computation is re-executed. Incremental updates avoid redundant computation of unchanging features, reducing computational overhead.

[0037] S36 employs model compression techniques to reduce the parameter size and computational complexity of graph convolutional networks (GCNNs). The trained GCNN is pruned by identifying and removing redundant connections with weights close to zero, resulting in a sparse network structure. The pruned network is then fine-tuned to recover from the performance degradation caused by pruning. Quantization techniques are used to convert network parameters from floating-point representations to low-bit integer representations, such as converting 32-bit floating-point quantization to 8-bit integers, reducing the model's storage space and computational cost. While maintaining relatively consistent prediction accuracy, pruning and quantization compress the model size to a fraction of its original size, enabling deployment on resource-constrained edge computing devices.

[0038] Through the above sub-steps, a deep feature representation integrating inter-device correlation information is obtained. This feature representation not only includes the operating status information of a single device but also integrates the status information of related devices and the interaction patterns between devices, enabling the identification of cross-device associated faults and fault propagation paths. Lightweight design and incremental update mechanism ensure the real-time performance and scalability of feature computation.

[0039] In some embodiments, since the importance and impact of failures vary among different devices in a group of devices, a graph convolution method based on device importance weighting can be used. The aim is to make the model focus more on learning the features of key devices. Specifically, each device is assigned an importance weight based on factors such as its position in the production process, its value, and the impact of its failure on production. During the feature aggregation process of graph convolution, the features of neighboring nodes are multiplied not only by the attention weight but also by the device importance weight, giving the features of key devices a greater weight in the aggregation. In the loss function of the model training, a higher penalty weight is assigned to the prediction error of key devices, prompting the model to prioritize the prediction accuracy of key devices.

[0040] S4. Obtain deep feature representations and use small sample fault identification and virtual sample generation methods to obtain a fault identification model; S41 addresses the class imbalance problem in equipment fault data, where normal samples far outnumber fault samples, and the problem of extremely scarce samples for certain rare fault types. A meta-learning framework is introduced to train the fault identification model. The core idea of ​​meta-learning is learning how to learn; by training on multiple related tasks, the model gains the ability to quickly adapt to new tasks. Identification tasks for different equipment types, operating conditions, and fault types are used as the training task set for meta-learning. For each training task, a support set and a query set are constructed. The support set contains a small number of labeled samples for model adaptation, and the query set contains samples to be predicted for evaluating model performance.

[0041] S42, design the prototype network as the foundational model for meta-learning. The core of the prototype network is to learn a prototype representation for each category in the feature space. The prototype representation is the center point of all sample features for that category. For samples in the support set, the deep feature representation obtained in step S3, which integrates inter-device association information, is used as input to calculate the prototype for each category. The prototype is equal to the average of the feature vectors of all samples in that category. For the samples to be predicted in the query set, the deep feature representation obtained in step S3 is used to calculate the distance between its feature vector and the prototypes of each category, using Euclidean distance or cosine distance as the distance metric. The query sample is classified into the category corresponding to the nearest prototype, and the negative value of the distance is converted into a category probability distribution using a softmax function.

[0042] S43 introduces a metric learning method to learn a distance metric function that can effectively distinguish different fault categories. Traditional Euclidean or cosine distance are fixed metrics and may not be suitable for the characteristics of fault identification tasks. A learnable metric function is adopted, and a neural network maps feature vectors to a metric space, calculating the distance between samples in the metric space. The training objective of the metric network is to make samples of the same class close in distance and samples of different classes far apart in the metric space. A triplet loss function is used for training. A triplet contains an anchor sample, a positive sample, and a negative sample. The anchor sample belongs to the same class as the positive sample, and the anchor sample belongs to a different class as the negative sample. The loss function ensures that the distance between the anchor point and the positive sample is less than the distance between the anchor point and the negative sample plus a boundary value.

[0043] S44 introduces a task adaptation mechanism to enable the model to quickly adapt to new equipment types and operating conditions. During meta-learning training, the model is trained on multiple tasks, each corresponding to a specific equipment type or operating condition. The model learns cross-task general knowledge and rapid adaptability. When faced with new equipment types or operating conditions, only a small number of labeled samples are needed as a support set for the model to quickly adjust its parameters to adapt to the new task. Task adaptation is achieved through several steps of gradient updates on the support set, and the updated model parameters are used for prediction on the query set. The training objective of meta-learning is to optimize the model's initial parameters so that it achieves good performance on any new task after a small number of gradient updates.

[0044] S45, a memory-enhanced network is established, storing feature representations of historical fault samples through an external memory module. The memory module is organized in key-value pairs, where the key is the feature vector of the fault sample, and the value is the category label and related information of the fault sample. During fault prediction, the feature vector of the current sample is used as the query to retrieve similar historical samples from the memory module. An attention mechanism is used to calculate the similarity between the current sample and each sample in memory; historical samples with high similarity are assigned higher attention weights. The information from the retrieved historical samples is weighted and aggregated, and used as auxiliary information input into the classifier to enhance the prediction of the current sample. The memory module is updated using a first-in, first-out (FIFO) strategy; when the memory capacity reaches its limit, the oldest stored sample is removed, and a new sample is added. For samples that are mispredicted or difficult to classify, their retention priority in the memory module is increased, enabling the model to learn from its mistakes.

