Industrial Internet of Things equipment fault prediction system driven by artificial intelligence

Abstract

The invention discloses an artificial intelligence-driven industrial Internet of Things equipment fault prediction system, and relates to the technical field of industrial Internet of Things and predictive maintenance, an edge-cloud collaborative architecture is adopted, real-time acquisition, preprocessing and online fault prediction of industrial equipment sensing data are realized, the system uses a Transform neural network to construct a hierarchical spatio-temporal model, and the fault prediction of the industrial equipment sensing data is realized. Time sequence data are processed in a segmented mode through a sliding window method, a causal reasoning enhancement mechanism is integrated, an industrial equipment causal atlas is constructed, attention masks are generated, a model is focused on key features, and therefore prediction accuracy and interpretability are improved, meanwhile, a closed-loop continuous optimization mechanism is established by the system, and prediction efficiency is improved. And an edge fault prediction result and actual operation feedback are uploaded to a cloud, a causal atlas and model parameters are updated, adaptive optimization of the model is realized, and the system provides efficient, accurate and explainable decision support for preventive maintenance of industrial equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial Internet of Things and predictive maintenance, and specifically to an artificial intelligence-driven industrial Internet of Things device fault prediction system. Background Art

[0002] At present, there are multiple technical solutions for predictive maintenance of industrial equipment. For example, Chinese invention patent CN111401661B discloses a mechanical equipment predictive maintenance method and maintenance system. This system collects the original vibration acceleration data and temperature data of the equipment, and through preprocessing, data analysis, and fault feature extraction, realizes the prediction of equipment faults and the display of maintenance information. This solution effectively avoids ineffective manual operations and realizes a certain degree of automated maintenance. However, it mainly relies on traditional data collection and statistical analysis methods, lacks the modeling of deep-level features of the complex dynamic behavior of the equipment, has limited ability to explain the cause of faults, and is difficult to meet the requirements of high-precision prediction and fine fault analysis.

[0003] On the other hand, Chinese invention patent CN118092404B discloses an artificial intelligence-based PLC controller network preventive maintenance method and system. This system deploys intelligent sensing nodes in the PLC network to collect the operation parameters of each controller in real time, and uses a preset machine learning model for preliminary analysis, and then constructs a fault prediction model based on a deep neural network to timely monitor the operation state of the equipment. However, this technology mainly relies on traditional deep learning methods in model training and data processing, does not fully consider the mining of multi-scale spatio-temporal features and causal relationships, resulting in insufficient interpretability of the prediction results. At the same time, it lacks a closed-loop adaptive optimization mechanism and is difficult to cope with the complex and changeable actual working conditions in the industrial field.

[0004] In summary, although the existing technologies have achieved predictive maintenance to a certain extent, there are still the following deficiencies: the capture of spatio-temporal features of complex equipment is insufficient, and the prediction accuracy is difficult to meet high requirements; the prediction results lack clear fault explanations and are difficult to accurately point out the key fault reasons; data processing and model update lack a closed-loop adaptive optimization mechanism and cannot adapt to the dynamic changes of the industrial field environment in real time. Summary of the Invention

[0005] The purpose of the present invention is to overcome the deficiencies of the existing technologies and provide a new artificial intelligence-driven industrial Internet of Things device fault prediction system. By integrating Transformer hierarchical spatio-temporal modeling, causal reasoning enhancement, and edge-cloud collaborative closed-loop optimization, it effectively makes up for the deficiencies in the existing technologies and realizes high-precision, real-time, interpretable, and adaptively optimized fault prediction, providing reliable technical support for the preventive maintenance of industrial equipment.

[0006] The object of the present invention is achieved through the following technical solutions: An artificial intelligence-driven industrial Internet of Things device fault prediction system, the system includes edge devices and a cloud server, and adopts an edge-cloud collaborative architecture; the edge devices are used to collect sensor data of industrial devices and perform preprocessing; the cloud server is used to train a fault prediction model and perform online fault prediction inference based on the preprocessed data; the fault prediction model uses a Transformer neural network to achieve hierarchical spatio-temporal modeling, and performs segmented processing on sensor time series data through sliding window calculation; the fault prediction model integrates a causal reasoning enhancement mechanism; the causal reasoning enhancement mechanism generates an attention mask by constructing a causal graph of the operation of industrial devices, so as to guide the self-attention mechanism of the Transformer model to focus on key features with causal relevance; the system can monitor the operation status of industrial devices in real time and predict the probability of fault occurrence, and provide dynamic explanations for the prediction results based on the causal graph.

