Fault detection method and device based on digital twinning

Abstract

The invention discloses a fault detection method and device based on digital twinning. The method comprises the steps of obtaining physical entity operation data and extracting feature parameters; based on the characteristic parameters, generating simulation fault data through a diffusion model; and establishing a training data set based on the physical entity operation data and the simulation fault data, and constructing a complete closed-loop application for fault diagnosis. The device comprises a processor and a memory. The problem that fault data is insufficient in an existing fault diagnosis technology is solved, in the aspect of simulation data generation, the technology combining physical feature constraint and a depth generation model is adopted, the diversity and physical authenticity of generated data are improved at the same time, and the limitation of an existing simulation data generation method is overcome.

Description

Technical Field

[0001] This invention relates to the field of fault detection, and in particular to a fault detection method and apparatus based on digital twins. Background Technology

[0002] Ensuring stable equipment operation is crucial in key sectors such as industrial production, transportation, and energy supply. However, equipment failures are difficult to completely avoid, and traditional fault diagnosis methods face two core challenges: the scarcity of actual fault data and insufficient data diversity. These limitations severely restrict the adaptability and accuracy of diagnostic models in complex and ever-changing industrial scenarios.

[0003] First, traditional methods face the problem of scarce actual faults. Equipment faults are typically characterized by low incidence and high data collection costs. Taking large wind power equipment as an example, its normal operating time is long, resulting in extremely limited actual fault samples. Obtaining effective fault data often requires professionals to use specialized equipment for on-site collection and annotation, a time-consuming, labor-intensive, and costly process. Furthermore, some faults are sudden and non-reproducible, making it difficult to reproduce them under identical conditions to obtain more data. This data scarcity makes it difficult for traditional data-driven diagnostic models to obtain sufficient training samples, affecting their generalization ability and diagnostic accuracy.

[0004] Secondly, another serious problem is insufficient data diversity. The operating conditions of industrial equipment (such as speed, load, and environment) are complex and variable, causing their operating characteristics (such as vibration and temperature) to change accordingly. Traditional diagnostic models typically assume that training and testing data come from the same fixed operating condition distribution. However, in real-world applications, dynamic changes in operating conditions lead to data distribution drift, meaning the data distribution used during model training is inconsistent with the data distribution used in actual applications. This lack of diversity causes a significant drop in model performance when operating conditions change, making it difficult to maintain stable diagnostic capabilities.

[0005] Digital twin technology offers a new approach to solving the aforementioned problems by constructing virtual mappings of physical devices. Its core value lies in:

[0006] 1. Alleviate data scarcity: It can simulate the operation of equipment in various states (including fault states) in virtual space, generate a large amount of simulation data, and make up for the lack of actual fault samples.

[0007] 2. Enhanced data diversity: It can simulate the operation of equipment under different working conditions, loads and environments, and theoretically can cover a wider range of operating scenarios, providing more diverse training data sources for the model.

[0008] However, existing fault diagnosis methods based on digital twins still have significant limitations and have failed to fully realize their potential. These limitations correspond to the challenges faced by traditional methods. On the one hand, the fault simulation capabilities of existing technologies are relatively limited. Research is mostly focused on basic system modeling, and the generated simulation data (especially fault data) is often limited to preset, relatively simple fault modes, making it difficult to accurately simulate complex, rare, or compound faults. This results in limited diversity of generated simulation data, failing to comprehensively cover various fault modes and operating states that may occur in actual equipment under complex and changing conditions, directly corresponding to the problem of insufficient data diversity in traditional methods. On the other hand, a more critical challenge lies in the insufficient realism and fidelity of simulation data. Currently used simulation data generation methods, such as Generative Adversarial Networks (GANs), have inherent defects: GANs are prone to pattern collapse, resulting in data with a single pattern and poor diversity; more importantly, there is often a significant deviation (noise / distortion) between the generated simulation data and the data collected from real physical equipment, making it difficult to accurately simulate the subtle feature changes of real equipment under complex physical conditions. This "realism gap" makes models trained based on these simulation data perform poorly in practical applications and cannot effectively replace scarce real fault data.

[0009] Furthermore, current applications of digital twin technology in fault diagnosis tend to prioritize model development over practical application. Most research focuses on digital twin model construction and simulation algorithms themselves (conceptual and algorithmic levels), lacking practical exploration of building a closed-loop operating system using digital twin technology. How to effectively utilize the continuously generated simulation data from digital twins, combined with online data streams from real-world scenarios, to achieve adaptive updates to the diagnostic model, and ultimately translate this capability into a workable diagnostic system tailored to specific industrial scenarios, remains a weak link in current research.