[0045] S46. A virtual fault sample generator is constructed using a generative adversarial network (GAN) to learn the distribution characteristics of fault samples and generate realistic virtual samples. The GAN consists of a generator and a discriminator. The generator aims to generate virtual samples similar to real fault samples, while the discriminator aims to distinguish between real and virtual samples. The generator takes a random noise vector and a fault category label as input and generates feature vectors for virtual samples through a neural network. The discriminator takes the sample feature vector as input and outputs the probability that the sample is a real or virtual sample. The generator and discriminator improve each other through adversarial training. The generator strives to generate virtual samples that the discriminator cannot recognize, while the discriminator strives to improve its ability to recognize virtual samples. After training convergence, the generator can generate high-quality virtual fault samples to expand the training set and alleviate the sample imbalance problem.

[0046] Through the above sub-steps, a high-precision identification model for rare faults is obtained. This model achieves rapid adaptability through meta-learning, utilizes historical fault cases through memory enhancement, and alleviates the sample imbalance problem through virtual sample generation, enabling high-precision fault identification even when fault samples are scarce.

[0047] S5. Based on the fault identification model and deep feature representation, an adaptive evolution fault prediction model is obtained by using continuous model optimization and performance monitoring methods. S51. Establish an online learning mechanism to enable the fault identification model trained in step S4 to continuously learn and optimize from new data after deployment. During model deployment and operation, continuously collect equipment operation data and actual fault occurrences. For newly collected operation data, extract deep feature representations that fuse inter-device correlation information using the graph convolutional network in step S3. When a device malfunctions and is labeled by maintenance personnel, the fault sample and its features are used as new training data. An incremental learning algorithm is used to update the model parameters. Incremental learning does not require retraining the entire model but rather fine-tuning the existing model parameters. Specifically, new samples and some historical samples are combined into a small batch of data. The loss function of the model on this batch of data is calculated, and the model parameters are updated using the gradient descent method. To avoid the model overfitting to new data and forgetting old knowledge, an elastic weight consolidation method is used. A regularization term is added to the loss function to limit the variation of important parameters, which are those that have a significant impact on historical task performance.

[0048] S52, design a sample importance assessment mechanism to evaluate the value of new samples, prioritizing samples with high information content for model updates. The information content of a sample is comprehensively evaluated using multiple indicators, including sample uncertainty, sample representativeness, and sample novelty. Sample uncertainty is measured by the entropy of the probability distribution predicted by the model; a higher entropy value indicates greater uncertainty in the model regarding the sample, and thus greater information content. Sample representativeness is measured by calculating the similarity between the sample and other samples; if a sample has a large number of similar samples around it, its representativeness is high. Sample novelty is measured by comparing the sample with the samples in the training set; if a sample is dissimilar to all samples in the training set, it is novel and may represent a new failure mode. The importance score of the sample is calculated by combining these three indicators, and samples with high importance scores are prioritized for model updates.

[0049] S53 introduces a forgetting mechanism to reduce the weight of outdated or no longer applicable historical samples. As equipment is used and maintained, its state and failure modes may change, and early failure samples may no longer be applicable to the current equipment state. Each historical sample is assigned a time-decay weight, which decays exponentially over time, with samples older than the current time having lower weights. During model training, the sample weights are used as coefficients in the loss function; samples with lower weights have less impact on model updates. For samples that are clearly no longer applicable, such as failure samples from before a major overhaul or component replacement, they are directly removed from the training set or their weights are set to zero. This forgetting mechanism prevents the model from being misled by stale data and maintains the model's adaptability to the current equipment state.

[0050] S54 employs an active learning strategy, enabling the model to proactively identify samples with high uncertainty and request manual annotation. During model operation, the uncertainty of the prediction is calculated for each predicted sample, measured by the entropy of the predicted probability distribution or the variance of predictions from multiple models. When the uncertainty exceeds a preset threshold, the sample is marked as unlabeled and submitted to operations personnel for manual judgment and annotation. The operations personnel determine the true category of the sample based on the actual situation, and the labeled sample is immediately used for incremental model updates. Through active learning, the model can achieve maximum performance improvement with minimal annotation cost, focusing on annotating the most valuable samples for model improvement and avoiding the blind annotation of a large number of low-value samples.

[0051] S55 establishes a model performance monitoring mechanism to track the model's predictive performance metrics across various fault categories in real time. Performance metrics include accuracy, recall, precision, and F1 score. Accuracy reflects the proportion of correct predictions made by the model overall; recall reflects the proportion of true faults identified by the model out of all true faults; precision reflects the proportion of true faults among the samples predicted as faults; and the F1 score is the harmonic mean of recall and precision. Performance metrics are calculated separately for each fault category to identify which categories the model performs poorly in. Performance thresholds are set; when the recall or F1 score for a category falls below the threshold, model retraining or parameter tuning is triggered. Performance monitoring results are displayed in the form of visual charts, including performance metric curves over time and bar charts comparing performance across categories, facilitating understanding of the model's status by operations personnel.