[0007] The cloud server is provided with a causal reasoning module for constructing a causal graph based on the historical operation data and expert knowledge of industrial devices, and the causal graph depicts the causal association between the sensor parameters of industrial devices and fault events; the causal reasoning module provides the causal graph to the Transformer fault prediction model to enhance the causal reasoning ability of the model.

[0008] An attention mask is applied in the self-attention mechanism of the Transformer model, and the attention mask is generated based on the causal graph; by applying the attention mask to the self-attention calculation of the Transformer model, the model only focuses on relevant sensor features indicated by the causal graph at each time step, thereby improving the accuracy of fault prediction and the interpretability of model output.

[0009] The preprocessing module of the edge device performs sliding window calculation, and divides the continuous sensor data stream into multiple time segments according to a predetermined window length and step size; each time segment is used as a sample to be input into the Transformer fault prediction model for analysis, and there is an overlap between adjacent time segments to ensure continuous monitoring of fault symptoms.

[0010] The hierarchical spatio-temporal modeling of the Transformer fault prediction model is achieved through a multi-level structure: the bottom layer is used to capture the short-term temporal patterns of each sensor data sequence; the upper layer is used to fuse the correlation and long-term trends between multiple sensors; thus, it can accurately model the complex system behavior of industrial devices.

[0011] The system also includes a fault explanation module, which is used to determine the key sensor parameters or causal paths related to the prediction result by using the causal graph when it is predicted that the device has a fault risk, and output corresponding dynamic explanation information to assist the operation and maintenance personnel in analyzing the cause of the fault.

[0012] A local fault prediction unit is deployed on the edge device to run the Transformer fault prediction model trained by the cloud server, perform real-time analysis and fault prediction on the preprocessed sensing data, thereby reducing the prediction latency and reducing the dependence on network connection.

[0013] The cloud server includes a model training module for collecting operation data from multiple industrial Internet of Things edge devices and performing offline training and periodic updates on the Transformer fault prediction model; the cloud server also includes a central analysis module for receiving and integrating the fault prediction results or alarm information from the edge device, and performing centralized fault diagnosis and trend analysis.

[0014] A closed-loop continuous optimization mechanism is established between the edge device and the cloud server; the edge device sends the fault prediction result and the corresponding actual operation status feedback to the cloud server; the cloud server updates the causal graph based on the feedback and adjusts the parameters of the fault prediction model, and then sends the updated model to the edge device; thereby realizing the continuous improvement and adaptive optimization of the performance of the fault prediction model.

[0015] Its usage method includes the following steps: S1. The edge device collects the sensor data of the industrial device and performs preprocessing, and divides the sensor data into a continuous sliding window sequence; S2. The cloud server trains the Transformer fault prediction model based on the collected historical data and constructs the corresponding causal graph of the industrial device; S3. Deploy the trained fault prediction model and causal graph to the edge device; S4. The edge device uses the fault prediction model to perform online fault prediction inference on the real-time acquired sensor data window; S5. When the edge device detects possible fault symptoms or meets the preset conditions, upload the relevant sensor data and fault prediction results to the cloud server; S6. The cloud server updates the causal graph according to the feedback data received from the edge device, and retrains or adjusts the parameters of the fault prediction model; S7. Send the updated fault prediction model to the edge device to form a closed loop for continuous optimization of the model.

[0016] The beneficial effects of the present invention are: 1. By constructing a causal graph of industrial equipment operation and generating an attention mask, the Transformer model only focuses on sensor features closely related to faults during the self-attention calculation process, effectively filtering out irrelevant information, reducing the risk of false attribution, and improving the accuracy of fault prediction; a multi-level Transformer structure is adopted to capture the short-term temporal changes and long-term trends of sensor data respectively, enhancing the model's ability to express the dynamic behavior of complex industrial systems and further improving the prediction performance.

[0017] 2. Edge devices independently collect, preprocess data and run local fault prediction models to achieve real-time inference of on-site data, ensuring rapid response at the initial stage of fault occurrence and meeting the strict requirements of industrial sites for low-latency prediction; through the collaborative working mode of centralized training in the cloud and real-time inference at the edge, both the high-performance computing resources in the cloud are fully utilized and the dependence on network transmission is reduced, ensuring the normal operation of the system even when the network conditions are unstable.

[0018] 3. Based on the constructed causal graph and attention mask, the system can output detailed dynamic explanation information for each fault prediction, clearly indicating the key sensor data or causal paths, helping maintenance personnel quickly understand the root cause of faults and guiding targeted maintenance; the dynamic explanation mechanism not only improves the model transparency but also provides a strong basis for the traceability and location of industrial equipment faults, thus reducing the ambiguity of maintenance decisions.

