Predictive maintenance decision method and system based on cooperation of plc and cloud platform

Abstract

The application discloses a predictive maintenance decision method and system of PLC and cloud platform cooperation, which deploys a light AI model on the edge side of PLC to output preliminary decisions and decision confidence scores in real time, and makes adaptive gating decisions based on the same to obtain reporting data packets. On this basis, the local model is adaptively updated through a decision closed-loop mechanism, thereby realizing dynamic balance between real-time and accuracy. In this way, the inherent contradiction between decision real-time and diagnostic accuracy in predictive maintenance is effectively solved. The rapid response capability of the edge side ensures timely processing of equipment state changes, and the confidence gating avoids the limitations of single-point judgment of the edge model, realizes on-demand deep diagnosis, and significantly improves diagnostic accuracy and system robustness.

Description

Technical Field

[0001] This application relates to the field of intelligent decision-making, specifically to a predictive maintenance decision-making method and system that integrates PLC and cloud platform. Background Technology

[0002] The stable operation of industrial production equipment is crucial for ensuring production efficiency, product quality, and operational safety. Traditional maintenance strategies, such as reactive repair and periodic preventative maintenance, often lead to unplanned downtime, high costs, and wasted resources, and are ineffective in preventing sudden failures. With the development of Industrial Internet of Things (IIoT), big data, and artificial intelligence (AI) technologies, predictive maintenance has become an inevitable trend for improving equipment reliability and reducing operating costs. It aims to provide early warnings of failures by monitoring equipment data in real time and intelligently analyzing its health status, achieving precise maintenance and maximizing equipment uptime and lifespan. However, in complex industrial environments, efficient and accurate predictive maintenance requires not only millisecond-level real-time response capabilities to handle urgent operating conditions, but also deep and high-precision diagnostic capabilities to accurately identify potential faults.

[0003] Existing technological solutions often face the core issue of the conflict between real-time decision-making and diagnostic accuracy when attempting to address this challenge. On the one hand, while edge-side PLC models can meet the real-time response requirements of industrial sites, their diagnostic accuracy and generalization ability to complex operating conditions are insufficient due to limitations in computing resources, easily leading to misjudgments or omissions; simple threshold or statistical models are insufficient to capture fault characteristics under complex operating conditions. On the other hand, while cloud platform deep models offer accurate diagnosis, network latency and lag caused by complex calculations cannot meet the millisecond-level real-time response requirements of industrial control, rendering them useless at critical moments. Existing attempts, such as model compression or simple division of labor between real-time control and offline analysis, have failed to build an organic and dynamic collaborative decision-making mechanism, and cannot effectively bridge the gap between edge-side real-time performance and cloud-based diagnostic accuracy. This architecture, lacking intelligent collaboration, makes it difficult for the system to balance rapid response and accurate judgment when dealing with complex decision-making needs in industrial sites, severely restricting the practical application effect of predictive maintenance.

[0004] Therefore, we look forward to an optimized predictive maintenance decision-making method that integrates PLCs and cloud platforms. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a predictive maintenance decision-making method and system that integrates a PLC and a cloud platform. It deploys a lightweight AI model at the PLC edge to output preliminary decisions and decision confidence scores in real time. Based on these, adaptive gating decisions are made to obtain reported data packets. Furthermore, a decision-making closed-loop mechanism adaptively updates the local model, thereby achieving a dynamic balance between real-time performance and accuracy. This effectively resolves the inherent contradiction between real-time decision-making and diagnostic accuracy in predictive maintenance. The rapid response capability at the edge ensures timely handling of changes in equipment status, while confidence gating avoids the limitations of single-point judgment by the edge model, enabling on-demand in-depth diagnosis and significantly improving diagnostic accuracy and system robustness.

[0006] According to one aspect of this application, a predictive maintenance decision-making method in collaboration between a PLC and a cloud platform is provided, comprising: Acquire raw sensor data; Real-time feature extraction of raw sensor data to obtain local feature vectors; Perform edge-side lightweight model inference on local feature vectors to obtain local preliminary decisions and local confidence scores; Confidence-gated decisions are made on the local preliminary decision and local confidence score to obtain the edge final decision and the reported data packet; Perform in-depth cloud platform diagnostics and model optimization on the reported data packets to obtain detailed diagnostic results and model update instructions from the cloud. Based on the detailed diagnostic results and model update instructions from the cloud, a decision-making closed loop is performed, and the local model is adaptively updated to obtain the final control instructions and the updated local model.