[0010] Therefore, there is an urgent need for an innovative solution that can effectively combine the advantages of digital twin simulation, overcome the limitations of existing simulation data (diversity and realism), and realize the transformation from conceptual algorithms to practical system applications. Summary of the Invention

[0011] This invention provides a fault detection method and apparatus based on digital twins. It addresses the problem of insufficient fault data in existing fault diagnosis technologies. In terms of simulation data generation, it employs a technique combining physical feature constraints and deep generative models, aiming to simultaneously improve the diversity and physical realism of the generated data, thereby overcoming the limitations of existing simulation data generation methods. Details are described below:

[0012] Firstly, a fault detection method based on digital twins, the method comprising:

[0013] Acquire physical entity operation data and extract feature parameters; based on the feature parameters, generate simulation fault data through a diffusion model;

[0014] A training dataset is established based on physical entity operation data and simulation fault data, and a complete closed-loop application is constructed for fault diagnosis.

[0015] The closed-loop application is: a digital twin five-dimensional model that is deeply integrated and operates in a closed loop. The digital twin five-dimensional model includes: a physical entity model, a virtual entity model, a twin data model, a service model, and a connection model.

[0016] The process of generating simulation fault data through a diffusion model involves applying a diffusion model and combining it with a residual learning mechanism to generate high-fidelity simulation data. The diffusion model learns the residual distribution between real equipment data and initial simulation data generated based on basic features, thereby generating highly diverse simulation data.

[0017] The diffusion model employs a linear noise scheduler strategy, defining the noise addition process as follows:

[0018] ;in, Total diffusion steps; denoising net

[0019] The network adopts a U-Net structure and accepts noisy input signals. and conditional embedding Predicting noise residuals ;

[0020] The network consists of: 4 downsampling convolutional layers, each followed by a temporal embedding and conditional embedding fusion module; and 4 upsampling deconvolutional layers, which preserve high-frequency details through skip connections.

[0021] The learning residual mechanism is as follows:

[0022] The signal generation is transformed into a gradual correction of the noise distribution, expressed as: The goal is to predict the noise residual of the current step. ;

[0023] In data-scarce scenarios, the generation process is broken down into T steps, with each step requiring only a small correction to the current noise.

[0024] The key diagnostic features of bearing fault vibration signals are concentrated, and through explicit modeling, the model focuses on local changes in the high-frequency region.

[0025] The method further includes: a diffusion step. The following steps are executed iteratively:

[0026]

[0027] in, For standard noise difference, This is a random perturbation term used to increase generation diversity.

[0028] The service model employs Transformer Encoder and MCF Layer to identify and diagnose faults.

[0029] A second aspect is a fault detection device based on digital twins, the device comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to perform the method described in any of the first aspects.

[0030] Third aspect, a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of the first aspects.

[0031] The beneficial effects of the technical solution provided by this invention are:

[0032] 1) By adopting a digital twin data generation method based on the diffusion model and combining it with a learning residual mechanism, the core problems of scarce fault data and insufficient fidelity of simulation data are effectively solved. On the basis of ensuring authenticity through feature extraction (such as statistical mathematical models), complex residuals are efficiently learned and generated, which solves the shortcomings of existing simulation methods in terms of data diversity (covering complex / rare faults) and authenticity (high-fidelity restoration of physical features). It effectively generates high-quality, high-coverage fault data, greatly alleviates the problem of scarce actual fault samples, and improves the accuracy and generalization ability of diagnostic models.

[0033] 2) A five-dimensional digital twin model incorporating algorithms was constructed, and a closed-loop operation mechanism of "model construction -> data-driven -> diagnostic decision -> feedback optimization" was established. This system deeply integrates advanced diffusion model-generated data with fault diagnosis deep learning models, enabling dynamic adaptive updates of the diagnostic model in real-world scenarios. It successfully transforms digital twins and intelligent algorithms from a conceptual level into a workable industrial-grade fault diagnosis solution, solving the problem of "heavy on model, light on application". Attached Figure Description

[0034] Figure 1 The five-dimensional digital twin model constructed for this invention mainly includes a schematic diagram of five dimensions: physical entity, virtual entity, twin data, service, and connection.