[0052] S56 implements a model version management mechanism that records the time, data used, and performance changes of each model update. A unique version number is assigned to each model version, and the model's parameter files and configuration information are stored. When performance degrades or anomalies occur after a model update, rollback to a previous version is supported. A model comparison testing framework is established, running multiple model versions simultaneously on the same test data to compare the prediction results and performance metrics of each version, selecting the version with the best performance as the current deployment version. The evolution history of the model is recorded, long-term performance trends are analyzed, and key factors affecting model performance are identified, providing guidance for model optimization.

[0053] Through the above sub-steps, an adaptive evolutionary fault prediction model is obtained. This model continuously learns from new data through online learning, efficiently utilizes manually labeled resources through active learning, and promptly identifies and resolves model performance issues through performance monitoring. It can adapt to dynamic changes in equipment status and fault modes, maintaining long-term stable predictive performance.

[0054] S6. Based on the deep feature representation and adaptive evolution fault prediction model, a health assessment and fault prediction collaborative optimization method is adopted to obtain the quantitative results of equipment health status and the prediction of remaining life. S61, construct a multi-task learning framework, jointly training the equipment health assessment task and the fault prediction task as interrelated optimization objectives. Design a shared feature extraction network, taking the equipment group association features obtained in step S3 as input, and extracting higher-level abstract features through a multi-layer neural network. The output of the shared network is connected to two task branches: one branch for fault classification prediction and the other for health regression assessment. The fault classification branch uses fully connected layers and a softmax activation function to output the probability distribution of each fault category. The health regression branch uses fully connected layers and a sigmoid activation function to output a health score, ranging from 0 to 1, where 1 represents complete health and 0 represents complete failure.

[0055] S62, design a joint loss function comprising three parts: fault classification loss, health regression loss, and consistency constraint loss. The fault classification loss uses the cross-entropy loss function to measure the difference between the predicted fault category probability distribution and the true category label. The health regression loss uses the mean squared error loss function to measure the difference between the predicted health score and the true health label. The consistency constraint loss ensures that the outputs of the two tasks are logically consistent; specifically, devices with low health should have a higher failure probability, and devices with high health should have a lower failure probability. The consistency constraint loss is defined as a correlation penalty between the health score and the failure probability; a larger penalty is applied when the health is low but the failure probability is also low, or vice versa. The joint loss function is a weighted sum of the three losses, with the weight coefficients determined through cross-validation.

[0056] S63 establishes a multi-dimensional assessment system for equipment health, including performance, reliability, safety, and economic dimensions. The performance dimension evaluates the equipment's working efficiency and processing accuracy, calculating the degree of performance degradation by comparing the equipment's actual output with standard output; a greater degree of performance degradation indicates lower health. The reliability dimension is calculated based on the output probability of the fault prediction model obtained in step S5 and the equipment's historical fault frequency; equipment with a high fault probability or frequent historical faults has low reliability. The safety dimension assesses whether the equipment's key parameters are close to safety thresholds, such as whether the temperature is close to the overheating threshold or the vibration is close to the danger threshold; equipment close to safety thresholds has low safety. The economic dimension evaluates the equipment's maintenance costs and residual value; equipment with high maintenance costs or low residual value has poor economic performance.

[0057] S64 normalizes the indicators for each dimension, mapping indicators with different dimensions and numerical ranges to a unified 0-1 interval. It employs a minimum-maximum normalization method, where the normalized indicator value equals the original indicator value minus the minimum value, then divided by the difference between the maximum and minimum values. For negative indicators such as performance degradation (i.e., higher values ​​indicate lower health), the normalization process is reversed; the reversed value equals 1 minus the normalized value, making all indicators positive—higher values ​​indicate higher health.

[0058] The S65 employs a deep learning-based weight learning method to automatically determine the fusion weights of each dimension's indicators. A weight learning network is designed, taking the normalized indicators of each dimension as input and learning the weight coefficient for each indicator through neural network layers. The weight coefficients are normalized using a softmax function to ensure that the sum of all weights equals 1. The weight learning network is jointly trained with a health regression branch, optimizing the weight coefficients by minimizing the health prediction error. After training, the weight coefficients reflect the contribution of each dimension's indicators to the health score, avoiding the subjectivity of manually setting weights. The device's overall health score is calculated, which is equal to the sum of the products of the normalized indicators of each dimension and their corresponding weight coefficients.

[0059] S66. Establish a health degradation trend model to predict the health trajectory of equipment over future time periods. Collect historical health data sequences of the equipment and analyze the patterns of health changes over time. For equipment exhibiting a linear degradation trend, use linear regression to fit the degradation curve. For equipment exhibiting a non-linear degradation trend, use exponential smoothing or multinomial regression to fit the degradation curve. Based on the fitted degradation curve, extrapolate and predict the health values ​​at future time points. Calculate the rate of change of health, defined as the difference in health values ​​between adjacent time points divided by the time interval. Identify accelerated health degradation stages. When the absolute value of the rate of change exceeds a preset acceleration threshold, the equipment is determined to have entered an accelerated degradation stage, which is usually a sign of impending failure, and an accelerated degradation warning signal is generated.

[0060] S67 uses the Weibull distribution model to predict the remaining useful life of equipment. The Weibull distribution is a commonly used lifetime distribution model in reliability engineering, describing the time distribution from normal operation to failure of equipment. Based on the equipment's historical failure data and current health status, the shape and scale parameters of the Weibull distribution are estimated. The shape parameter reflects the change in failure rate over time, while the scale parameter reflects the characteristic life of the equipment. Based on the estimated Weibull distribution, the remaining useful life of the equipment under the current health status is calculated. The remaining useful life is defined as the expected time from the current moment to the occurrence of failure. The remaining useful life prediction results are used as auxiliary information and input into the failure prediction model to enhance the model's ability to predict the time of failure.