[0007] According to another aspect of this application, a predictive maintenance decision-making system that coordinates PLC and cloud platform is provided, comprising: Raw data acquisition module, used to acquire raw data from the sensor; The real-time feature extraction module is used to extract features from the raw sensor data in real time to obtain local feature vectors. The edge-side lightweight model inference module is used to perform edge-side lightweight model inference on local feature vectors to obtain local preliminary decisions and local confidence scores. The confidence gating decision module is used to perform confidence gating decisions on the local preliminary decision and local confidence score to obtain the edge final decision and the reported data packet; The cloud platform deep diagnostics and model optimization module is used to perform deep diagnostics and model optimization on the reported data packets to obtain detailed diagnostic results and model update instructions in the cloud. The decision-making closed-loop and local model adaptive update module is used to perform decision-making closed-loop and local model adaptive updates based on the cloud-based fine diagnostic results and model update instructions to obtain the final control instructions and the updated local model.

[0008] Compared with existing technologies, this application provides a predictive maintenance decision-making method and system that integrates PLC and cloud platform. It deploys a lightweight AI model at the PLC edge to output preliminary decisions and decision confidence scores in real time. Based on these, adaptive gating decisions are made to obtain reported data packets. Furthermore, a decision-making closed-loop mechanism adaptively updates the local model, achieving a dynamic balance between real-time performance and accuracy. This effectively resolves the inherent contradiction between real-time decision-making and diagnostic accuracy in predictive maintenance. The rapid response capability at the edge ensures timely handling of equipment status changes, while confidence gating avoids the limitations of single-point judgment by the edge model, enabling on-demand in-depth diagnosis and significantly improving diagnostic accuracy and system robustness. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a predictive maintenance decision-making method for PLC and cloud platform collaboration according to an embodiment of this application; Figure 2 This is a data flow diagram illustrating a predictive maintenance decision-making method for PLC and cloud platform collaboration according to an embodiment of this application. Figure 3 This is a block diagram of a predictive maintenance decision system that integrates a PLC and a cloud platform according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0016] The technical solution of this application proposes a predictive maintenance decision-making method that coordinates PLC and cloud platform. Figure 1 This is a flowchart of a predictive maintenance decision-making method for PLC and cloud platform collaboration according to an embodiment of this application. Figure 2 This is a data flow diagram illustrating a predictive maintenance decision-making method for PLC and cloud platform collaboration according to an embodiment of this application. Figure 1 and Figure 2 As shown, the predictive maintenance decision-making method for PLC and cloud platform collaboration according to an embodiment of this application includes the following steps: S1, acquiring raw sensor data; S2, performing real-time feature extraction on the raw sensor data to obtain a local feature vector; S3, performing edge-side lightweight model inference on the local feature vector to obtain a local preliminary decision and a local confidence score; S4, performing confidence-gated decision-making on the local preliminary decision and the local confidence score to obtain an edge final decision and a reported data packet; S5, performing cloud platform deep diagnosis and model optimization on the reported data packet to obtain cloud-based refined diagnosis results and model update instructions; S6, performing decision closure and local model adaptive update based on the cloud-based refined diagnosis results and model update instructions to obtain final control instructions and an updated local model.

[0017] Specifically, S1 involves acquiring raw sensor data. Raw sensor data refers to the unprocessed physical quantity signals collected in real time by sensors installed in key parts of industrial equipment, such as vibration sensors, temperature sensors, pressure sensors, and current / voltage sensors. This data typically exists in time-series form, meaning each data point has a corresponding timestamp, reflecting the equipment's operating parameters at a specific point in time. For example, a vibration sensor might provide the vibration acceleration values ​​of a bearing at different frequencies, a temperature sensor might provide the real-time temperature of a motor winding, and a pressure sensor might provide the working pressure of a hydraulic system. This raw data usually exhibits characteristics such as continuity, temporality, and diversity, representing the most fundamental digital representation of the equipment's operating status. Acquiring raw sensor data provides unprocessed raw information input for subsequent data processing and analysis, ensuring the accuracy and reliability of the basis for subsequent processing.

[0018] In practical implementation, this can be achieved by directly connecting various sensors to the input module of a PLC (Programmable Logic Controller). The sensors convert the monitored physical quantities (such as vibration, temperature, pressure, and current) into analog signals (such as 4mA-20mA current signals and 0V-10V voltage signals) or digital signals and transmit them to the PLC. The PLC's input module is responsible for receiving and digitizing these analog signals (via an A / D converter) or directly reading the digital signals, performing high-speed acquisition within a specific sampling period. These acquired digital quantities, accompanied by precise timestamps, constitute the raw sensor data. This raw data is then initially stored in the PLC's memory and used as direct input for the next stage. This process ensures the real-time nature and accuracy of the data, laying a solid foundation for subsequent lightweight model inference based on the edge.

