Self-adaptive predictive maintenance system and method for industrial equipment

Abstract

The invention relates to the technical field of industrial equipment intelligent operation and maintenance, and discloses a self-adaptive predictive maintenance system and method for industrial equipment. A physical mechanism and neural residual parallel mixed digital twinborn model is constructed, predictive deviation is explicitly modeled into a residual, and through a meta-learning technology, the prediction deviation and the residual are subjected to prediction prediction; a twin mother set with universal knowledge is trained offline from historical data of an equipment group, for new equipment, a small amount of initial data can be used for rapid fine adjustment based on the mother set, personalized twin is efficiently generated, a residual sequence is continuously monitored in an online diagnosis stage and a system, probability modeling is performed on health residual distribution, and the probability modeling result is obtained. Sensitive detection of early and unknown faults is achieved, and after individual twins obtain new knowledge through self-adaptive learning, the knowledge is shared through a federated network in a model updating quantity mode and used for iteratively optimizing a global twinborn mother set, and intelligent population evolution between devices is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent operation and maintenance of industrial equipment, and specifically relates to an adaptive predictive maintenance system and method for industrial equipment. Background Art

[0002] Predictive maintenance of industrial equipment is a key link in modern industrial production. Its core goal is to predict potential failures in advance by analyzing equipment operating data, thereby achieving a shift from traditional scheduled maintenance or post-failure maintenance to a condition-based, more cost-effective, situation-based maintenance model.

[0003] Existing technologies for achieving predictive maintenance can be categorized into two main approaches: physical model-based analysis and data-driven analysis. Physical model-based methods describe the operating mechanisms of equipment by establishing mathematical equations (such as kinetic and thermodynamic equations). While these methods offer good interpretability, their drawback is that the real-world industrial environment is extremely complex, making accurate physical modeling difficult and leading to significant discrepancies between the model and reality.

[0004] Therefore, data-driven approaches, particularly machine learning and deep learning methods, which have emerged in recent years with the development of artificial intelligence, have become mainstream in the field of predictive maintenance. These methods automatically map equipment status to faults by learning from large amounts of historical operating data. However, despite their success in many scenarios, data-driven approaches still face a series of deep and interconnected technical bottlenecks in practical industrial applications.

[0005] First, existing data-driven models typically require extensive, comprehensive historical data for training, particularly datasets rich in failure samples. This makes it difficult to effectively build models when a new piece of equipment is first put into production due to a lack of dedicated, long-term operational data, particularly failure data. This creates the so-called "cold start" problem. This often renders predictive maintenance systems inoperable during the initial stages, when equipment monitoring is most needed.

[0006] Furthermore, even if a model is successfully deployed initially, its performance will be severely tested over time. As a dynamically evolving physical entity, industrial equipment's performance will slowly but irreversibly change due to ongoing wear, aging, and material fatigue. Its operating conditions may also drift. However, once trained and deployed, most existing predictive models have their parameters fixed, becoming static entities. These models are unable to proactively adapt to this "concept drift" of the equipment's physical characteristics, leading to a gradual decline in prediction accuracy and ultimately a disconnect from the equipment's true state.

[0007] Furthermore, in industrial scenarios with large numbers of similar devices, existing maintenance strategies often analyze each device as an independent entity, creating information silos. When a new fault occurs in one device and is successfully diagnosed, this valuable experience and knowledge is difficult to systematically and automatically transfer to other devices to improve the "immunity" of the entire device population. This lack of mechanisms for knowledge sharing and collaborative evolution across the population results in wasted experience and repeated learning costs, limiting the overall improvement of the maintenance system's intelligence level.

[0008] At the same time, many existing predictive maintenance systems have limitations in their fault identification capabilities. Most are trained to recognize a predefined, known fault library. These systems often fail to recognize novel or rare fault modes that have never appeared in the training data, rendering them unable to provide effective early warnings. The fundamental reason is that these systems learn the "labels of specific faults" rather than the "essence of deviations from a healthy state."