[0019] 4. Edge devices feedback the real-time fault prediction results and actual operation status to the cloud, and the cloud server updates the causal graph according to the feedback data and retrains or adjusts the parameters of the Transformer model to form a continuously optimized closed-loop system, enabling the model to adapt to the ever-changing operating environment of industrial sites; through centralized data aggregation and offline training, the cloud server can regularly update the model and causal graph to ensure that the system always adopts the latest and optimal prediction strategies, enhancing the long-term stability and adaptability of the overall system.

[0020] 5. By integrating the fault prediction data and alarm information from multiple edge devices, the cloud server can conduct centralized fault diagnosis and trend analysis for the entire industrial system, providing a macro decision-making basis for the preventive maintenance and global monitoring of large-scale industrial equipment; global monitoring and trend analysis enable maintenance personnel to identify potential fault risks in advance and take preventive measures in a timely manner, thereby reducing equipment downtime and maintenance costs and improving production safety and efficiency.

[0021] 6. By performing real-time data processing and local fault prediction at the edge, the need for real-time data transmission to the cloud is reduced, effectively lowering the network transmission burden and enhancing the reliability of the system in bandwidth-constrained scenarios. The system design fully considers the resource limitations of the industrial site and the compatibility with existing PLC systems, enabling deployment without large-scale renovation, reducing the implementation cost, and having good scalability, making it suitable for various industrial environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 System interaction of the present invention Figure 1 ; Figure 2 System interaction of the present invention Figure 2 . DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present invention.

[0024] It should be noted that in the following solutions, the orientation concepts of "left", "right", "up", "down", "front", "back", "inside", and "outside" are all relative directions, and will not be listed one by one here.

[0025] Embodiment 1: As Figure 1 shown, this embodiment adopts an edge-cloud collaborative architecture, in which the key is to implement the real-time fault prediction function at the edge. The cloud server is responsible for offline training and updating the fault prediction model and causal graph, while the edge device collects data on-site, performs preprocessing, local fault prediction, and dynamic interpretation, thereby reducing the prediction latency and dependence on network connections.

[0026] Composition of the edge device Sensor data acquisition module. The edge device collects real-time data through various sensors (such as temperature, pressure, vibration sensors, etc.) installed on industrial equipment to form a continuous data stream.

[0027] The preprocessing module uses a sliding window calculation method to segment the continuous data stream. The specific method is as follows: Preset a fixed window length (such as 100 ms, 1 second, etc.) and an overlapping step size, and segment the original data into multiple time segments; There is partial overlap between adjacent time segments to ensure continuous capture of sudden fault symptoms; Each time segment is used as an independent sample for subsequent model fault prediction analysis.

[0028] The local fault prediction unit is deployed at the edge side and is used to run the Transformer fault prediction model trained by the cloud server. After offline training, this model integrates a causal reasoning enhancement mechanism, which can generate an attention mask using the constructed causal graph of industrial equipment to guide the model to focus on the key features related to faults.

[0029] The fault prediction model uses a Transformer neural network to implement hierarchical spatio-temporal modeling. The model can encode each time segment of the input to capture local high-frequency changes and long-term trends. The model integrates a causal reasoning mechanism: First, construct a causal graph of the operation of industrial equipment through historical data and device expert knowledge; then, generate an attention mask according to the causal graph and embed it into the Transformer self-attention layer, so that the model focuses on sensor data with causal relevance during information interaction, thereby improving prediction accuracy and interpretability.

[0030] Let the sensing data of industrial equipment be: where x t represents the sensor data at each time t; construct a causal graph G ∈ R n×n whose element G ij represents the causal influence intensity of sensor i on sensor j. For example, G ij = 0.8G means that the change of sensor i has a strong causal influence on j.

[0031] For the data within the sliding window W (k) ⊂ X (window length Δ), calculate the query matrix Q, key matrix K, and value matrix V through the Transformer model.

[0032] W (k) The data subsequence (time segment) within the k-th sliding window, usually with a length of Δ, and there can be overlap between windows. It is used for local modeling of time series features.

[0033] The length of the Δ sliding window, that is, how many time steps each subsequence contains (e.g., 100 means 0.5 seconds).

[0034] k represents the number of the k-th sliding window, and its value ranges from 1 to T - Δ + 1, which is used to traverse the entire time series.