[0019] Specifically, step S2 involves real-time feature extraction from the raw sensor data to obtain a local feature vector. It should be understood that raw sensor data is typically a high-dimensional, continuous, and potentially noisy time-series signal. This unprocessed raw data is large and redundant; directly inputting it into a model for analysis would significantly increase the model's computational burden and inference time, making it almost infeasible in edge-constrained computing environments such as PLCs. Furthermore, excessive noise and irrelevant information could reduce the model's predictive accuracy. Therefore, in the technical solution of this application, real-time feature extraction is performed on the raw sensor data to accurately extract the core information that best reflects the equipment's operating status and potential fault modes from the massive amount of raw data. This reduces data dimensionality, improves data quality, and enables the subsequent lightweight model to achieve effective fault mode identification while maintaining high efficiency. This step is a prerequisite for achieving real-time decision-making at the edge and reducing data transmission pressure.

[0020] In practice, the process begins by extracting time-domain and frequency-domain features from the raw sensor data. Specifically, for continuous raw data streams, the system uses a sliding window or fixed-size data frames for processing. Within each data frame, a series of statistics are calculated as time-domain features. These features include, but are not limited to: mean (reflecting the average level of the signal), variance or standard deviation (characterizing the dispersion and volatility of the signal), root mean square (RMS, reflecting the energy of the signal), peak value (reflecting the instantaneous maximum intensity of the signal), kurtosis (reflecting the sharpness of the signal waveform, which can be used for early fault detection), skewness (reflecting the symmetry of the signal waveform), waveform factor, and peak factor. These time-domain features directly capture the signal's temporal changes, vibration intensity, impact characteristics, and abnormal fluctuations, obtained in real-time through mathematical operations on the raw values ​​within the time window. Simultaneously, frequency-domain features are extracted from the raw sensor data. Specifically, to capture periodic fault information during equipment operation (such as bearing failure, gear wear, rotor imbalance, etc.), the time-domain signal needs to be converted to the frequency domain using Fourier transform (usually Fast Fourier Transform, FFT). In the frequency domain, equipment faults often manifest as an enhancement of specific frequency components. Extracted frequency domain features can include: dominant frequency, amplitude of each frequency component, energy spectral density, sideband information, harmonic content, bandwidth, etc. These frequency domain features can reveal structural anomalies, resonance phenomena, and frequency characteristics caused by periodic impacts within the equipment's internal components, and can be obtained in real-time through spectral analysis of the time-domain data. Furthermore, the time-domain and frequency-domain features are vectorized to obtain local feature vectors. That is, the various time-domain and frequency-domain features calculated in real-time are arranged and combined in a predefined order to form a fixed-length numerical vector, i.e., the local feature vector. For example, a local feature vector may contain RMS values, peak values, and the amplitude of the dominant frequency. This vector is an abstraction and representation of the original sensor data in the feature space, containing all the key information required by the model, and significantly reducing the data volume, meeting the input data requirements of lightweight AI models at the edge.

[0021] Specifically, S3 involves performing edge-side lightweight model inference on the local feature vector to obtain a preliminary local decision and a local confidence score. It should be understood that industrial sites have extremely high real-time requirements for responding to changes in equipment status; any diagnostic delay can lead to serious production accidents or property damage. Deploying the pre-trained lightweight model on edge devices such as PLCs allows the system to directly and quickly analyze the extracted local feature vectors at the data source, achieving millisecond-level real-time decision response and meeting the deterministic time requirements of industrial control. Furthermore, in addition to providing a preliminary judgment result, calculating the confidence score of the decision is crucial. It provides a basis for subsequent confidence-gated decisions, allowing the system to intelligently decide whether deeper cloud-based diagnosis is needed based on the reliability of the current judgment. This effectively balances the contradiction between real-time performance and diagnostic accuracy, avoiding the risk of misjudgment or missed judgment on the edge side.

[0022] Lightweight edge models refer to optimized and compressed machine learning or deep learning models deployed on PLCs or adjacent edge computing devices. These models are designed with computational efficiency and memory constraints in mind, typically featuring small model size and low computational complexity. They can quickly process local feature vectors and provide diagnostic results, such as decision trees, support vector machines, small neural networks, or quantized deep models.

[0023] In practice, the first step is to input the local feature vector into a lightweight AI model deployed on the PLC to obtain the raw score vector. This step aims to leverage edge computing power to quickly make preliminary, real-time judgments on the feature representation of the equipment status, thereby capturing potential anomalies at the first moment and meeting the requirements of industrial control for deterministic, low-latency responses. By obtaining the raw score vector, the system can provide an accurate, unnormalized internal evaluation benchmark for subsequent probability calculations and confidence assessments.