[0009] Finally, with the increasing complexity of models, especially with the widespread use of deep neural networks, their "black box" nature presents new challenges. While these models may produce accurate predictions, their decision-making processes lack transparency, making it difficult to provide engineers with explanations that align with physical intuition and engineering logic. This lack of explainability makes it difficult for maintenance personnel to fully trust the model's diagnostic conclusions, leading to doubts when making high-risk, high-cost maintenance decisions and hindering the in-depth application of advanced technologies in critical industrial scenarios. Summary of the Invention

[0010] In response to the shortcomings of the existing technology, the present invention aims to solve the technical problems in the existing predictive maintenance technology, such as static models, poor adaptability, reliance on massive historical data, and inability to co-evolve among equipment groups.

[0011] To achieve the above objectives, the present invention is implemented through the following technical solutions: an adaptive predictive maintenance system for industrial equipment, comprising:

[0012] A hybrid digital twin modeling module is used to build a personalized digital twin representing the state of the target industrial equipment based on its physical mechanism and real-time operating data;

[0013] A personalized twin generation module is used to generate a twin master that can quickly adapt to the target industrial equipment based on the historical data of the equipment family, and to generate the personalized digital twin using the twin master and the initial operating data of the target industrial equipment;

[0014] An online diagnosis module, configured to monitor the real-time operating status of the target industrial equipment using the personalized digital twin and perform fault diagnosis based on the residual between the prediction result of the personalized digital twin and the actual operating data of the equipment;

[0015] The group knowledge evolution module is used to share new knowledge among other personalized digital twins in the device group when the personalized digital twin of any device learns new knowledge, so as to update the twin master.

[0016] Preferably, the hybrid digital twin modeling module is specifically used to:

[0017] Establishing a physical mechanism model based on the physical laws of the target industrial equipment;

[0018] Establishing a neural residual model for learning and compensating for the deviation between the physical mechanism model and the actual operating state of the target industrial equipment;

[0019] The output of the physical mechanism model is combined with the output of the neural residual model to form the prediction result of the personalized digital twin.

[0020] Preferably, the personalized twin generation module is specifically used to:

[0021] Offline, a meta-learning method is used to process historical data of the device population and train to obtain the twin master, which contains a set of optimized model meta-initial parameters.

[0022] Preferably, the personalized twin generation module is further specifically used to:

[0023] Online, the model element initial parameters of the twin master are used as the initial parameters of the personalized digital twin of the target industrial equipment, and the initial parameters are fine-tuned using the initial operating data of the target industrial equipment to generate the final personalized digital twin.

[0024] Preferably, the fine-tuning process is achieved by the following formula:

[0025] ;

[0026] in, : represents the parameters of the personalized digital twin obtained after fine-tuning and ultimately applied to the target industrial equipment; : represents the model initial parameters obtained by meta-learning method; : represents the learning rate during fine-tuning of the personalized digital twin; : represents the gradient operator, which here refers to the calculation of the gradient of the loss function with respect to the initial parameters of the model element; : represents the loss function calculated on the initial operating data of the target industrial equipment; : Represents the neural residual model part in the hybrid digital twin model; : Represents the initial operating data of the target industrial equipment used for fine-tuning.

[0027] Preferably, the online diagnosis module is specifically used to: continuously calculate the residual sequence between the prediction result of the personalized digital twin and the real-time operation data of the target industrial equipment;

[0028] Establishing a healthy residual space based on the historical residual sequence of the target industrial equipment in a healthy state;

[0029] When the residual generated in real time deviates from the healthy residual space, it is determined that an abnormality has occurred and an early warning is issued.

[0030] Preferably, the group knowledge evolution module is specifically used to:

[0031] Identifying the model parameter updates generated when a personalized digital twin adapts to a new operating condition or a new failure mode as the new knowledge;

[0032] broadcasting the new knowledge to other personalized digital twins in the device population via the federated network;

[0033] The twin master that receives the new knowledge performs fusion update.

[0034] Preferably, the fusion update process is implemented by the following formula:

[0035] ;

[0036] in, : represents the initial parameters of the model element to be updated; : represents the meta-knowledge learning rate; : Represents a preset aggregation function that processes the update amount of model parameters; : represents the amount of model parameter updates extracted from the personalized digital twin of a device that has learned new knowledge.