[0061] Through the above sub-steps, the quantitative results of equipment health status and the prediction of remaining service life are obtained. The multi-dimensional health assessment comprehensively reflects the equipment's health status, the health degradation trend prediction provides information on the evolution of equipment status, and the remaining service life prediction provides a time reference for maintenance planning. The multi-task learning framework enables health assessment and failure prediction to reinforce each other, improving the performance of both tasks.

[0062] In some embodiments, since the degradation process of equipment may be influenced by various factors such as operating load, environmental conditions, and maintenance history, a covariate-based degradation modeling method can be employed to improve the accuracy of degradation trend prediction. Specifically, covariate factors affecting equipment degradation are incorporated into the degradation model. These covariates include cumulative operating time, cumulative workload, ambient temperature and humidity, and historical maintenance frequency. A proportional hazards model or accelerated time to failure model is used to establish the relationship between covariates and degradation rate or remaining lifespan. Regression analysis is used to estimate the influence coefficients of the covariates, which reflect the degree of contribution of each covariate to equipment degradation. When predicting the future health and remaining lifespan of the equipment, the parameters of the degradation model are adjusted according to the expected operating load and environmental conditions to obtain more accurate prediction results.

[0063] S7: Obtain quantitative results of equipment health status and prediction of remaining lifespan, and use intelligent handling suggestion generation method to obtain early warning decisions and executable maintenance plans; S71 models the optimization of the early warning strategy as a reinforcement learning problem to train an intelligent early warning decision-making agent. A state space for reinforcement learning is defined, where each state contains information across multiple dimensions, including the fault prediction probability obtained in step S5, the health score and degradation rate obtained in step S6, the equipment importance level, the urgency of the current production task, the availability of maintenance resources, and the accuracy of historical early warnings. These state variables are combined into a state vector, which serves as the input to the intelligent agent. An action space is defined, where actions represent the choice of early warning decisions, including four actions: not issuing an early warning, issuing a level-of-concern warning, issuing a level-of-concern warning, and issuing an emergency warning. Different levels of early warning correspond to different handling requirements and response times: level-of-concern warnings require continuous monitoring, level-of-concern warnings require scheduling inspections, and emergency warnings require immediate action.

[0064] S72, a reward function is designed to evaluate the effectiveness of early warning decisions and guide the learning of intelligent agents. The reward function comprehensively considers the accuracy, timeliness, and cost-effectiveness of early warnings. A positive reward is given when an early warning is correct and the fault is handled promptly, preventing losses; the reward amount is proportional to the amount of loss avoided. A negative reward is given when a false alarm occurs (equipment malfunction but no early warning is issued); the penalty is proportional to the actual loss caused by the fault. A negative reward is given when a false alarm occurs (early warning is issued but the equipment is not malfunctioning); the penalty is proportional to the unnecessary maintenance costs and production disruptions caused by the false alarm. A reward for early warning lead time is introduced; the longer the lead time, the more preparation time for maintenance personnel, and the greater the reward. However, early warnings may lead to a higher false alarm rate, requiring a balance between the two. The reward function also considers the reasonableness of the early warning level; issuing low-level warnings for serious faults or high-level warnings for minor anomalies will incur penalties.

[0065] S73 employs a deep Q-network algorithm to train an early warning decision-making agent. The deep Q-network approximates the state-action value function (SAF) through a neural network, which represents the long-term cumulative reward expectation of taking a certain action in a given state. The neural network takes the state vector as input and outputs the Q-value of each action; a larger Q-value indicates a better action. The intelligent agent selects actions based on the Q-values, employing an epsilon greedy strategy to balance exploration and exploitation. It randomly selects actions for exploration with epsilon probability and selects the action with the highest Q-value for exploitation with a probability of 1 minus epsilon. After executing an action, the environment returns a new state and reward, and the experience tuple consisting of the state, action, reward, and new state is stored in the experience replay buffer.

[0066] S74: A batch of experience tuples is randomly sampled from the experience replay buffer for training the deep Q-network. For each experience tuple, a target Q-value is calculated, which equals the immediate reward plus a discount factor multiplied by the maximum Q-value in the new state. The mean squared error between the current Q-network's predicted Q-value for that experience and the target Q-value is calculated as the loss function. The parameters of the Q-network are updated using gradient descent to minimize the loss function. To improve training stability, a target network technique is employed, maintaining two Q-networks: an online network for action selection and updates, and a target network for calculating the target Q-value. The parameters of the target network are periodically copied from the online network, rather than updated every step, to avoid drastic fluctuations in the target value.

[0067] S75 establishes an early warning effectiveness evaluation and feedback mechanism to track the follow-up handling and actual failure occurrence of each early warning. For each early warning, record the warning time, warning level, predicted failure type, and recommended handling measures. Track the actions and results of maintenance personnel, recording whether inspections were conducted, whether failures were discovered, whether repairs were performed, and the effectiveness of the repairs. Track the subsequent operation of the equipment, recording whether failures occurred within the time window following the early warning, whether the failure type was consistent with the prediction, and the magnitude of the losses caused by the failure. Based on this information, calculate performance indicators such as early warning accuracy, recall, false alarm rate, false negative rate, and average early warning lead time. Use the early warning effectiveness as a reward signal to feed back to the reinforcement learning system for continuous optimization of the early warning strategy.