[0035] Figure 2This is a schematic diagram of the method for constructing digital twin data based on a diffusion model proposed in this invention;

[0036] Figure 3 This is a schematic diagram of the core structure of the diffusion model in the data construction method proposed in this invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0038] Example 1

[0039] This invention is primarily used for intelligent fault detection and diagnosis of industrial equipment. Its core lies in constructing a digital twin five-dimensional model (physical entity model, virtual entity model, twin data model, service model, and connection model) that deeply integrates advanced algorithms and achieves closed-loop operation, thereby overcoming the bottlenecks of existing methods in data quality and practical application. The specific implementation follows these steps.

[0040] Step 101: Obtain physical entity operation data and extract feature parameters;

[0041] Collect the status data of the equipment during operation, and extract features from the raw vibration signals to obtain feature parameters that can reflect the key physical state of the equipment (i.e., statistical features such as operating frequency, amplitude, and phase), laying a realistic foundation for subsequent data generation;

[0042] Step 102: Based on the feature parameters extracted in Step 101, generate simulation fault data through a diffusion model;

[0043] The key innovation of this invention lies in the application of a diffusion model combined with a residual learning mechanism to generate high-fidelity simulation data. This diffusion model, based on extracted feature parameters, focuses on learning the residual distribution between real equipment data and initial simulation data generated from basic features. This enables the efficient generation of highly diverse simulation data covering complex, rare, and compound fault modes, while significantly improving the physical fidelity (realism) of the generated data. It effectively overcomes the mode collapse and distortion problems of traditional simulation methods, and the generated high-quality, high-coverage twin data greatly alleviates the scarcity problem of actual fault samples.

[0044] Step 103: Establish a training dataset based on physical entity operation data and simulation fault data, and build a complete closed-loop application system for fault diagnosis applications.

[0045] The high-fidelity data generation capability based on the diffusion model is deeply embedded into the five-dimensional digital twin model to construct a complete closed-loop application system. The physical entity operation data obtained in step 101 and the simulation fault data generated in step 103 are mixed at a preset ratio to form a training dataset for training the fault diagnosis model based on machine learning algorithms. During system operation, continuous data-driven and model optimization is performed: simulation data is dynamically generated / supplemented using the diffusion model, combined with real-time acquired equipment data, to train and optimize the fault diagnosis model; the trained fault diagnosis model performs fault diagnosis; and the diagnosis results and newly acquired data features are fed back to the digital twin model to update the virtual model state, optimize the simulation data generation process, and iteratively improve the fault diagnosis model. This closed-loop mechanism based on real-time interaction and data-driven operation of the digital twin ensures that the system can continuously optimize performance using the latest information and generated high-quality data, improving its adaptability and diagnostic reliability in dynamic industrial environments.

[0046] Example 2

[0047] The following description, with reference to specific examples and accompanying drawings, further illustrates the solution in Embodiment 1.

[0048] I. Constructing Digital Twin Data Based on a Diffusion Model:

[0049] This invention utilizes a diffusion model to construct digital twin data, addressing the difficulties encountered in practical data collection, enriching the dataset, and improving the model's performance and reliability. The methods described are summarized below. Figure 2 As shown.

[0050] 1. Physical entity data based on datasets

[0051] This invention uses a publicly available bearing fault dataset as the source of physical entity data. This dataset simulates the operating state of bearings under different fault types (including inner ring damage, outer ring damage, and rolling element damage) and fault sizes (i.e., 0.007 inches to 0.028 inches) using an experimental bench. Vibration signals from a drive-end accelerometer are collected at a sampling frequency of 12 kHz, covering various load conditions (i.e., 0 HP to 3 HP). This invention selects data including normal conditions and typical fault types (including single-point damage to the inner ring, outer ring, and rolling elements), with a uniform fault size of 0.021 inches and load conditions of 0 HP (light load) and 2 HP (medium load). Continuous vibration signals are divided into segments of a fixed window length (1024 sampling points), with each sample corresponding to a time segment. Each sample is labeled with the fault type (including: normal, inner ring, outer ring, and rolling element), fault location (clockwise position), and operating parameters (load and speed).