[0037] A maintenance method for an adaptive predictive maintenance system for industrial equipment, comprising the following steps:

[0038] Based on the physical mechanism and real-time operating data of the target industrial equipment, a personalized digital twin representing the status of the equipment is constructed;

[0039] Based on the historical data of the device population, a twin master that can quickly adapt to the target industrial equipment is generated, and the personalized digital twin is generated using the twin master and the initial operating data of the target industrial equipment;

[0040] Using the personalized digital twin to monitor the real-time operating status of the target industrial equipment, and performing fault diagnosis based on the residual between the prediction result of the personalized digital twin and the actual operating data of the equipment;

[0041] And when the personalized digital twin of any device learns new knowledge, the new knowledge is shared among other personalized digital twins in the device group to update the twin master.

[0042] The present invention provides an adaptive predictive maintenance system and method for industrial equipment. It has the following beneficial effects:

[0043] 1. This invention effectively solves the "cold start" problem of new industrial equipment, where it is difficult to establish effective predictive models due to a lack of historical data, by introducing a meta-learning-based twin master and combining it with an online rapid adaptation mechanism. Rather than learning from scratch for each device, the system first extracts universal prior knowledge from the device population offline and solidifies it in the twin master. When a new device is connected, only a minimal amount of its initial operating data is needed to quickly fine-tune the master, generating a highly personalized, exclusive digital twin. This enables predictive maintenance capabilities to be activated early in the device lifecycle.

[0044] 2. This invention achieves continuous adaptive tracking of device state changes throughout its lifecycle by constructing a hybrid digital twin model that combines physical mechanisms with neural residuals. This hybrid model structure explicitly decouples the device's ideal behavior from its actual deviations. By continuously monitoring and learning the dynamic evolution of these deviations, the system continuously adjusts its data-driven components to ensure that the digital twin remains synchronized with state drift caused by factors such as wear and aging in the physical entity, thereby overcoming the drawback of traditional static models that gradually fail over time.

[0045] 3. This invention establishes a group knowledge evolution mechanism based on a federated network, enabling the collaborative evolution of the intelligence level of the entire device group, greatly improving the system's robustness and knowledge reuse efficiency. When any personalized digital twin successfully responds to a new failure mode through adaptive learning, the core knowledge it has learned can be securely broadcast to the entire network without leaking any original production data, and used to optimize the global twin master, achieving the effect of "single learning, group benefit" and accelerating the entire system's ability to respond to unknown risks.

[0046] 4. This invention utilizes an online diagnostic method based on residual space, significantly improving the sensitivity of detecting unknown and early-stage weak faults. This method does not rely on predefined specific fault characteristics. Instead, it defines a "healthy state space" by modeling the probability distribution of deviations between the physical model and reality. Any early anomaly that deviates from the inherent characteristics of the equipment will cause the residual vector to deviate from this healthy state space. The system captures these statistical deviations and triggers early warnings, thereby discovering new fault modes that have never appeared in historical data.

[0047] 5. The prediction results of this invention possess greater engineering interpretability and reliability, overcoming the "black box" nature of purely data-driven models. Because the core of the hybrid digital twin model is based on clear physical mechanisms, its predictive behavior has a solid theoretical foundation. When a fault occurs, analysis of residual patterns can more easily link it to anomalies in specific physical parameters, assisting engineers in inferring the root cause of the fault, rather than simply providing vague conclusions based on data correlations, thereby enhancing the credibility of maintenance decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a system architecture diagram of the present invention;

[0049] Figure 2 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] Please see the attached Figure 1-2 , an embodiment of the present invention provides an adaptive predictive maintenance system for industrial equipment, including:

[0052] A hybrid digital twin modeling module is used to build a personalized digital twin representing the state of the target industrial equipment based on its physical mechanism and real-time operating data;

[0053] In this embodiment, the core task of the hybrid digital twin modeling module is to create a personalized mathematical representation for each target industrial device that accurately reflects its current operating state. This model is not a single, monolithic structure, but rather an innovative parallel fusion of two different models, each performing its own specific functions and complementing each other.