[0068] S76 trains dedicated early warning strategy models for different equipment types and business scenarios, enabling personalized early warning decisions. Different equipment types exhibit varying fault characteristics and impact levels. For example, faults in critical main equipment have a significant impact, requiring more sensitive early warnings, while faults in auxiliary equipment have a smaller impact and can be addressed with more conservative early warnings. Different business scenarios have different early warning needs. For instance, during urgent order production, more timely early warnings are needed to prevent downtime, while during normal production, early warnings can be appropriately delayed to reduce false alarms. Independent early warning strategy models are trained for each equipment type and business scenario, or a general model is trained and its parameters are adjusted based on equipment type and scenario. In practical applications, the appropriate early warning strategy model is selected for decision-making based on the current equipment and scenario.

[0069] S77 constructs a knowledge graph of equipment faults to generate intelligent handling suggestions. The knowledge graph organizes equipment fault-related knowledge in a graph structure. Nodes represent knowledge entities, including equipment type entities, fault type entities, fault cause entities, fault phenomenon entities, handling measures entities, spare parts entities, and tool entities. Edges represent relationships between entities, including relationships such as equipment has a fault, the fault is caused by a cause, the fault manifests as a phenomenon, the fault is handled by a measure, the measure requires spare parts, and the measure uses a tool. Natural language processing technology is used to extract knowledge entities and relationships from historical maintenance records, equipment technical manuals, and expert experience documents to automatically construct the knowledge graph. For structured maintenance record databases, data is converted into nodes and edges of the knowledge graph through database queries and mapping rules. For unstructured text documents, named entity recognition technology is used to identify entities in the text, and relation extraction technology is used to identify relationships between entities, adding the extracted knowledge to the graph.

[0070] S78 designs a knowledge graph-based reasoning mechanism to generate response suggestions based on current early warning information. After generating an early warning decision using a trained early warning strategy model, the early warning information, including equipment type, fault type, and fault symptoms, is used as input to the knowledge graph. A path search is performed within the knowledge graph, starting from equipment type and fault type nodes and traversing along relational edges to find associated response measure nodes. Graph traversal algorithms, such as breadth-first search or depth-first search, are used to search all reachable response measure nodes. The found response measures are scored and ranked, with scoring considering multiple factors, including the historical success rate, implementation difficulty, required time, required cost, and availability of required spare parts. The historical success rate is calculated by comparing the number of times the measure has been effective in historical cases to the total number of uses. Information such as implementation difficulty, required time, and required cost is obtained from the attributes of the knowledge graph or through expert annotation.

[0071] S79 introduces a case-based reasoning method to retrieve historical cases similar to the current fault and their corresponding solutions. A historical fault case database is established, with each case containing information such as the equipment status at the time of the fault, sensor data, fault type, cause, corrective measures taken, and the effectiveness of the corrective measures. The similarity between the current fault and historical cases is calculated, comprehensively considering the matching degree of equipment type, fault type, fault phenomenon, and equipment status. The matching degree of equipment type and fault type is judged by category consistency; if they match, the matching degree is 1; otherwise, it is 0. The similarity of fault phenomena is calculated by comparing the feature vectors of sensor data using cosine similarity or Euclidean distance. The similarity of equipment status is calculated by comparing the closeness of parameters such as health, running time, and operating conditions. An overall similarity score is obtained by combining the similarities across all dimensions. The historical cases with the highest similarity are selected, and their corresponding solutions are extracted as candidate suggestions.

[0072] S710 comprehensively evaluates and makes decisions on multiple candidate solutions obtained from knowledge graph reasoning and case-based reasoning. A multi-criteria decision-making model is established, with evaluation dimensions including solution effectiveness, implementation difficulty, required time, required cost, and risk level. Solution effectiveness is measured by historical success rate or expert ratings; the better the effectiveness, the higher the score. Implementation difficulty is measured by required skill level and operational complexity; the lower the difficulty, the higher the score. Less required time and cost, the higher the score. Risk level assesses potential new problems introduced during the solution process or its impact on production; the lower the risk, the higher the score. Each evaluation dimension is normalized, and a weighted summation method is used to calculate the comprehensive score of each solution. The weighting coefficients are set according to current business needs; for example, time has a higher weight in emergency situations, and cost has a higher weight in cost-sensitive situations. The solution with the highest comprehensive score is selected as the recommended primary solution, while several alternative solutions are provided for maintenance personnel to refer to.

[0073] S711 generates a detailed handling recommendation report, including fault diagnosis conclusions, recommended handling solutions, alternative solutions, a list of required spare parts, a list of required tools, estimated repair time, operating procedures, and precautions. The fault diagnosis conclusions are based on the fault prediction results from step S5 and the health assessment results from step S6, providing a judgment on the fault type, cause, and severity. The recommended handling solution details the specific steps of the handling process, with each step including operational content, key points, and safety precautions. The list of required spare parts lists the parts that need to be replaced during the handling process, including the spare part name, specifications, quantity, and inventory status. The list of required tools lists the tools and equipment needed. The estimated repair time is estimated based on the average repair time of historical cases. The precautions remind users of potential problems and corresponding solutions.