[0024] In this process, the local feature vector is transmitted as input to a lightweight AI model pre-deployed in the PLC. The PLC's processor or its integrated dedicated inference chip loads and executes this lightweight AI model. After receiving the local feature vector, the model performs high-speed, parallel mathematical operations on the input vector based on its internal parameters (such as weights and biases) and algorithmic structure (such as matrix multiplication and activation functions between neural network layers, and hyperplane calculation in SVM), resulting in a raw score vector containing multiple values. Each element of this vector represents the strength of the model's original numerical judgment that the input feature vector belongs to a predefined equipment state or fault category. For example, in a binary classification problem, the raw score vector may have only one value, whose sign or magnitude reflects the tendency to belong to a certain class; in a multi-class classification problem, it is a vector equal to the number of classes, with each value corresponding to the raw score of a class. This raw score vector is a direct intermediate result of the model's classification judgment, providing core data for subsequent generation of probability distributions and extraction of confidence scores. Next, the original score vector is processed to calculate the category probability distribution vector. It should be understood that the original score vector output by the lightweight AI model on the edge is usually an unnormalized numerical value. These only represent the model's relative tendency for the input to belong to each category; the numerical value is not directly equivalent to the probability and may even be negative. This raw form is not convenient for direct comparison between multiple categories, nor can it intuitively measure the certainty of the model's judgment. By transforming it into a category probability distribution vector, the model's output becomes more interpretable. Each value falls between 0 and 1, and the sum of the probabilities of all categories is 1. Each element in this vector represents the posterior probability that the model judges the input feature vector to belong to the corresponding category. This vector intuitively shows the confidence distribution of the model's predictions across all possible categories. This not only provides a clear basis for subsequent initial decisions—selecting the category with the highest probability as the decision result—but more importantly, it provides a quantified confidence score, i.e., the maximum probability value, thus indicating the reliability of the model's judgment and providing direct input for subsequent confidence-gated decisions.

[0025] In this process, the category probability distribution is calculated using the original score vector according to the following formula:

[0026] in, Let z be each element in the original fraction vector, where z is the original fraction vector.

[0027] Furthermore, preliminary decision-making and confidence extraction are performed on the probability distribution vector to obtain local preliminary decisions and local confidence scores. It should be understood that although the probability distribution vector obtained in the previous step has quantified the likelihood that the model considers the input feature vector to belong to each category, for practical applications, a single, clear decision result is usually needed to guide the next step. Moreover, having only a decision result without its reliability assessment is insufficient. Therefore, in the technical solution of this application, preliminary decision-making and confidence extraction are further performed on the probability distribution vector. The generated confidence score enables the system to intelligently judge the degree of certainty of the current edge-side decision. In practical applications, the decision is directly adopted only when the system has sufficiently high confidence; otherwise, a higher-level diagnostic process needs to be initiated, thus effectively balancing real-time performance and diagnostic accuracy, and avoiding blind decision-making or unnecessary resource consumption.

[0028] In this process, firstly, a maximum value indexing operation is performed on the probability distribution vector to obtain a local preliminary decision. Specifically, after receiving the probability distribution vector generated in the previous step, the system scans this vector to find the element with the largest value. The index corresponding to this maximum value (i.e., the position of the probability value in the vector) is identified. Since each index is pre-mapped to a specific device state or fault category, the category corresponding to this maximum value index is determined as the final result of this edge-side inference, i.e., the local preliminary decision. For example, if the probability distribution vector is [0.1, 0.85, 0.05], and index 1 corresponds to "normal", index 2 corresponds to "bearing wear", and index 3 corresponds to "motor overheating", then the system detects that 0.85 is the maximum value, its index is 2, and the local preliminary decision is "bearing wear". Secondly, the maximum probability value in the probability distribution vector is used as the local confidence score. That is, upon determining the initial local decision, the system directly uses the maximum probability value extracted in the previous step as the local confidence score for this decision. This value directly quantifies the model's confidence in the decision made. For example, in the above example, 0.85 is the "local confidence score," indicating that the model has an 85% confidence in the decision regarding "bearing wear." This score will serve as the direct input for subsequent confidence-gated decisions, determining the final processing path of this initial decision.

[0029] Specifically, in S4, a confidence-gated decision is performed on the local preliminary decision and local confidence score to obtain the edge final decision and the reported data packet. It should be understood that while the lightweight AI model on the edge side can provide real-time preliminary decisions and their corresponding confidence scores, if the judgment relies solely on the confidence score at a single point in time, the system will be highly susceptible to transient disturbances, such as electromagnetic noise, sensor jitter, or brief changes in operating conditions. This transient judgment mechanism leads to the system being overly sensitive to these non-persistent events, and consequently, overly sensitive to transient disturbances. For example, an isolated, non-persistent drop in confidence score caused by electromagnetic noise, sensor jitter, or brief changes in operating conditions (such as fluid splashing) may be incorrectly interpreted as a signal of equipment malfunction. Therefore, the system will overreact to such events, triggering unnecessary safety mode interventions and cloud data reporting. This not only increases the load on the network and cloud platform but may also cause minor fluctuations in production cycle time, potentially negatively impacting overall equipment efficiency (OEE). Relying solely on confidence scores at a single point in time completely misses the trend information inherent in the confidence scores over time, failing to effectively utilize the strong time-series correlation between confidence scores across multiple consecutive time points. This inherent correlation, compared to values ​​at a single time point, more realistically and stably reflects the true evolution trend of the equipment's state. To address these technical shortcomings, this application proposes an optimal selection mechanism for adaptive gating decision-making based on stable time confidence scores.