[0054] Specifically, the hybrid digital twin model includes a physical mechanism model ( ) and a neural residual model ( ).

[0055] Physical mechanism model ( ), which serves as the "skeleton" of the entire digital twin. This model is based on clear, recognized first principles of physics. For example, for an electric motor, it can be constructed based on its electromagnetic, thermodynamic, and rotor dynamics equations. The model inputs are the device's operating instructions and measurable environmental parameters, and its output is the theoretical response values ​​of the sensors of each key component of the device under ideal operating conditions. Therefore, the physical mechanism model provides a stable and interpretable foundation for predictions of the entire system, ensuring that the model does not deviate from the basic laws of physics under any circumstances.

[0056] Its mathematical form can be expressed as:

[0057]

[0058] In this formula, Represents a set of time series data input to the model, such as the speed command, load, ambient temperature and other characteristics of the equipment. Symbol It represents a set of physical parameters included in the physical model, such as the elastic modulus of the material, the mass of the component, the thermal conductivity, etc. Some of these parameters are known constants, and some can be calibrated according to the equipment manual or preliminary experiments. is the sequence of sensor readings predicted by the model under ideal conditions.

[0059] However, in real industrial environments, the behavior of equipment is far more complex than idealized physical equations. In order to capture all the complex and nonlinear dynamic characteristics that cannot be described by physical mechanism models, this embodiment introduces a neural residual model ( ).

[0060] It's worth noting that the neural residual model isn't designed to directly learn the complete behavior of a device. Instead, it's specifically designed to learn and compensate for the deviations, or residuals, between the physical mechanism model and the real world. These residuals encompass the combined effects of a complex set of factors, including component wear, material fatigue, unmodeled friction, measurement noise, and various random perturbations. By focusing the data-driven model on this learning component, the complexity of the learning task is greatly reduced, and its output has a clear physical meaning: a correction to the ideal model.

[0061] In a preferred embodiment, the neural residual model can adopt a neural network structure with strong temporal information processing capabilities, such as a long short-term memory network (LSTM) or a gated recurrent unit (GRU). Its mathematical form can be expressed as: In this formula, input Keep it consistent with the physical mechanism model to ensure that both make judgments based on the same working condition information. Represents all the weights and bias parameters in the neural network that need to be trained through data. Its output This is the residual vector sequence predicted by the model.

[0062] Ultimately, the hybrid digital twin modeling module in this embodiment linearly sums the outputs of the two parallel models to produce a precise prediction of the device sensor readings. This fusion approach ensures that the model's predictions retain the stability and interpretability of the physical mechanism model while also achieving high-precision adaptation to the complexities of the real world through the neural residual model.

[0063] Final prediction output of the personalized digital twin The complete expression is:

[0064]

[0065] The hybrid digital twin model constructed in this way not only provides highly accurate state predictions but, more importantly, quantifies the deviation (residual) between the physical model and reality as a clear, analyzable object. This explicitly modeled residual sequence becomes a valuable source of information for subsequent online monitoring, fault diagnosis, and adaptive learning, laying a solid foundation for the subsequent technical steps of this invention.

[0066] A personalized twin generation module is used to generate a twin master that can quickly adapt to the target industrial equipment based on the historical data of the equipment family. It also uses the twin master and the initial operating data of the target industrial equipment to generate a personalized digital twin.

[0067] In this embodiment, after building the aforementioned hybrid digital twin model infrastructure, the core task of this invention shifts to efficiently and accurately generating unique, personalized twin parameters for each specific piece of industrial equipment. To achieve this goal, the present invention provides an innovative personalized twin generation module. Rather than relying on traditional, isolated training for a single piece of equipment, this module employs a two-stage paradigm combining offline meta-learning and online rapid adaptation.

[0068] In this embodiment, this two-stage paradigm aims to fundamentally solve two major problems in industrial applications: first, the "cold start" problem when new equipment is initially online, that is, the inability to establish an effective model due to the lack of its own historical failure data; second, the "concept drift" problem caused by long-term operation wear or changes in operating conditions of the equipment, that is, the gradual failure of the original model.