[0074] S712 establishes a feedback and knowledge update mechanism for handling effectiveness. After implementing a handling plan, maintenance personnel record the actual handling process and results, including whether the suggested plan was followed, whether any problems were encountered, whether the fault was ultimately resolved, and the actual time and cost spent. The handling results are fed back into the knowledge graph and case base. For effective handling plans, their success rate is increased, raising their priority in subsequent recommendations. For ineffective handling plans, their recommendation priority is lowered or they are marked as not recommended. Newly discovered fault types or new handling methods are added as new knowledge to the knowledge graph, enriching the knowledge base. Through continuous feedback and updates, the knowledge graph is constantly improved, and the quality of handling recommendations is continuously enhanced.

[0075] Through the above sub-steps, personalized early warning decisions and executable maintenance plans are obtained. The reinforcement learning-driven early warning strategy can adaptively select the optimal early warning level and timing based on equipment status, business needs, and historical results, balancing accuracy, timeliness, and cost-effectiveness. The knowledge graph-driven handling suggestion generation mechanism can comprehensively utilize expert knowledge and historical experience to provide maintenance personnel with detailed, reliable, and executable maintenance plans, improving the efficiency and success rate of fault handling.

[0076] In some embodiments, since certain novel or rare faults lack relevant knowledge in the knowledge graph and case base, reasoning methods based on equipment physical models and fault mechanism models can be used to generate handling suggestions, aiming to improve the ability to handle unknown faults. Specifically, a physical simulation model of the equipment is established to describe the structure, function, and interaction relationships of each component. A fault mechanism model is established to describe the occurrence mechanism, evolution process, and impact range of different fault types. When encountering a fault type not found in the knowledge graph, causal reasoning is performed based on the physical model and mechanism model to analyze the possible causes and impact paths of the fault. Based on the reasoning results, hypothetical handling solutions are generated; for example, if the reasoning indicates that the fault is caused by the failure of a certain component, it is recommended to check and replace that component. The handling suggestions based on model reasoning are marked as hypothetical suggestions to remind maintenance personnel to verify them carefully. After the handling is implemented, the correctness of the hypothesis is verified based on the actual effect, and the verified knowledge is added to the knowledge graph to achieve incremental accumulation of knowledge.

[0077] In one embodiment of the present invention, the focus is on the application of fault prediction and early warning for a group of CNC machine tools in a large machining workshop in the field of intelligent manufacturing. The workshop is equipped with 20 high-end CNC machining centers for the precision machining of aero-engine parts. The group of equipment operates continuously 24 / 7, and the failure of any critical equipment may lead to production line shutdown and cause significant economic losses.

[0078] The workshop has established an equipment failure prediction and early warning system using the method of this invention. The system's operation after deployment is as follows: In the data acquisition and processing phase, the system collects multimodal sensor data from 20 machining centers in the equipment group. Each machine is equipped with spindle temperature sensors, spindle vibration sensors, feed axis current sensors, tool wear monitoring sensors, coolant temperature sensors, etc., totaling more than 100 sensor nodes. The sensor sampling frequencies range from 1 Hz to 10 kHz, generating approximately 500 GB of raw data per day.

[0079] Through the dynamic topology discovery mechanism in step S1, the system automatically identifies the process relationships between equipment and constructs an equipment relationship map. For example, the processing output of equipment A is the processing input of equipment B, and there is a strong correlation between the two. Through multi-scale feature extraction in step S2, the system extracts more than 2,000 dimensions of features from the original sensor data, including time-domain statistical features, frequency-domain spectral features, wavelet decomposition features, etc.

[0080] Regarding data quality, the system detected drift in some sensors and corrected it using an adaptive calibration mechanism. The system also identified several data transmission interruptions that resulted in missing data, which were filled using interpolation. After processing, the data integrity rate reached over 98.5%, providing a reliable data foundation for subsequent analysis.

[0081] During the training and deployment phase of the fault prediction model, the system collected equipment operation data and fault records from the workshop over the past two years, including approximately 1 million normal operation samples and approximately 2,000 fault samples, covering more than 10 fault types such as spindle bearing failure, abnormal tool wear, cooling system leakage, and servo system failure. The fault sample ratio was only 0.2%, showing a severe class imbalance.

[0082] Through the graph convolutional network in step S3, the system learns the correlation features between devices, enabling it to identify cross-device fault propagation patterns. Through meta-learning and virtual sample generation in step S4, the system still achieves high recognition accuracy even with scarce fault samples. After model training, the overall accuracy on the test set reaches 96.8, and the recall rate for rare fault types reaches 89.3, outperforming traditional methods.

[0083] The online operation and continuous optimization phase: After system deployment, the model is continuously optimized through the online learning mechanism in step S5. Within three months of operation, the system processed 15 new fault samples and updated model parameters through incremental learning. The system's active learning mechanism identified 120 high-uncertainty samples requiring manual annotation; the annotated samples were used for model optimization, further improving model performance.