[0035] G ij The edge weight in the causal graph. The larger the value, the stronger the influence of variable i on variable j. It can be a positive number (promoting causality), a negative number (inhibiting causality), or 0 (no direct causal relationship).

[0036] The causal fusion operator ⊕ is defined as: It is used to fuse the traditional dot - product attention scores with causal influence information. This operation has a normalization and co - regulation effect, enhancing the non - linear expression of the synergy strength between two variables. |ϵ| is a small constant used to prevent the denominator from being zero or too small, typically taking a value such as 1e−6.

[0037] Design a causal transformation function Γ for G ij to perform non - linear compression and smoothing, making it adapt to the attention weight range and enhancing the sensitivity to medium - strength causal information. The function formula is as follows: It fuses positive and negative causal information and has strong differentiability and smoothness.

[0038] The improved self - attention weight (at time step t), at time t, the attention weight of position i to position j, which fuses the attention score and causal information. It is used to measure which sensor channel features the Transformer should focus on at a specific time position. Its definition is as follows: Compared with the traditional Transformer, the causal weight is directly involved in the weighted mechanism of the normalized softmax. The output representation of the model at time step t and channel i is composed of a linear combination of all attention - weighted Value vectors, that is: T is the total number of time steps, representing the length of the sensor data sequence collected. For example, if the data is continuously collected for 10 seconds and sampled 10 times per second, then T = 100.

[0039] t is the current time step, 1 ≤ t ≤ T, representing a specific time point in the entire time - series data.

[0040] x t ∈Rn is the data vector collected by n sensors at time t, where each component is the observation value of the i - th sensor.

[0041] The entire original sensor data sequence, with a length of T, is the main input of the model.

[0042] n is the total number of sensors deployed on industrial equipment, i.e., the data dimension at each time point. Commonly, there are multiple channels such as temperature, pressure, current, voltage, rotational speed, vibration, etc.

[0043] ∈R d is the query vector at the t-th time step and position i in the Transformer, obtained from the input through a linear transformation or an embedding module, with a dimension of d. It is used to initiate an attention query for information at other positions.

[0044] ∈Rd is the key vector at the t-th time step and position j in the Transformer, used to calculate the attention matching degree with the query vector.

[0045] ∈Rd is the value vector at the t-th time step and position j, which participates in the final representation output after being weighted by the attention weights.

[0046] d is the dimension of the vector, usually the hidden space dimension set by the model (such as 64, 128, etc.), used to control the representation ability and the calculation scale.

[0047] i, j, k respectively represent different feature channel indices, with the range 1 ≤ i, j, k ≤ n, and are used for cross-channel modeling in the attention calculation.

[0048] Overall structural flow of the formula Input data: Multi-channel time series sensor data X = {xt} Sliding window processing: Segmentally extract local time slices W (k) Transformer encoding: Extract Q, K, V vectors for each segment of data Introduce the causal graph G: Constructed based on experience or data derivation Causal enhanced attention calculation: Dot product attention scoring Q T K√d Nonlinear transformation of causal weights Γ(G ij ) Fusion operation ⊕ Softmax normalization to obtain Output representation : As the input basis for prediction and interpretation Working process Data acquisition and preprocessing Edge devices collect sensor data of industrial equipment in real time; The preprocessing module uses the sliding window method to divide the continuous data stream into multiple overlapping time segments according to the predetermined window length and step size, and each time segment constitutes a sample.

[0049] Local model inference Each time segment sample is input into the Transformer fault prediction model deployed at the edge; The attention mask generated by the pre-embedded causal graph is used inside the model to guide the self-attention mechanism to focus on key sensor features and extract local spatio-temporal patterns closely related to faults; The model outputs the real-time probability of fault occurrence, and through the dynamic interpretation function of the attention mechanism, provides key feature information related to the prediction result.

[0050] Fault prediction and response The edge-side fault prediction unit triggers local alarms or control instructions according to the fault probability output by the model, supporting the immediate monitoring and fault emergency handling of industrial equipment; Meanwhile, the prediction results and dynamic interpretation information can be uploaded to the cloud server when network conditions permit, for centralized management, historical data analysis, and continuous model optimization.

[0051] By deploying a local fault prediction unit at the edge, the immediate analysis and fault judgment of the segmented samples are realized, greatly reducing data transmission and processing delays, and meeting the requirements of the industrial site for low-latency response.

[0052] The edge-side independently executes fault prediction, avoiding prediction interruption caused by network latency or instability, and ensuring continuous monitoring and safe operation of on-site equipment even in an environment with poor network connection.