[0069] The first stage of the personalized twin generation module involves offline twin master training. The goal of this stage is not to train a "mean model" that applies universally to all devices, but rather to develop a "learning capability" through meta-learning, ultimately producing optimized model initial parameters, also known as the "twin master." This master itself is not directly used for prediction, but it incorporates universal prior knowledge learned from the entire device population, serving as an efficient starting point for rapidly customizing a unique model for any individual device.

[0070] In a preferred embodiment, the specific implementation steps of this stage are as follows:

[0071] First, data preparation is performed. A large number of operating data segments with varying lengths are collected from a group of equipment, such as all centrifugal pumps of the same model in multiple factories. These data segments are structured into a series of independent learning tasks. . Every task Each of them represents the behavior pattern of a specific device within a specific time window, thus ensuring that the training data covers the widest possible range of individual differences, wear levels, and operating conditions. The data is further divided into support set and query sets , the former is used to simulate the rapid adaptation process of the model, and the latter is used to evaluate the performance of the model after adaptation.

[0072] Secondly, perform meta-learning optimization. In this embodiment, the model-independent meta-learning (MAML) algorithm is preferably used to optimize the neural residual model in the aforementioned hybrid digital twin model. The training process follows a two-layer loop optimization structure:

[0073] In the inner loop, the system simulates the process of deploying the model to a new device. For each task in the training set , the system starts from the current meta parameters Starting from, only using the support set data of the task Calculate the loss and perform one or more gradient descents to obtain a parameter temporarily optimized for this specific task The mathematical expression of this process is:

[0074]

[0075] in, is the learning rate of the inner loop, It's on a mission The loss function calculated on the support set of , such as mean squared error.

[0076] In the outer loop, the goal of the meta-learning algorithm is to optimize the meta-parameters itself, making it a better "starting point". To do this, the system evaluates all of these temporarily optimized models In their respective querysets Then, the performance loss of all tasks is accumulated, and the gradient is calculated based on this total loss to update the meta-parameters The mathematical expression of this process is:

[0077]

[0078] in, is the learning rate of the external meta-update. By repeating this internal and external cycle, is continuously optimized until it is such that any fine-tuning process starting from it can quickly reach the desired performance.

[0079] The second phase of the personalized twin generation module is online instantiation and rapid adaptation. This phase is activated when a new target industrial device is connected to the system.

[0080] First, instantiation is performed. The edge computing unit deployed on the new device side downloads the offline trained "twin master" from the central platform, that is, the optimized initial parameters. , and use it to initialize the parameters of the neural residual part in the local hybrid digital twin model.

[0081] Secondly, it conducts online rapid adaptation. The system uses the small amount of data collected in real time during the initial operation of the new device as its exclusive support set. , then, a fine-tuning process identical to the inner loop of meta-learning is performed on the edge.

[0082]

[0083] In this formula, These are the highly personalized twin model parameters that are unique to the new device and are ultimately obtained through this rapid adaptation process.

[0084] Through this two-stage design, the present invention's personalized twin generation module transforms the traditional, lengthy, independent data accumulation and model training process for each device into a one-time, reusable, offline learning of common knowledge and highly efficient online customization. This enables the present invention's predictive maintenance capabilities to be rapidly extended to new devices, and also enables efficient self-correction through this rapid adaptation mechanism when device characteristics change.

[0085] An online diagnostic module, which uses personalized digital twins to monitor the real-time operating status of target industrial equipment and performs fault diagnosis based on the residual between the personalized digital twin's prediction results and the actual operating data of the equipment;

[0086] In this embodiment, after the personalized twin generation module successfully generates and deploys a unique, highly adaptable, personalized digital twin for the target industrial equipment, the system transitions to a continuous online diagnostic phase. This invention provides a sophisticated online diagnostic module whose core purpose is to provide 24 / 7 real-time monitoring of the equipment's health status and to provide sensitive and reliable early warnings of faults.