[0030] In practice, the first step is the maintenance and updating of the dynamic confidence sequence, aiming to transform isolated, instantaneous data points into continuous data sequences with historical context. During execution, the system maintains a first-in, first-out (FIFO) queue of length M in the edge-side memory to form the historical confidence sequence. Whenever the upstream model generates a new local confidence score, this score data is pushed to the head of the queue; simultaneously, if the queue length exceeds the preset value M, the oldest confidence score at the tail of the queue is automatically popped. In industrial scenarios, such as for a continuously running CNC machine tool, this step is equivalent to establishing a dynamic short-term memory cache for its prediction system, enabling it to capture the complete evolution of the model's decision determinism over a recent period. In this way, by continuously maintaining a data window containing the historical confidence history of the most recent M periods, the necessary and continuous input data is provided for subsequent trend quantification calculations, thus enabling the system to perceive the dynamic changes in decision confidence over time. Secondly, based on the stable time confidence score and local confidence score of the previous time point, the stable time confidence score of the current time point is calculated. This step aims to generate a quantitative indicator that can effectively smooth instantaneous noise and accurately reflect the inherent trend of decision confidence. During execution, the system uses the Exponential Weighted Moving Average (EWMA) algorithm, combining the latest local confidence score with the stable time confidence score calculated at the previous time point, to recursively calculate the stable time confidence score of the current time point; this process is expressed by the formula:

[0031] in, This represents the steady-state time confidence level at the current point in time. Represents the latest ConfidenceScore_Local (local confidence score) at the current moment; Represents t The steady-state time confidence calculated at time 1 already implies t. All historical information prior to time 1; It is an adjustable smoothing factor (ranging from 0 to 1) used to control the weight of new and old data in the calculation. This mechanism transforms the decision-making benchmark from a susceptible instantaneous value to a more robust metric that reflects the trend of the time series. .this The indicator not only quantifies the trend of confidence level changes but also constructs a metacognitive mechanism for the system to dynamically assess the stability of its own decisions, thereby endowing the decision-making system with inertia and enabling it to withstand short-term shocks. This allows for the generation of a stable-time confidence level indicator that effectively filters out occasional and meaningless confidence level fluctuations. This indicator more realistically reflects the continuous changes in the determinism of model judgments caused by gradual changes in equipment status, providing high-quality and robust input for subsequent decision-making.

[0032] Furthermore, based on the stable time confidence level and uncertainty threshold at the current time point, adaptive gating decision-making is applied to the local preliminary decision to obtain the final edge decision and reported data packet. That is, the stable trend index generated in the preceding steps is used to execute the final, more reliable decision gating logic. During execution, the system uses the calculated stable time confidence level at the current moment as the core judgment criterion, replacing the local confidence score in the original mechanism. Specifically, its core decision logic is as follows: only when the confidence trend is represented by the stable time confidence level... The system only determines that the model has entered a persistent and noteworthy low-confidence state when the value continues to decline and eventually falls below a preset uncertainty threshold. In this case, the system triggers a reporting mechanism to the cloud and executes corresponding local safety control strategies. In specific industrial scenarios, even if a sensor data point on a device experiences a sharp but brief jump causing a momentary drop in confidence, it is not sustained, and the stable-time confidence will not immediately fall below the threshold, thus preventing false alarms. In this way, by adopting trend-based judgment criteria, the intelligence and reliability of the decision-making logic are greatly improved. It enables the system to accurately distinguish between genuine, progressive equipment failures (typically manifested as a continuous decline in confidence) and random, insignificant field disturbances, ensuring that higher-level diagnostic and intervention processes are only initiated when a credible abnormal trend is detected.