[0053] After integrating the causal inference enhancement mechanism, the Transformer model uses the attention mask generated by the causal graph, enabling the model to only focus on features with actual physical and causal relevance, thereby improving the accuracy of fault prediction; meanwhile, the dynamic interpretation function helps maintenance personnel quickly understand the prediction basis and facilitates the timely adoption of targeted maintenance measures.

[0054] The sliding window is used to calculate and segment the data stream, making adjacent time segments overlap, ensuring continuous capture of equipment state changes, timely discovery of minor anomalies, and early warning of fault trends.

[0055] Generally speaking, this embodiment effectively realizes low-latency prediction and on-site rapid response of industrial IoT device faults by implementing preprocessing, real-time model inference, and dynamic interpretation of sensor data at the edge, providing a reliable technical guarantee for the safe and stable operation of industrial equipment.

[0056] Embodiment 2: As Figure 2 shown, based on Embodiment 1, the entire system in this embodiment still adopts an edge-cloud collaborative architecture, but the focus is on the centralized processing of the cloud server. The cloud server integrates a causal reasoning module, a model training module, a central analysis module, and a fault explanation module, and uses large-scale historical data and multi-edge device feedback to achieve offline training, periodic update, and causal reasoning enhancement of the Transformer fault prediction model.

[0057] Composition of the cloud server The causal reasoning module constructs a causal map of the operation of industrial equipment using historical operation data and expert knowledge.

[0058] The causal map depicts the causal associations between various sensor parameters and fault events, and generates corresponding attention masks for use by the Transformer model to enhance the model's causal reasoning ability.

[0059] The model training module collects historical operation data from multiple edge devices and performs offline training of the Transformer fault prediction model.

[0060] This model adopts a hierarchical spatio-temporal modeling structure: The bottom layer captures the temporal patterns of short-term sensor data sequences; The upper layer fuses the correlations and long-term trends between multiple sensors to accurately model the complex behaviors of industrial equipment.

[0061] During the training process, the attention masks generated by the causal map are embedded into the self-attention layer of the Transformer to ensure that the model only focuses on the relevant features indicated by the causal map at each time step.

[0062] The fault explanation module, when the model predicts that there is a fault risk in the device, uses the pre-constructed causal map to determine the key sensor parameters or causal paths related to the prediction result, and generates dynamic explanation information to assist the operation and maintenance personnel in analyzing the cause of the fault.

[0063] The central analysis module aggregates and integrates the fault prediction results and alarm information from multiple edge devices, performs centralized fault diagnosis and trend analysis, and provides macro fault situation monitoring and prediction support for the entire industrial system.

[0064] Collaboration between the edge and the cloud Data upload and feedback. The edge device uploads the preprocessed data and real-time fault prediction results to the cloud.

[0065] The cloud server continuously updates the causal graph and the Transformer model based on the feedback data from each edge device, realizing the periodic retraining and performance optimization of the model.

[0066] Model distribution and update: The latest Transformer fault prediction model obtained by training in the cloud model training module will be distributed to the edge devices to ensure that the on-site fault prediction module always uses the latest and optimized model.

[0067] Working process Causal graph construction and attention mask generation The cloud server first constructs the causal graph of industrial equipment through the causal reasoning module using a large amount of historical data and expert input.

[0068] According to the constructed causal graph, the system generates the corresponding attention mask and provides it to the Transformer model, enabling the model to only focus on the sensor features related to fault prediction during the self-attention calculation process.

[0069] Offline model training and update The model training module collects historical data from multiple edge devices and performs offline training on the data using a multi-level Transformer model.

[0070] During the training process, using the attention mask and the hierarchical spatio-temporal modeling structure, the model can accurately capture short-term signal fluctuations and long-term trends.

[0071] After training is completed, the cloud server evaluates and validates the model, and periodically performs model retraining and update.

[0072] Fault explanation and central analysis When the cloud model detects a potential fault risk in the device, the fault explanation module uses the causal graph to explain the model prediction results and generates a dynamic explanation report of key sensor parameters or causal paths.

[0073] At the same time, the central analysis module integrates the fault prediction and alarm information from each edge device, conducts centralized fault diagnosis and trend analysis, and forms a global fault monitoring report to provide data support for decision-making.

[0074] Model distribution and edge collaboration The latest trained model is distributed to the edge devices through a secure network channel.

[0075] The edge devices use the latest model to perform real-time fault prediction and feedback the results back to the cloud, forming a closed loop of continuous data and model optimization.