[0087] In this embodiment, the design concept of the online diagnostic module represents a significant breakthrough from the traditional alarm approach, which relies on fixed thresholds for raw sensor signals. Instead, this module focuses diagnostics on a deeper, more informative subject: the residual between the personalized digital twin's predictions and the device's actual operating data. This residual-based diagnostic approach effectively filters out interference caused by normal operating fluctuations, significantly improving the ability to detect early, weak fault signals.

[0088] The workflow of the online diagnosis module can be broken down into the following interrelated steps:

[0089] First, continuous residual sequence generation is performed. During the online operation of the device, the personalized digital twin runs synchronously with its physical entity. The online diagnostic module collects the device's real sensor reading sequence in real time at high frequency. , and simultaneously receive data from the personalized twin model The module then calculates the vector difference between the two at each moment, thereby generating a multi-dimensional residual time series .

[0090] Its mathematical expression is:

[0091]

[0092] in, are the deployed, personalized twin parameters specific to the device. When the device is running in a healthy state, since the personalized twin can accurately reproduce its behavior, the residual sequence The performance should be a low-amplitude random noise that fluctuates around zero and has no specific structure. When some physical degradation or early failure begins to occur inside the device, its behavior will first deviate from the "ideal personalized" mode defined by the twin. This deviation will directly lead to the residual sequence Exhibits non-random, patterned drift or fluctuation.

[0093] Secondly, to accurately determine when residuals are "abnormal," the module must first model "normal." This embodiment, rather than employing simple upper and lower thresholds, constructs a "healthy residual space" that describes the complex distribution of residuals under healthy conditions. In a preferred embodiment, this space is learned using an unsupervised probabilistic generative model. Specifically, a variational autoencoder (VAE) is trained using a large amount of historical residual data generated by the device during well-defined healthy operating phases.

[0094] VAE consists of an encoder and a decoder Its training goal is to maximize the evidence lower bound (ELBO), and the objective function is:

[0095]

[0096] In this formula, is the input health residual vector, is its representation in the low-dimensional latent space, and are the network parameters of the encoder and decoder respectively, Represents KL divergence, which is used to measure the difference between the encoded distribution and a standard prior distribution (such as the standard normal distribution). Through this training process, the VAE learns how to compress the healthy residual vector into a compact latent space and accurately reconstruct the original healthy residual from this space. This mapping and reconstruction capability defined by the VAE together defines a high-dimensional "healthy residual space."

[0097] Finally, real-time anomaly detection and warning are performed. For each new residual vector generated in real time during online operation , the system inputs it into the pre-trained VAE. The degree of anomaly is quantified by its reconstruction error, that is, the difference between the input residual and the output residual after encoding and decoding is calculated. This difference is the anomaly score The internal logic is: if the real-time residual Belongs to the healthy distribution, VAE is able to reconstruct it well, resulting in anomaly scores Very low; on the contrary, if Including new, unseen failure modes, the VAE reconstruction effect will be poor, resulting in anomaly scores Significantly increased. When the anomaly score continuously exceeds a pre-set dynamic or static threshold When the device status deviates significantly, the online diagnosis module will determine that the device status has deviated significantly, thereby triggering an early fault warning.

[0098] Through the precise coordination of the above steps, the online diagnosis module of the present invention transforms the fault diagnosis problem from a fuzzy, experience-based judgment to an objective statistical inference process based on quantification of model deviations, thereby achieving sensitive discovery of unknown faults and reliable assessment of equipment health status.

[0099] And the group knowledge evolution module is used to share the new knowledge among other personalized digital twins in the device group when the personalized digital twin of any device learns new knowledge, so as to update the twin master.

[0100] In this embodiment, to enable the predictive maintenance system of the present invention to transcend passive adaptation to a single device and achieve proactive, collaborative intelligent evolution across the entire device population, the present invention provides a core group knowledge evolution module. This module's fundamental purpose is to break down information silos between devices and establish an efficient and secure knowledge-sharing network. This allows the valuable experience gained by any device in its battle against failures or new operating conditions to be rapidly refined, disseminated, and empowered across the entire device population.