[0033] Through the aforementioned optimization mechanism, the edge decision-making logic of predictive maintenance is transformed from a passive, event-driven model to a proactive, trend-intuitive model, fundamentally solving various problems caused by the temporal isolation of the original mechanism. Specifically, by introducing and utilizing stable time confidence as a core technical means, this mechanism achieves significant technical effects: First, it greatly enhances the robustness of the decision-making system in complex industrial environments filled with electrical noise and mechanical vibration, effectively avoiding misjudgments and erroneous operations caused by instantaneous data disturbances; second, by significantly reducing the frequency of unnecessary cloud reporting, this mechanism optimizes the utilization efficiency of network bandwidth and reduces the consumption of computing resources on the cloud platform; finally, by minimizing unnecessary production interventions and downtime, this mechanism can effectively guarantee and even improve the overall equipment efficiency (OEE), thereby constructing a more intelligent, efficient, and deeply integrated PLC-cloud collaborative predictive maintenance decision-making system that is closely aligned with the physical reality of the industrial manufacturing environment.

[0034] Specifically, S5 involves performing deep diagnostics and model optimization on the reported data packets via the cloud platform to obtain detailed diagnostic results and model update instructions from the cloud. It should be understood that while lightweight AI models on the edge side can meet real-time response requirements, their diagnostic accuracy and generalization ability to complex and unknown fault modes are limited. When the edge model has low confidence in its judgment of a device, or encounters previously unknown anomalies, relying solely on edge-side judgments may lead to misjudgments or omissions. In such cases, reporting relevant data to the cloud platform for deep diagnostics fully utilizes the cloud's vast computing resources, storage capacity, and more complex and comprehensive AI models for high-precision complex analysis, thereby obtaining detailed diagnostic results from the cloud to address anomalies that the edge side cannot effectively handle. Simultaneously, the cloud platform has the ability to aggregate and analyze massive amounts of historical data and device group data, enabling the discovery of new fault modes or optimization of existing model parameters. By feeding back model update instructions, local models can be adaptively updated, continuously improving the overall performance and intelligence level of the entire predictive maintenance system. This compensates for the shortcomings of a single computing node in high-precision complex analysis, thus resolving the conflict between real-time decision-making and diagnostic accuracy.

[0035] A cloud platform refers to a computing environment based on the internet, consisting of high-performance server clusters and distributed storage systems. It provides virtually unlimited computing and storage capabilities, enabling the deployment and operation of resource-intensive large-scale deep learning models and complex data analysis algorithms, and supporting centralized processing and analysis of massive amounts of device data.

[0036] In practice, the first step is to send a data packet containing key information such as raw data fragments, local feature vectors, preliminary decisions, and confidence levels to the cloud platform via industrial communication networks (such as 5G and Ethernet). Upon receiving this data, the cloud platform first verifies and decodes it, then securely stores it in a distributed database or data lake for subsequent analysis and long-term archiving. This data may also be integrated with historical operational data and manufacturing data from other equipment to enrich the dimensions of the analysis. Next, the cloud platform takes the received reported data packets as input and uses its deployed advanced AI models (e.g., large-scale deep neural network models, ensemble learning models, transfer learning models, etc.) and big data analytics tools for in-depth diagnosis. These models have ample computing resources and memory support in the cloud, enabling them to run more complex algorithms, undergo longer training periods, and access larger-scale labeled datasets. In this process, firstly, the reported data is integrated with operational data from other related devices, historical fault records, expert knowledge bases, etc., to identify fault patterns from a more macroscopic and detailed perspective. Then, through source tracing analysis and correlation mining, the root causes of equipment anomalies are explored in depth. Furthermore, combining the equipment's operational history and current status, the future development trend and remaining lifespan of the fault are predicted. Subsequently, unsupervised or semi-supervised learning methods are used to identify new abnormal behavior patterns that edge models may not have seen before. The generated detailed cloud-based diagnostic results include detailed fault reports, recommended maintenance measures, and more accurate equipment health assessments.

[0037] Ultimately, the cloud platform utilizes the refined diagnostic results generated by deep diagnostics, along with the continuously accumulating, labeled new measured data, to continuously optimize the model. This typically includes: Continuous training and fine-tuning: Once enough new and representative data is collected (especially cases where the edge model has high uncertainty or makes mistakes), the cloud platform will retrain or fine-tune the lightweight AI model deployed at the edge. This can correct the model's biases and improve its ability to identify new operating conditions and new failure modes. Model structure optimization: Based on the model's performance and data characteristics, the cloud platform may adjust the network structure of the edge model or explore new lightweight model architectures. Generating Model Update Instructions: The optimized model undergoes compression and quantization to ensure efficient operation on PLCs or edge devices. Finally, the cloud platform generates model update instructions containing new model parameters, model files, or relevant update strategies. These instructions are encapsulated and sent to the edge PLC via a secure communication link for local model adaptive updates.