[0052] 2. Feature Extraction and Learning

[0053] 1) Temporal characteristics

[0054] The following statistical features are extracted from each signal segment:

[0055] a) Mean:

[0056] b) Root Mean Square (RMS):

[0057] c) Kurtosis:

[0058] d) Crest Factor:

[0059] 2) Frequency domain characteristics

[0060] a) Fast Fourier Transform (FFT): Calculates the signal spectrum and extracts the energy proportion of the main frequency components (such as the characteristic frequency of bearing failure).

[0061] b) Envelope spectrum analysis: High-frequency components of the modulation signal are extracted through Hilbert transform to enhance the identifiability of fault features.

[0062] 3) Time-frequency domain characteristics

[0063] The signal is decomposed into 16 sub-bands using wavelet packet transform (WPT), and the energy entropy of each sub-band is calculated.

[0064]

[0065] The above features are combined into a multi-dimensional vector, which serves as the input condition for the conditional diffusion model.

[0066] 3. Construction and Training of Conditional Diffusion Model

[0067] 1) Model Architecture Design

[0068] a) Input condition definition

[0069] The vibration signal features extracted from the dataset (time-domain statistics, frequency-domain energy distribution, and wavelet packet energy entropy) are used as conditional vectors to guide the diffusion model in generating simulation signals consistent with the actual data distribution. In the feature extraction stage, a 20-dimensional feature vector is constructed (i.e., 4-dimensional time-domain statistics + 4-dimensional frequency-domain dominant frequency energy + 12-dimensional wavelet packet subband energy entropy). A conditional encoder is then constructed to generate the initial simulation signal. A 1D convolutional neural network (CNN) is used to map the feature vectors to a 128-dimensional conditional embedding vector.

[0070]

[0071] b) Core components of the diffusion model

[0072] Using a linear noise scheduler strategy, the noise addition process is defined as follows:

[0073]

[0074] in, This represents the total number of diffusion steps.

[0075] The denoising network adopts a U-Net structure, with a noisy input signal. and conditional embedding Predicting noise residuals The network consists of: 4 downsampling convolutional layers (number of channels: 64→128→256→512), each followed by a time embedding and conditional embedding fusion module; and 4 upsampling deconvolutional layers (number of channels: 512→256→128→64), which preserve high-frequency details through skip connections.

[0076] 2) Residual learning strategy

[0077] Directly generating the complete signal may cause the model to ignore high-frequency details, while residual learning, by predicting the noise distribution, focuses more on the difference between the real data and the noise signal. The essence of residual learning is to decompose the complex generation task into a multi-step denoising process, transforming signal generation into a gradual correction of the noise distribution. Its core formula can be expressed as:

[0078]

[0079] The goal of the model is to predict the noise residual of the current step. Instead of directly outputting the complete signal.

[0080] In data-scarce scenarios (e.g., the CWRU dataset contains only a few hundred samples), direct generative models (e.g., GANs) are prone to mode collapse, manifesting as low diversity and high repetitiveness of generated samples. For example, the model might only be able to generate vibration signals for a single fault mode, failing to cover combinations of inner race, outer race, and rolling element damage. By decomposing the generation process into T steps (e.g., T=1000), each step only needs to predict a small correction to the current noise. This "divide and conquer" strategy reduces the complexity of each step, making it easier for the model to learn general patterns from limited data.

[0081] Key diagnostic features of bearing fault vibration signals are concentrated in high-frequency transient impacts (such as periodic pulses from inner race faults). These components manifest as abrupt amplitude changes in the time domain and correspond to harmonic energy at characteristic frequencies in the frequency domain. Traditional generative models (such as VAEs) tend to obscure these details due to the smoothing tendency of the mean square error (MSE) loss function. Noise residuals Essentially, this corresponds to subtle differences in the signal that are not masked by noise. This is achieved through explicit modeling. The model is forced to focus on local changes in high-frequency regions. Taking CWRU inner ring fault data as an example, the signal generated by residual learning can clearly retain the fault characteristic frequencies in the envelope spectrum. And its harmonics, while the harmonic energy attenuation rate of the directly generated model is over 40%. Let the real signal be... ( For a pure signal, (for noise), the diffusion model predicts The distribution, indirectly reconstructed Due to fault characteristics The energy is concentrated in a specific frequency band, and residual learning is equivalent to... The bandpass filter suppresses irrelevant noise interference.

[0082] 4. Simulation data generation and fusion

[0083] Backpropagation process

[0084] Sample pure noise signal from standard Gaussian distribution The length matches the actual vibration signal. Following the diffusion steps... The following steps are executed iteratively:

[0085]

[0086] in, For standard noise difference, This is a random perturbation term used to increase generation diversity. It is adjusted... The randomness of the fault generates multiple vibration modes under the same fault type (such as different impact intensities or noise interference).