[0076] Through cloud-based centralized model training, leveraging a vast amount of historical data and feedback from multiple edge devices, the modeling ability of the Transformer model for complex industrial equipment behavior can be significantly improved. Meanwhile, the causal reasoning enhancement mechanism ensures that the model only focuses on the truly relevant key features in fault prediction, thereby improving the prediction accuracy.

[0077] By integrating the causal graph and the attention mask mechanism, the fault explanation module can output detailed dynamic explanation information for each prediction, helping maintenance personnel quickly locate the root cause of the fault and enhancing the interpretability and operational confidence of the system.

[0078] The cloud-based central analysis module aggregates data from multiple edge devices, enabling fault diagnosis and trend analysis of the entire industrial system, providing a macro decision-making basis for equipment maintenance and preventive maintenance, and reducing the overall operation and maintenance risk.

[0079] Based on the continuously collected feedback data from edge devices, the cloud server can perform periodic retraining on the causal graph and the Transformer model to achieve model adaptive optimization and continuously improve the overall performance and stability of the system.

[0080] Through cloud-based centralized fault diagnosis and trend analysis, maintenance personnel can take targeted maintenance measures more timely, detect potential faults in advance and carry out preventive maintenance, thereby significantly reducing equipment downtime and maintenance costs.

[0081] Generally speaking, this embodiment realizes high-precision, global management and dynamic response for the fault prediction of industrial Internet of Things devices by centrally constructing a causal graph, offline training a Transformer model and dynamic fault explanation in the cloud. The cloud and the edge work together, not only improving the accuracy and interpretability of fault prediction, but also providing a strong guarantee for the safe and stable operation of large-scale industrial systems.

[0082] Embodiment 3: As Figure 1 and Figure 2 shown, this embodiment still adopts the edge-cloud collaborative architecture on the basis of Embodiments 1 and 2, but focuses on the centralized processing and closed-loop optimization of the cloud server. The entire system consists of edge devices and a cloud server. The edge side is mainly responsible for data collection, preprocessing and real-time online fault prediction, while the cloud server undertakes the tasks of offline model training, causal graph construction, fault explanation and model adaptive optimization. The main component modules include: The cloud-based causal inference module constructs a causal graph using the historical operation data of industrial equipment and expert knowledge, depicts the causal associations between various sensor parameters and fault events, and generates corresponding attention masks. These attention masks are applied in the self-attention layer of the Transformer model, enabling the model to focus only on the relevant features indicated by the causal graph at each time step, thereby improving the accuracy and interpretability of fault prediction.

[0083] The cloud-based model training module collects historical data from multiple edge devices and uses a Transformer neural network to achieve hierarchical spatio-temporal modeling, where: The bottom layer is used to capture the short-term temporal patterns of sensor data sequences; The upper layer is used to fuse the correlations and long-term trends between multiple sensors.

[0084] During the training process, a causal inference enhancement mechanism is embedded in the model to optimize the model parameters and ensure that the model has high prediction performance.

[0085] The closed-loop continuous optimization mechanism: Edge devices upload the real-time fault prediction results and the actual device operation status to the cloud server. The cloud server uses this feedback data to update the causal graph and retrain or adjust the parameters of the Transformer model. The updated model is sent to the edge devices through a secure channel to achieve continuous improvement and adaptive optimization of the model performance.

[0086] The central analysis module integrates the fault prediction data and alarm information from multiple edge devices, conducts centralized fault diagnosis and trend analysis on the entire industrial system, and provides macro support for system optimization and decision-making.

[0087] In addition, the following method steps are adopted in this embodiment to achieve the closed-loop operation of the entire system: S1: Data collection and preprocessing Edge devices collect the sensor data of industrial equipment and divide the data into continuous sliding window sequences to ensure the time continuity and overlap of data samples.

[0088] S2: Offline model training and causal graph construction The cloud server uses historical data to train the Transformer fault prediction model and simultaneously constructs the corresponding causal graph of industrial equipment to provide causal prior information for model training.

[0089] S3: Model and causal graph deployment After training is completed, the cloud server sends the updated fault prediction model and causal graph to the edge devices to ensure that the on-site fault prediction module always uses the latest optimized model.

[0090] S4: Online Fault Prediction Inference The edge device performs online fault prediction on the real-time collected data window using the downloaded model, and outputs the fault probability and dynamic interpretation results.

[0091] S5: Abnormality Feedback Upload When the edge device detects a potential fault or reaches the preset alarm threshold, it uploads the relevant sensor data and prediction results to the cloud server in a timely manner.

[0092] S6: Cloud Feedback Processing and Model Update Based on the feedback data received from the edge side, the cloud server updates the causal graph and retrains or adjusts the parameters of the Transformer model to adapt to the device operating state and environmental changes.