[0038] Specifically, S6 involves a decision-making loop based on cloud-based detailed diagnostic results and model update instructions, along with adaptive updates to the local model to obtain the final control instructions and the updated local model. It should be understood that the initial lightweight edge-side model inference ensures real-time response, while the cloud platform's deep diagnostics provide high-precision fault analysis and more comprehensive insights. However, if the cloud-based diagnostic results cannot be effectively fed back to the field equipment to guide its actions, and the edge model cannot learn from this to improve its capabilities, the value of the entire collaborative architecture will be significantly diminished. Therefore, further decision-making loops based on cloud-based detailed diagnostic results and model update instructions are implemented to transform the advanced intelligence of the cloud into executable final control instructions, ensuring that the equipment can receive timely and effective intervention based on the most accurate judgment. Simultaneously, through adaptive updates to the local model, continuous learning and optimization in the cloud continuously improves adaptability to complex operating conditions and diagnostic accuracy, thus forming a self-optimizing, continuously evolving intelligent predictive maintenance system. This directly resolves the contradiction between high real-time performance and high precision requirements that a single node cannot simultaneously meet, achieving system intelligence and robustness.

[0039] In practice, firstly, the PLC or edge control unit receives detailed diagnostic results from the cloud platform. This result may be a detailed fault report or a direct indication of a problem with the equipment (e.g., "Severe bearing wear, immediate shutdown and inspection recommended"). Next, based on the received detailed cloud diagnostic results, the PLC combines preset control strategies, safety regulations, and the current operating conditions of the equipment to generate specific final control commands. For example, if the diagnostic results indicate that a critical component is about to fail, the command might be "immediately reduce speed"; if the diagnosis indicates wear on a minor component, it might trigger a maintenance work order to "schedule the next planned maintenance." This process may involve the PLC's internal logic program interpreting the cloud diagnostic information and converting it into specific I / O control signals, HMI (Human Machine Interface) alarm information, or work orders in the MES / ERP system. These instructions are executed immediately to prevent equipment failures or mitigate their consequences. Subsequently, the PLC or edge computing device receives model update instructions from the cloud platform. To ensure the security and stability of the industrial control system, these update instructions undergo rigorous integrity and security checks upon receipt to prevent malicious tampering or transmission errors. Once the checks pass, the edge device updates its local model according to the instructions. This process specifically includes: directly replacing the currently running old model with the new model file optimized by the cloud platform; if the instructions only contain updated model parameters, the edge device loads these new parameters into the existing model architecture; and for rule-based or hybrid models, the update instructions may adjust, add, or delete diagnostic rules. After the model update is complete, the system smoothly switches to the updated local model, meaning that future sensor raw data will be used for inference based on this new model, thereby improving the intelligence and decision-making capabilities of the edge side and forming an optimized closed-loop chain.

[0040] In summary, the predictive maintenance decision-making method based on PLC and cloud platform collaboration according to the embodiments of this application is explained. It deploys a lightweight AI model at the PLC edge to output preliminary decisions and decision confidence scores in real time, and performs adaptive gating decisions based on these to obtain reported data packets. Furthermore, through a decision-making closed-loop mechanism, the local model is adaptively updated, thereby achieving a dynamic balance between real-time performance and accuracy. This effectively resolves the inherent contradiction between real-time decision-making and diagnostic accuracy in predictive maintenance. The rapid response capability at the edge ensures timely processing of equipment status changes, while confidence gating avoids the limitations of single-point judgment by the edge model, enabling on-demand in-depth diagnosis and significantly improving diagnostic accuracy and system robustness.

[0041] Furthermore, a predictive maintenance decision-making system that integrates PLC and cloud platform is also provided.

[0042] Figure 3 This is a block diagram of a predictive maintenance decision-making system that integrates a PLC and a cloud platform according to an embodiment of this application. Figure 3 As shown, the predictive maintenance decision system 300, which integrates a PLC and a cloud platform according to an embodiment of this application, includes: a raw data acquisition module 310 for acquiring raw sensor data; a real-time feature extraction module 320 for extracting features from the raw sensor data in real time to obtain a local feature vector; an edge-side lightweight model inference module 330 for performing edge-side lightweight model inference on the local feature vector to obtain a local preliminary decision and a local confidence score; a confidence-gated decision module 340 for performing confidence-gated decision on the local preliminary decision and the local confidence score to obtain an edge final decision and a reported data packet; a cloud platform deep diagnosis and model optimization module 350 for performing cloud platform deep diagnosis and model optimization on the reported data packet to obtain a cloud-based fine diagnosis result and a model update instruction; and a decision closure and local model adaptive update module 360 ​​for performing decision closure and local model adaptive update based on the cloud-based fine diagnosis result and model update instruction to obtain a final control instruction and an updated local model.