[0087] In each step, the features are embedded (Time-frequency domain features from the CWRU dataset) and noise signals splicing, input to denoising network .

[0088] 5. Condition-controlled generation

[0089] By using the category of fault data as a conditional input, residual learning can precisely control the fault type and severity of the generated signal. For example, modifying the "fault size" feature value can generate a gradual signal ranging from slight wear (0.007 inches) to severe spalling (0.028 inches); modifying the fault label in the conditional vector cc (e.g., setting "inner ring fault" to 1 and "outer ring fault" to 0) or adjusting the load (0HP / 2HP) or speed feature value can generate simulated vibration data for different fault types and operating conditions.

[0090] The conditional diffusion model implicitly learns the physical mapping relationship between fault characteristics and vibration response. The time-domain statistics of the generated signal (such as kurtosis and peak factor) have an error of less than 5% compared with the input conditions, proving that it conforms to the laws of fault dynamics.

[0091] 6. Multimodal data fusion

[0092] Simulation data and real CWRU data were mixed in a 7:3 ratio to ensure the training set covered diverse operating conditions and failure modes, while balancing the diversity gain of simulation data with the distribution dominance of real data. A stratified sampling method was also used to balance failure types and match different operating conditions. Since rolling element damage samples accounted for only 5% of the CWRU dataset, while inner race failures accounted for 60%, the model underfitted the minority class. This was addressed by oversampling the minority class and controlling the generation of the majority class. Specifically, more simulation data of rolling element damage was generated, increasing its proportion from 5% to 20%; simulation data of inner race failures were generated in a 1:1 ratio to avoid overfitting. To ensure physical consistency, the operating condition distribution was matched. Specifically, simulation data was generated according to the real operating condition distribution (0HP:2HP = 1:4) to ensure the load conditions were consistent with the real environment; additional simulation data under 3HP load (accounting for 5%) was generated to cover high-load anomalies in actual industrial scenarios. The consistency of the operating condition distribution was verified using Kullback-Leible divergence (KL divergence). The results show that the operating condition distribution of the simulation data is highly matched with that of the real data (KL divergence <0.05 is excellent).

[0093] 7. Simulation quality verification

[0094] 1) Quantitative validation: Distribution consistency assessment

[0095] The consistency between virtual and physical entities is assessed by analyzing and calculating FID scores and Wasserstein distance.

[0096] The FID score extracts the time-frequency domain features (wavelet packet energy entropy, MFCC) of the virtual entity's output signal and the real CWRU data. These features are mapped to feature vectors using a pre-trained Inception network, and the FID distance between the two distributions is calculated. The Wasserstein distance is used to evaluate the similarity between the virtual entity's output and the real data in the frequency domain energy distribution by dividing the signal spectrum into multiple frequency bands and calculating the Wasserstein distance of the energy in each band. Verification shows that the virtual entity constructed using digital twin simulation data has an FID value of 12.4 and a Wasserstein distance of 0.08. An FID < 15 and a Wasserstein distance < 1 indicate a high degree of consistency between the virtual entity's output and the real data distribution.

[0097] 2) Qualitative verification: Physical consistency analysis

[0098] The consistency between virtual and physical entities is evaluated by comparing time-domain waveforms, spectral analysis, and separator tests. The output of the virtual entity and the real signal under the same operating conditions are selected, and time-domain waveforms are plotted to examine the periodicity, amplitude, and shape of the impact components. The envelope spectra of the virtual entity output and the real data are compared to examine fault characteristic frequencies (such as the inner ring fault frequency) and their harmonic energy proportions. A ResNet-18 fault classifier is trained using a mixed dataset (70% real data + 30% virtual entity output data); the classification performance is evaluated on an independent test set (real data only).

[0099] Example 3

[0100] The following describes the solutions in Examples 1 and 2 in more detail with specific examples: Constructing a five-dimensional digital twin model and a complete digital twin.

[0101] 1. Constructing a five-dimensional digital twin model

[0102] 1) Physical Entity Model (PE): Taking the physical bearing corresponding to the dataset as an example, its operating data, such as vibration signals and temperature signals, are collected by sensors to provide basic data for subsequent fault diagnosis.