[0101] In this embodiment, this module transforms the personalized digital twins distributed across various edge units from a series of isolated intelligent nodes into an interconnected intelligent organism capable of "sharing experience and co-evolving." Its specific workflow is meticulously designed into the following sequential, closed-loop steps.

[0102] First, this module needs to accurately identify and encapsulate "new knowledge". Not all parameter updates of the twin model have equal value. Therefore, the present invention defines a mechanism to determine whether an adaptive learning constitutes valuable "new knowledge" that can be shared. In a preferred embodiment, when a device (indexed by After a significant fault evolution or sudden change in working condition, the parameters of the neural residual model of the personalized twin (marked) change from the original stable state to Adaptively updated to a new stable state When , the system will calculate the incremental vector of this parameter update

[0103]

[0104] This parameter increments It is considered as the core product of this learning event. The system will further determine the norm of the incremental vector through preset criteria, such as Whether it exceeds a dynamic or fixed threshold determines the "significance" of this learning. Only significant updates are considered a successful acquisition of "new knowledge."

[0105] Once new knowledge is identified, the system will package it securely. Increment by parameter The core content is the knowledge package, which may also include some desensitized metadata, such as the type of abnormal pattern that triggered the learning. Crucially, this knowledge package does not contain any raw sensor data or specific operating instructions related to the device, thus technically ensuring data privacy and production security during the knowledge sharing process.

[0106] Secondly, this module performs the federation broadcast of new knowledge. Afterwards, it uploads the data package to the central platform. As the coordination hub of the entire knowledge network, the central platform verifies the received knowledge package, checking its format for compliance, and can even aggregate and generalize similar knowledge from multiple twins. Once verified, the central platform broadcasts the knowledge package to all other edge computing units of similar devices in the network that have subscribed to such updates.

[0107] Finally, and most importantly, this module performs the fusion update of group meta-knowledge. After that, the parameter is not simply incremented Instead of applying this new knowledge directly to its own personalized twin, which could destroy its already established, precise adaptation to its own characteristics, this new knowledge is used to optimize a higher-level, more fundamental object—the globally shared “twin master” parameters stored in a central platform. .

[0108] In a preferred embodiment, this fusion update process is implemented by a smooth update formula with a forgetting factor:

[0109]

[0110] In this formula, Represents a global meta-knowledge learning rate. It is a small value used to ensure the stable evolution of global meta-knowledge and avoid drastic fluctuations in the entire knowledge base due to single, local learning experiences. That is the core of the new knowledge received - parameter increment. It is a preset aggregation or transformation function, which can be a simple identity mapping or a function that weights or scales the knowledge package according to its metadata.

[0111] In this way, the "learned" failure response experience of a single device is abstracted, refined, and ultimately integrated into the "innate knowledge" of the entire device population. This means that the next time a new device is instantiated from this updated twin master, it will naturally possess the initial ability to recognize and adapt to this new failure mode. This mechanism establishes a positive, continuous evolutionary cycle, allowing the robustness and intelligence of the entire predictive maintenance system to continuously and systematically self-enhance over time and with the accumulation of experience.

[0112] Another embodiment of the present invention provides a maintenance method for an adaptive predictive maintenance system for industrial equipment, characterized by comprising the following steps:

[0113] S1. Based on the physical mechanism and real-time operation data of the target industrial equipment, a personalized digital twin representing the status of the equipment is constructed;

[0114] S2. Generate a twin master that can quickly adapt to the target industrial equipment based on the historical data of the equipment family. Then use the twin master and the initial operating data of the target industrial equipment to generate a personalized digital twin.

[0115] S3. Use the personalized digital twin to monitor the real-time operating status of the target industrial equipment and perform fault diagnosis based on the residual between the personalized digital twin's prediction results and the actual operating data of the equipment;

[0116] S4, and when the personalized digital twin of any device learns new knowledge, the new knowledge is shared among other personalized digital twins in the device group to update the twin master.

[0117] The method of this embodiment can be used to execute the above system embodiment. Its principles and technical effects are similar and will not be described in detail here.