[0093] S7: Model Download and Continuous Optimization The updated model is downloaded from the cloud server to the edge device, forming a continuously optimized closed loop to ensure the continuous improvement of the overall prediction performance of the system.

[0094] Working Process Initial Model Training and Causal Graph Construction The cloud server first constructs an initial causal graph using historical data and expert knowledge, and generates the corresponding attention mask.

[0095] At the same time, the model training module uses the Transformer neural network to perform offline training on the historical data to obtain an initial fault prediction model.

[0096] After the training is completed, the cloud deploys the fault prediction model and the causal graph to the edge device (Steps S2 and S3).

[0097] Real-time Online Fault Prediction The edge device collects real-time sensor data, and the preprocessing module divides it into data samples according to the sliding window (Step S1).

[0098] Each sample runs the downloaded Transformer model through the online fault prediction unit at the edge side for fault prediction inference, and outputs the fault occurrence probability and the corresponding dynamic interpretation (Step S4).

[0099] Abnormality Detection and Feedback Upload When the edge device detects that the prediction result reaches or exceeds the preset fault threshold, the edge side uploads the relevant data and prediction results to the cloud server (Step S5).

[0100] Closed-loop Optimization and Model Update The feedback processing module of the cloud server receives the abnormal data and actual operating conditions from the edge devices, updates the causal graph, and retrains or fine-tunes the parameters of the fault prediction model using this data (step S6).

[0101] The updated model is sent to the edge devices through a secure channel to replace the old model, achieving closed-loop continuous optimization (step S7).

[0102] Central Analysis and Global Monitoring At the same time, the central analysis module integrates the data of multiple edge devices, conducts global fault diagnosis and trend analysis, forms a comprehensive report, and provides support for operation and maintenance decisions.

[0103] Through cloud-based centralized training, using a large amount of historical data and feedback from multiple edge devices for model optimization, and the embedding of causal reasoning enhancement mechanisms, the Transformer model can more accurately capture the complex spatio-temporal behavior of industrial devices, significantly improving the accuracy of fault prediction.

[0104] Based on the constructed causal graph to generate an attention mask, the model focuses on the key sensor features related to faults during inference, enabling each prediction to output dynamic interpretation information, helping operation and maintenance personnel clarify the cause of faults and key influencing factors, thereby enhancing the overall interpretability of the system.

[0105] The closed-loop continuous optimization mechanism established between the edge devices and the cloud server enables the system to adjust the causal graph and model parameters in a timely manner according to the actual device operation feedback, realizing model self-adaptive improvement, and effectively coping with the changing working conditions and data distributions in the industrial environment.

[0106] The cloud server integrates the operation data of multiple edge devices, realizes centralized fault diagnosis and trend analysis of the entire industrial system, provides macro decision-making support for preventive maintenance and equipment management, and reduces operation and maintenance risks and downtime costs.

[0107] Using the edge-cloud collaborative architecture, it not only gives full play to the advantages of large-scale computing and storage in the cloud, but also realizes real-time inference at the edge, ensuring that the system can still operate normally under unstable network conditions, while saving transmission bandwidth and on-site processing resources.

[0108] Generally speaking, through cloud-based centralized causal graph construction, offline training of the Transformer model, dynamic fault interpretation, and closed-loop continuous optimization in this embodiment, high-precision prediction and global monitoring of industrial equipment faults are achieved. The cloud and the edge work together, not only improving the prediction performance and interpretability of the system, but also realizing self-adaptive optimization and centralized management, providing a strong technical guarantee for the stable operation of industrial Internet of Things devices.

[0109] The above are only the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein, and should not be regarded as excluding other embodiments. Instead, it can be used in various other combinations, modifications, and environments, and can be changed within the scope of the concept described herein through the above teachings or the technology or knowledge in related fields. As long as the changes and variations made by those skilled in the art do not depart from the spirit and scope of the present invention, they should all be within the protection scope of the appended claims of the present invention.

Claims

1. An artificial intelligence-driven industrial Internet of Things device fault prediction system, characterized in that The system includes edge devices and a cloud server, and adopts an edge-cloud collaborative architecture; The edge devices are used to collect sensor data of industrial devices and perform preprocessing; the cloud server is used to train a fault prediction model and perform online fault prediction inference based on the preprocessed data; the fault prediction model uses a Transformer neural network to implement hierarchical spatio-temporal modeling, and performs segmented processing on sensor time series data through sliding window calculation; the fault prediction model integrates a causal reasoning enhancement mechanism; the causal reasoning enhancement mechanism generates an attention mask by constructing a causal graph of the operation of industrial devices to guide the self-attention mechanism of the Transformer model to focus on key features with causal relevance; the system can monitor the operation status of industrial devices in real time and predict the probability of fault occurrence, and provide dynamic explanations for the prediction results based on the causal graph.