[0103] 2) Virtual Entity Model (VE): Abstract statistical features are extracted from the collected bearing data to construct a statistical model that reflects the health status of the bearing. This serves as a data-driven virtual model for simulating and predicting the operating status of the bearing.

[0104] 3) Twin Data Model (DD): This model utilizes a diffusion model-based generative technique, combining the CWRU dataset with statistical features to generate a large amount of twin data. By learning the distribution and characteristics of the data, the diffusion model can simulate real-world fault conditions during data generation. It can also generate fault data that is difficult to obtain in practice, thereby expanding the quantity of fault data and improving its diversity and representativeness.

[0105] 4) Service Model (Ss): Develop a fault diagnosis algorithm and visualization platform based on digital twin data. This platform utilizes high-fidelity simulation data and advanced machine learning algorithms. In this embodiment, the Transformer Encoder and MCF Layer are employed to achieve rapid fault identification and diagnosis. Simultaneously, the platform employs modern Web technologies such as React and Flask to ensure user-friendliness and convenience, making the fault diagnosis process more intuitive and efficient.

[0106] 5) Connectivity Model (CN): Enables ubiquitous industrial interconnection between physical entities, virtual entities, services, and data, specifically including:

[0107] CN_PD: Collects fault data of the physical bearing through sensors and transmits it to the twin data to realize the interaction between the physical entity and the twin data.

[0108] CN_PV: Transmits the fault data of the physical bearing to the virtual entity for updating and correcting the virtual model, enabling interaction between the physical and virtual entities.

[0109] CN_PS: Transmits real-time data of the physical bearing to the service, enabling service updates and optimizations. The service provides operational guidance, professional analysis, and other results to users in the form of application software, allowing manual operation to control the physical bearing and realize interaction between the physical entity and the service.

[0110] CN_VD: Stores the simulation analysis data of virtual entities in twin data, providing richer data support for subsequent fault diagnosis and enabling interaction between virtual entities and twin data.

[0111] CN_VS: Transmits the simulation analysis results of virtual entities to the service, which then processes and analyzes these results to generate more accurate fault diagnosis results, enabling interaction between virtual entities and the service.

[0112] CN_SD: Stores the operational guidance, professional analysis, and other results generated by the service in the twin data, providing more comprehensive data support for subsequent fault diagnosis and decision-making, and realizing the interaction between the service and the twin data.

[0113] 2. Closed-loop formation and operation of a complete digital twin

[0114] Another core value of this invention lies in the construction of an engineerable digital twin system. Through a closed-loop architecture, the five-dimensional digital twin model is transformed into a physical system. The following details the process of building the system and implementing its application.

[0115] 1) Physical layer architecture

[0116] The physical entity layer of this invention is entirely built upon authoritative industrial datasets, obtained through a precision laboratory acquisition system. Specifically, a high-sensitivity SKF MB ER-16K accelerometer is used to capture vibration signals of the bearing under four typical operating conditions (corresponding to 0 to 3 horsepower loads) at a sampling frequency of 12kHz. The acquired data covers the bearing's normal state and seven typical fault modes (including different locations and degrees of damage for inner ring faults, outer ring faults, and rolling element faults). The original simulated vibration signals are converted by a 24-bit high-precision ADC and stored as standard MAT format data files, with an overall data signal-to-noise ratio ≥60dB, providing a high-quality physical world data foundation for constructing a high-fidelity virtual twin model.

[0117] 2) Virtual entity modeling

[0118] Virtual entities are constructed through a data-driven approach. Their core lies in extracting multi-dimensional statistical features from the operational data of physical entities and building mathematical models for real-time fault diagnosis based on these features.

[0119] Specifically, the virtual entity receives real vibration signals from the physical layer and first performs multi-dimensional feature extraction. This process analyzes the signals simultaneously in the time, frequency, and video domains: in the time domain, statistics including mean, root mean square, kurtosis, and peak factor are calculated; in the frequency domain, the energy proportion of bearing fault characteristic frequencies is extracted using Fast Fourier Transform, and envelope spectrum analysis is used to enhance the periodic impact characteristics; in the time-frequency domain, wavelet packet transform is used to decompose the signal into multiple sub-bands and calculate their energy entropy to capture the dynamic distribution characteristics of the signal energy.