[0118] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An adaptive predictive maintenance system for industrial equipment, characterized by: include: A hybrid digital twin modeling module is used to build a personalized digital twin representing the state of the target industrial equipment based on its physical mechanism and real-time operating data; A personalized twin generation module is used to generate a twin master that can quickly adapt to the target industrial equipment based on the historical data of the equipment family, and to generate the personalized digital twin using the twin master and the initial operating data of the target industrial equipment; An online diagnosis module, configured to monitor the real-time operating status of the target industrial equipment using the personalized digital twin and perform fault diagnosis based on the residual between the prediction result of the personalized digital twin and the actual operating data of the equipment; The group knowledge evolution module is used to share new knowledge among other personalized digital twins in the device group when the personalized digital twin of any device learns new knowledge, so as to update the twin master.

2. The system according to claim 1, wherein: The hybrid digital twin modeling module is specifically used to: Establishing a physical mechanism model based on the physical laws of the target industrial equipment; Establishing a neural residual model for learning and compensating for the deviation between the physical mechanism model and the actual operating state of the target industrial equipment; The output of the physical mechanism model is combined with the output of the neural residual model to form the prediction result of the personalized digital twin.

3. The system according to claim 1, wherein: The personalized twin generation module is specifically used to: Offline, a meta-learning method is used to process historical data of the device population and train to obtain the twin master, which contains a set of optimized model meta-initial parameters.

4. The system according to claim 3, characterized in that The personalized twin generation module is further specifically used for: Online, the model element initial parameters of the twin master are used as the initial parameters of the personalized digital twin of the target industrial equipment, and the initial parameters are fine-tuned using the initial operating data of the target industrial equipment to generate the final personalized digital twin.

5. The system according to claim 4, characterized in that The fine-tuning process is achieved through the following formula: ； in, : represents the parameters of the personalized digital twin obtained after fine-tuning and ultimately applied to the target industrial equipment; : represents the model initial parameters obtained by meta-learning method; : represents the learning rate during fine-tuning of the personalized digital twin; : represents the gradient operator, which here refers to the calculation of the gradient of the loss function with respect to the initial parameters of the model element; : represents the loss function calculated on the initial operating data of the target industrial equipment; : Represents the neural residual model part in the hybrid digital twin model; : Represents the initial operating data of the target industrial equipment used for fine-tuning.

6. The system according to claim 1, wherein: The online diagnosis module is specifically used to: continuously calculate the residual sequence between the prediction result of the personalized digital twin and the real-time operation data of the target industrial equipment; Establishing a healthy residual space based on the historical residual sequence of the target industrial equipment in a healthy state; When the residual generated in real time deviates from the healthy residual space, it is determined that an abnormality has occurred and an early warning is issued.

7. The system according to claim 1, wherein: The group knowledge evolution module is specifically used for: Identifying the model parameter updates generated when a personalized digital twin adapts to a new operating condition or a new failure mode as the new knowledge; broadcasting the new knowledge to other personalized digital twins in the device population via the federated network; The twin master that receives the new knowledge performs fusion update.

8. The system according to claim 7, characterized in that The fusion update process is implemented by the following formula: ； in, : represents the initial parameters of the model element to be updated; : represents the meta-knowledge learning rate; : Represents a preset aggregation function that processes the update amount of model parameters; : represents the amount of model parameter updates extracted from the personalized digital twin of a device that has learned new knowledge.

9. A maintenance method for an adaptive predictive maintenance system for industrial equipment according to any one of claims 1 to 8, characterized in that: The following steps are involved: Based on the physical mechanism and real-time operating data of the target industrial equipment, a personalized digital twin representing the status of the equipment is constructed; Based on the historical data of the device population, a twin master that can quickly adapt to the target industrial equipment is generated, and the personalized digital twin is generated using the twin master and the initial operating data of the target industrial equipment; Using the personalized digital twin to monitor the real-time operating status of the target industrial equipment, and performing fault diagnosis based on the residual between the prediction result of the personalized digital twin and the actual operating data of the equipment; And when the personalized digital twin of any device learns new knowledge, the new knowledge is shared among other personalized digital twins in the device group to update the twin master.

Images