[0043] As described above, the predictive maintenance decision system 300 for PLC and cloud platform collaboration according to embodiments of this application can be implemented in various wireless terminals, such as servers with predictive maintenance decision algorithms for PLC and cloud platform collaboration. In one possible implementation, the predictive maintenance decision system 300 for PLC and cloud platform collaboration according to embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the predictive maintenance decision system 300 for PLC and cloud platform collaboration can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the predictive maintenance decision system 300 for PLC and cloud platform collaboration can also be one of many hardware modules of the wireless terminal.

[0044] Alternatively, in another example, the predictive maintenance decision system 300 that coordinates the PLC with the cloud platform and the wireless terminal can also be separate devices, and the predictive maintenance decision system 300 that coordinates the PLC with the cloud platform can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0045] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A predictive maintenance decision-making method that integrates PLC and cloud platform, characterized in that, include: Acquire raw sensor data; Real-time feature extraction of raw sensor data to obtain local feature vectors; Perform edge-side lightweight model inference on local feature vectors to obtain local preliminary decisions and local confidence scores; Confidence-gated decisions are made on the local preliminary decision and local confidence score to obtain the edge final decision and the reported data packet; Perform in-depth cloud platform diagnostics and model optimization on the reported data packets to obtain detailed diagnostic results and model update instructions from the cloud. Based on the detailed diagnostic results and model update instructions from the cloud, a decision-making closed loop is performed, and the local model is adaptively updated to obtain the final control instructions and the updated local model.

2. The predictive maintenance decision-making method for PLC and cloud platform collaboration according to claim 1, characterized in that, Real-time feature extraction from raw sensor data to obtain local feature vectors includes: Extract time-domain and frequency-domain features from raw sensor data; The time-domain and frequency-domain features are vectorized to obtain local feature vectors.

3. The predictive maintenance decision-making method for PLC and cloud platform collaboration according to claim 1, characterized in that, Lightweight edge-side model inference is performed on the local feature vectors to obtain local preliminary decisions and local confidence scores, including: Input the local feature vector into a lightweight AI model deployed on the PLC to obtain the original score vector; The probability distribution vector is obtained by calculating the category probability distribution of the original score vector; Preliminary decision-making and confidence extraction are performed on the probability distribution vector to obtain local preliminary decision and local confidence score.

4. The predictive maintenance decision-making method for PLC and cloud platform collaboration according to claim 3, characterized in that, Calculating the category probability distribution of the original score vector to obtain a probability distribution vector includes: calculating the category probability distribution of the original score vector using the following formula, where the formula is: in, Let z be each element in the original fraction vector, where z is the original fraction vector.

5. The predictive maintenance decision-making method for PLC and cloud platform collaboration according to claim 3, characterized in that, Preliminary decision-making and confidence extraction are performed on the probability distribution vector to obtain local preliminary decisions and local confidence scores, including: Perform a maximum indexing operation on the probability distribution vector to obtain a preliminary local decision; The maximum probability value in the probability distribution vector is used as the local confidence score.

6. The predictive maintenance decision-making method for PLC and cloud platform collaboration according to claim 1, characterized in that, Confidence-gated decision-making is performed on the local preliminary decision and local confidence score to obtain the edge final decision and the reported data packet, including: Calculate the stable time confidence score at the current time point based on the stable time confidence score at the previous time point and the local confidence score; Based on the stable time confidence and uncertainty threshold at the current time point, adaptive gating decision is performed on the local preliminary decision to obtain the edge final decision and the reported data packet.

7. The predictive maintenance decision-making method for PLC and cloud platform collaboration according to claim 6, characterized in that, Based on the stable time confidence score and the local confidence score at the previous time point, the stable time confidence score at the current time point is calculated, including: calculating the stable time confidence score at the current time point using the following formula, where the formula is: in, This represents the steady-state time confidence level at the current point in time. Represents the local confidence score; Represents t The stationary time confidence level calculated at time 1; It is an adjustable smoothing factor.

8. A predictive maintenance decision-making system that integrates PLC and cloud platform, characterized in that, include: Raw data acquisition module, used to acquire raw data from the sensor; The real-time feature extraction module is used to extract features from the raw sensor data in real time to obtain local feature vectors. The edge-side lightweight model inference module is used to perform edge-side lightweight model inference on local feature vectors to obtain local preliminary decisions and local confidence scores. The confidence gating decision module is used to perform confidence gating decisions on the local preliminary decision and local confidence score to obtain the edge final decision and the reported data packet; The cloud platform deep diagnostics and model optimization module is used to perform deep diagnostics and model optimization on the reported data packets to obtain detailed diagnostic results and model update instructions in the cloud. The decision-making closed-loop and local model adaptive update module is used to perform decision-making closed-loop and local model adaptive updates based on the cloud-based fine diagnostic results and model update instructions to obtain the final control instructions and the updated local model.

Images