[0120] Subsequently, based on all the features extracted above, a fault diagnosis mathematical model based on multi-threshold interlocking rules is constructed. This model performs a comprehensive logical judgment by combining the feature values ​​calculated in real time with preset diagnostic thresholds. For example, when the system simultaneously identifies that the kurtosis exceeds the threshold, the energy proportion of a specific fault frequency is abnormally increased, and the time-domain waveform exhibits periodic impacts, the model immediately outputs a deterministic diagnostic conclusion of "moderate fault in the outer ring".

[0121] To ensure long-term diagnostic accuracy, the virtual entity incorporates a self-correcting mechanism. This mechanism receives real-world feedback from the physical entity via the CN_PV connection. When diagnostic consistency declines due to device state drift, the system automatically triggers optimization updates to feature thresholds and rule logic, ensuring that the diagnostic model continuously converges with the actual health status of the physical entity.

[0122] 3) Construction of the digital twin layer

[0123] The system uses feature data extracted from the Virtual Entity Layer (VE) as the physical basis for data generation. It applies a residual learning data generation framework based on a diffusion model, using physical parameters from the feature repository as input. Through a step-by-step noise addition and denoising process, it focuses on learning the residual distribution between the real equipment feature data and the initial simulation features generated from the basic physical model. This mechanism enables the model to accurately capture minute shifts in feature parameters under complex operating conditions (such as abrupt changes in envelope spectrum harmonic phase caused by composite faults), feature combinations of rare fault modes (such as the concurrent superposition of features from rolling element spalling and cage fracture), and nonlinear distortions under physical boundary conditions (such as frequency modulation effects caused by high loads). During generation, the model is iteratively optimized through a 20-layer U-Net residual network, outputting a highly diverse simulation feature set with physical fidelity ≥95% (measured by a Wasserstein distance ≤0.1 in the feature space) and fault mode coverage of 92%. The generated data is dynamically injected into physical feature partitions via Kafka message queues according to operating condition tags (1750rpm / 1HP load), achieving online mixed updates in a 7:3 ratio.

[0124] 4) Industrial-grade delivery model of the service layer (Ss)

[0125] The service architecture adopts a microservice design, encapsulating the diagnostic engine as an independent container deployed on a Kubernetes cluster. A dynamic scaling mechanism supports a single node's diagnostic processing capacity of 200 times / second, effectively handling high-concurrency requests. Diagnostic results are transmitted to the factory's SCADA system via a RESTful API, automatically converted into standard IEC 61131-3 control commands (such as the "SET Motor_Speed=0" stop command) for precise equipment control. Workshop inspectors operate the system via a React progressive web application on tablets. This application integrates a WebGL engine to render a real-time 3D model of bearing damage, supporting gesture-based rotation and zoom interactions. The system workflow starts with an Nginx load balancer, distributing diagnostic requests to multiple diagnostic nodes (based on the Flask framework integrating multi-category feature processing (MCF) and LightGBM algorithms). Each node accesses the pre-trained diagnostic model cached in Redis in parallel, triggering a Bayesian network for root cause analysis after computation. The analysis results are proactively pushed to the React frontend via the WebSocket protocol, driving the WebGL engine to generate a 3D damage visualization model.

[0126] 5) System Integration

[0127] This system is designed to achieve an end-to-end integrated intelligent diagnostic and control closed loop. Its core value lies in the integration of microservice architecture, industrial control standards, and modern web visualization technology: a dynamically orchestrated diagnostic container cluster using Kubernetes ensures a high-concurrency real-time computing capability of 200 times per second; deep integration with the SCADA system based on RESTful interfaces allows diagnostic conclusions to be directly converted into IEC 61131-3 standard control commands (such as specific actions like equipment start / stop and parameter adjustment); and the workshop terminal leverages the offline availability of React PWA and the 3D rendering capabilities of WebGL to transform abstract bearing damage data into an interactive 3D model. The entire process begins with Nginx distributing requests, proceeds through parallel processing by distributed diagnostic nodes, accelerates model inference using Redis, and finally utilizes Bayesian network root cause analysis and a WebSocket real-time push mechanism to construct a seamless link from data input and intelligent analysis to visualization and control execution.

[0128] II. Continuing Conclusions

[0129] This invention constructs an industrial-grade digital twin system based on statistical feature-driven and diffusion residual learning. It builds a lightweight virtual entity model by integrating multi-dimensional time-frequency domain features. It innovatively adopts a diffusion model residual generation mechanism to cover 17 complex fault modes, breaking through the bottleneck of small sample data. It relies on a five-dimensional architecture to connect the entire chain from sensing acquisition, dynamic mixed data (physical and simulation 7:3), MCF intelligent diagnosis to control execution. It relies on the five-dimensional digital twin model to build a diagnostic system to complete the closed loop, forming the transformation of digital twin technology from laboratory algorithms to production line requirements.