[0084] Through the health assessment in step S6, the system calculates a health score for each piece of equipment in real time. The health score of a certain machining center gradually decreased from 0.92 to 0.75 within 3 months. The system identified that the equipment had entered the accelerated degradation stage and predicted its remaining service life to be approximately 15 days.

[0085] Early warning and response phase; In a certain early warning case, the system detected an abnormal increase in the vibration of the spindle of machining center with equipment number MC-08, and the health level dropped to 0.68. Through the reinforcement learning early warning strategy in step S7, the system comprehensively considers factors such as failure probability, health level, equipment importance, and current production tasks, and decides to issue an early warning level warning with a lead time of 5 days.

[0086] The system generates a handling suggestion report through knowledge graph reasoning and case retrieval. The fault type is spindle bearing wear. The recommended handling solution is to replace the spindle bearing. The required spare parts are listed as two sets of angular contact ball bearings of model 7014C. The required tools include a special puller, torque wrench, etc. The estimated repair time is 8 hours.

[0087] The maintenance personnel arranged a planned shutdown for maintenance as advised, and completed the bearing replacement without affecting urgent orders. After maintenance, the equipment returned to normal operation, and its health level recovered to 0.95. Without prior warning, this malfunction could have occurred two days later, causing at least 24 hours of unplanned downtime and resulting in a production loss of approximately 500,000 yuan.

[0088] Overall system performance; statistical data from 6 months of system operation is shown in Table 1: Table 1. System Operation Statistics

[0089] As can be seen from the table, after system deployment, the accuracy of fault prediction and the recall rate of rare faults improved, and the average early warning time reached 5.2 days, providing ample preparation time for maintenance personnel. The number of unplanned downtimes decreased from 18 to 3, and the average equipment availability increased from 91.5% to 97.8%, resulting in improved production efficiency and economic benefits.

[0090] This application example verifies the effectiveness and practicality of the method of the present invention in real industrial scenarios, and demonstrates the technical advantages of the system in multi-source data fusion, rare fault identification, adaptive early warning decision-making, and intelligent handling suggestions.

[0091] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for device fault prediction and early warning using IoT data mining, characterized in that, Includes the following steps: S1, based on dynamic topology discovery and multi-source data standardization access through device communication behavior analysis, obtains device association map and multimodal data stream; S2, acquire device association map and multimodal data stream, and use sensor data feature extraction and quality enhancement methods to obtain multidimensional feature matrix and data quality assessment results; S3: Obtain the device association map, multi-dimensional feature matrix and data quality assessment results, and use the association feature learning and incremental feature update method to obtain deep feature representation; S4. Obtain deep feature representations and use small sample fault identification and virtual sample generation methods to obtain a fault identification model; S5. Based on the fault identification model and deep feature representation, an adaptive evolution fault prediction model is obtained by using continuous model optimization and performance monitoring methods. S6. Based on the deep feature representation and adaptive evolution fault prediction model, a health assessment and fault prediction collaborative optimization method is adopted to obtain the quantitative results of equipment health status and the prediction of remaining life. S7 acquires quantitative results of equipment health status and prediction of remaining lifespan, and uses an intelligent handling suggestion generation method to obtain early warning decisions and executable maintenance plans.

2. The method for device fault prediction and early warning based on IoT data mining according to claim 1, characterized in that, S1 includes: Real-time communication data streams of each device in the device group are collected, and the communication source address, destination address, interaction frequency and data flow direction features are extracted to construct a device communication behavior matrix. The association strength value between devices is calculated based on the communication behavior matrix. When the association strength value exceeds the preset association threshold, a connection edge is established in the device association graph. Establish a device metadata registry and develop data parsing adapter modules for different communication protocols to convert the parsed data into a unified internal format.

3. The method for device fault prediction and early warning based on IoT data mining according to claim 1, characterized in that, S2 includes: Wavelet transform is used to decompose the sensor time-series data in the multimodal data stream into multiple layers to obtain approximation coefficients and detail coefficients at different scales. Construct a sliding time window, calculate statistical characteristics within each time window, and perform a fast Fourier transform on the data to obtain the frequency domain amplitude spectrum; A sensor baseline model is established. When the deviation between the measured value and the baseline exceeds the preset drift threshold and lasts for a preset time, the sensor is determined to have drifted and adaptive calibration is performed.

4. The method for device fault prediction and early warning through IoT data mining according to claim 1, characterized in that, S3 includes: Each device is treated as a node in the graph, and the feature vector of the node is composed of a multi-dimensional feature matrix. The associated edges in the device association graph are used as the edges of the graph. The design uses a graph convolutional network to learn the associated features of device groups. A neighbor sampling strategy is adopted to randomly sample neighbor nodes of the target node for feature aggregation. An attention mechanism is introduced to calculate the attention score of the neighbor nodes and aggregate the features of the neighbor nodes in a weighted manner. A time attention module is introduced to calculate the time attention scores of the current moment and each historical moment, and to aggregate the features of each historical moment in a weighted manner.