2. The artificial intelligence-driven industrial Internet of Things device fault prediction system according to claim 1, wherein: The cloud server is equipped with a causal reasoning module for constructing a causal graph based on the historical operation data and expert knowledge of industrial devices, and the causal graph depicts the causal associations between the sensor parameters of industrial devices and fault events; The causal reasoning module provides the causal graph to the Transformer fault prediction model to enhance the causal reasoning ability of the model.

3. An artificial intelligence-driven industrial Internet of Things device fault prediction system according to claim 2, characterized in that: An attention mask is applied in the self-attention mechanism of the Transformer model, and the attention mask is generated based on the causal graph; by applying the attention mask to the self-attention calculation of the Transformer model, the model only focuses on the relevant sensor features indicated by the causal graph at each time step, thereby improving the accuracy of fault prediction and the interpretability of model output.

4. An artificial intelligence-driven industrial Internet of Things device fault prediction system according to claim 3, characterized in that: The preprocessing module of the edge device performs sliding window calculation, and divides the continuous sensor data stream into multiple time segments according to a predetermined window length and step size; each time segment is used as a sample to be input into the Transformer fault prediction model for analysis, and there is an overlap between adjacent time segments to ensure continuous monitoring of fault symptoms.

5. An artificial intelligence-driven industrial Internet of Things device fault prediction system according to claim 2, characterized in that: The hierarchical spatio-temporal modeling of the Transformer fault prediction model is achieved through a multi-level structure: the bottom layer is used to capture the short-term temporal patterns of each sensor data sequence; the upper layer is used to fuse the correlations and long-term trends between multiple sensors; thus, it can accurately model the complex system behavior of industrial devices.

6. An artificial intelligence-driven industrial Internet of Things device fault prediction system according to claim 5, characterized in that: The system also includes a fault explanation module, which is used to determine the key sensor parameters or causal paths related to the prediction result using the causal graph when a fault risk of the device is predicted, and output corresponding dynamic explanation information to assist the operation and maintenance personnel in analyzing the cause of the fault.

7. An artificial intelligence-driven industrial Internet of Things device fault prediction system according to claim 6, characterized in that: A local fault prediction unit is deployed on the edge device, which is used to run the Transformer fault prediction model trained by the cloud server, and perform real-time analysis and fault prediction on the preprocessed sensing data, thereby reducing the prediction delay and reducing the dependence on network connection.

8. An artificial intelligence-driven industrial Internet of Things device fault prediction system according to claim 7, characterized in that: The cloud server includes a model training module for collecting operation data from multiple industrial Internet of Things edge devices and performing offline training and periodic updates on the Transformer fault prediction model; the cloud server also includes a central analysis module for receiving and integrating fault prediction results or alarm information from edge devices and performing centralized fault diagnosis and trend analysis.

9. An artificial intelligence-driven industrial Internet of Things device fault prediction system according to claim 7, characterized in that: A closed-loop continuous optimization mechanism is established between the edge devices and the cloud server; the edge devices feedback and send the fault prediction results and the corresponding actual operating conditions to the cloud server; Based on the feedback, the cloud server updates the causal graph and adjusts the parameters of the fault prediction model, and then sends the updated model to the edge devices; thus realizing the continuous improvement and adaptive optimization of the performance of the fault prediction model.

10. An artificial intelligence-driven industrial Internet of Things device fault prediction system according to claim 1, characterized in that: Its usage method includes the following steps: S1. The edge device collects sensor data of industrial equipment and performs preprocessing, and divides the sensor data into continuous sliding window sequences; S2. The cloud server trains a Transformer fault prediction model based on the collected historical data and constructs a corresponding causal graph of industrial equipment; S3. Deploy the trained fault prediction model and the causal graph to the edge device; S4. The edge device uses the fault prediction model to perform online fault prediction inference on the sensor data window obtained in real time; S5. When the edge device detects possible fault signs or meets preset conditions, it uploads relevant sensor data and fault prediction results to the cloud server; S6. The cloud server updates the causal graph according to the feedback data received from the edge device, and retrains or adjusts the parameters of the fault prediction model; S7. Send the updated fault prediction model to the edge device to form a closed loop for continuous model optimization.

Images