5. The method for device fault prediction and early warning based on IoT data mining according to claim 1, characterized in that, S4 includes: A meta-learning framework is introduced to train the fault identification model. The identification tasks of different equipment types, different operating conditions, and different fault types are used as the training task set. For each training task, a support set and a query set are constructed. Design a prototype network to compute the prototype of each category for samples in the support set, compute the distance between the feature vector of the sample to be predicted in the query set and the prototype of each category, and classify the query sample into the category corresponding to the nearest prototype. A metric learning method is introduced, and a triplet loss function is used for training. The loss function makes the distance between the anchor point and the positive sample less than the distance between the anchor point and the negative sample plus a preset boundary value. A memory enhancement network is established, which stores the feature representations of historical fault samples through an external memory module. When performing fault prediction, the feature vector of the current sample is used as the query, and the information of the retrieved historical samples is weighted and aggregated as auxiliary information. A virtual fault sample generator is constructed. A generative adversarial network is used to learn the distribution characteristics of fault samples. The generator generates feature vectors of virtual samples, and the discriminator outputs the probability that a sample is a real sample or a virtual sample.

6. The method for device fault prediction and early warning based on IoT data mining according to claim 1, characterized in that, S5 includes: An online learning mechanism is established. When a device malfunctions and is labeled, the malfunction sample is used as new training data. An incremental learning algorithm is used to update the model parameters. An elastic weight consolidation method is used to add a regularization term to the loss function to limit the variation of important parameters. The design of the sample importance assessment mechanism comprehensively evaluates the information content of the sample by considering its uncertainty, representativeness, and novelty, and calculates the importance score of the sample; a forgetting mechanism is introduced to assign a time decay weight to each historical sample. An active learning strategy is adopted, and when the uncertainty of a predicted sample exceeds a preset threshold, the sample is submitted to the operation and maintenance personnel for annotation. Establish a model performance monitoring mechanism to track the model's performance metrics in each fault category in real time, and trigger model retraining or parameter adjustment when the performance metrics fall below the threshold.

7. The method for device fault prediction and early warning based on IoT data mining according to claim 1, characterized in that, S6 includes: A multi-task learning framework is constructed, and a shared feature extraction network is designed to take deep feature representations as input. The output of the shared network is connected to the fault classification branch and the health regression branch. Design a joint loss function that includes fault classification loss, health regression loss, and consistency constraint loss; establish a multi-dimensional evaluation system for equipment health, including performance, reliability, safety, and economic dimensions, and use a weight learning method to automatically determine the fusion weights of each dimension index; A health degradation trend model was established, and the degradation curve was fitted by regression method and extrapolated to predict the health value at future time points; the Weibull distribution model was used to predict the remaining service life of the equipment.

8. The method for device fault prediction and early warning based on IoT data mining according to claim 1, characterized in that, The early warning decision in S7 includes: The optimization of the early warning strategy is modeled as a reinforcement learning problem. The state space is defined to include the fault prediction probability, health score and degradation rate, equipment importance level, urgency of the current production task, and availability of maintenance resources. The action space is defined to include not issuing an early warning, issuing a warning of concern level, issuing a warning level warning, and issuing an emergency level warning. The reward function is designed to comprehensively consider the accuracy, timeliness, and economy of the early warning. A deep Q-network algorithm is used to train the early warning decision agent. The neural network takes the state vector as input and outputs the Q value of each action.

9. The method for device fault prediction and early warning based on IoT data mining according to claim 1, characterized in that, The maintenance schemes that can be executed in S7 include: Construct a knowledge graph of equipment failures, where nodes represent knowledge entities including equipment type entities, failure type entities, failure cause entities, handling measures entities, and spare parts entities, and edges represent the relationships between entities; The design incorporates a knowledge graph-based reasoning mechanism. Warning information is used as the input to the knowledge graph, and path search is performed within the knowledge graph to find related response measures. The found response measures are scored and ranked based on historical success rate, implementation difficulty, required time, and required cost. A case-based reasoning method is introduced to establish a historical failure case database. The similarity between the current failure and historical cases is calculated, and the response plan is extracted from the historical case with the highest similarity. The candidate solutions are comprehensively evaluated, and a weighted summation method is used to calculate the comprehensive score of each solution. The solution with the highest comprehensive score is selected as the recommended main solution.

10. A device fault prediction and early warning system for Internet of Things (IoT) data mining, used to execute the steps of a device fault prediction and early warning method for IoT data mining as described in any one of claims 1-9, characterized in that, include: The data access module is used for dynamic topology discovery and standardized access of multi-source data based on device communication behavior analysis, to obtain device association maps and multimodal data streams; The feature extraction module is used to acquire device association maps and multimodal data streams. It employs sensor data feature extraction and quality enhancement methods to obtain multidimensional feature matrices and data quality assessment results. The feature learning module is used to obtain device association maps, multi-dimensional feature matrices and data quality assessment results. It uses association feature learning and incremental feature update methods to obtain deep feature representations. The fault identification module is used to obtain deep feature representations and adopts a small sample fault identification and virtual sample generation method to obtain a fault identification model. The model optimization module is used to obtain an adaptive evolution fault prediction model based on the fault identification model and deep feature representation, using continuous model optimization and performance monitoring methods. The health assessment module is used to obtain the quantitative results of equipment health status and the prediction of remaining life based on the deep feature representation and the adaptive evolution fault prediction model, using a collaborative optimization method of health assessment and fault prediction. The early warning and decision-making module is used to obtain quantitative results of equipment health status and prediction of remaining lifespan. It adopts an intelligent handling suggestion generation method to obtain early warning decisions and executable maintenance plans.

Images