[0130] Example 4

[0131] An apparatus comprising: a processor and a memory storing program instructions, wherein the processor invokes the program instructions stored in the memory to cause the apparatus to perform the following method steps of Embodiment 1:

[0132] Acquire physical entity operation data and extract feature parameters; based on the feature parameters, generate simulation fault data through a diffusion model;

[0133] A training dataset is established based on physical entity operation data and simulation fault data, and a complete closed-loop application is constructed for fault diagnosis.

[0134] The closed-loop application is: a digital twin five-dimensional model that is deeply integrated and operates in a closed loop. The digital twin five-dimensional model includes: a physical entity model, a virtual entity model, a twin data model, a service model, and a connection model.

[0135] Among them, generating simulation fault data through diffusion model is as follows: applying diffusion model and combining it with learning residual mechanism to generate high-fidelity simulation data. The diffusion model learns the residual distribution between real equipment data and initial simulation data generated based on basic features to generate highly diverse simulation data.

[0136] The diffusion model employs a linear noise scheduler strategy, defining the noise addition process as follows:

[0137]

[0138] in, The total number of diffusion steps; the denoising network adopts a U-Net structure, with the input signal containing noise. and conditional embedding Predicting noise residuals ;

[0139] The network consists of: 4 downsampling convolutional layers, each followed by a temporal embedding and conditional embedding fusion module; and 4 upsampling deconvolutional layers, which preserve high-frequency details through skip connections.

[0140] The residual learning mechanism is as follows:

[0141] The signal generation is transformed into a gradual correction of the noise distribution, expressed as: The goal is to predict the noise residual of the current step. ;

[0142] In data-scarce scenarios, the generation process is broken down into T steps, with each step requiring only a small correction to the current noise.

[0143] The key diagnostic features of bearing fault vibration signals are concentrated, and through explicit modeling, the model focuses on local changes in the high-frequency region.

[0144] This also includes: diffusion steps The following steps are executed iteratively:

[0145]

[0146] in, For standard noise difference, This is a random perturbation term used to increase generation diversity.

[0147] The service model employs Transformer Encoder and MCF Layer to identify and diagnose faults.

[0148] It should be noted that the device descriptions in the above embodiments correspond to the method descriptions in the embodiments, and the embodiments of the present invention will not be repeated here.

[0149] The execution entities of the aforementioned processor and memory can be devices with computing functions such as computers, microcontrollers, and single-chip microcomputers. In specific implementations, the embodiments of the present invention do not limit the execution entities and can select them according to the needs of actual applications.

[0150] Data signals are transmitted between the memory and the processor via a bus, which will not be elaborated upon in this embodiment of the invention.

[0151] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0152] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0153] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0154] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0155] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0156] Unless otherwise specified, the model numbers of the various devices in this embodiment of the invention are not limited, and any device that can perform the above functions is acceptable.

[0157] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0158] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for fault detection based on digital twinning, characterized in that, The method comprises: Obtaining physical entity operation data and extracting feature parameters; based on the feature parameters, generating simulation fault data through a diffusion model; Based on the physical entity operation data and the simulation fault data, a training data set is established, and a complete closed-loop application is constructed for fault diagnosis.

2. The fault detection method based on digital twinning according to claim 1, characterized in that, The closed-loop application is a digital twin five-dimensional model that adopts deep fusion and realizes closed-loop operation, and the digital twin five-dimensional model comprises a physical entity model, a virtual entity model, a twin data model, a service model and a connection model.

3. The fault detection method based on digital twinning according to claim 1, characterized in that, The simulation fault data generated by the diffusion model is high-fidelity simulation data generated by applying the diffusion model and combining a learning residual mechanism, and the diffusion model learns the residual distribution between the real equipment data and the initial simulation data generated based on the basic features to generate high-diversity simulation data.

4. The fault detection method based on digital twinning of claim 1, wherein, The diffusion model adopts a linear noise scheduler strategy, and defines the noise addition process as: ； wherein, is the total number of diffusion steps; the denoising network adopts a U-Net structure, and the input is a noisy signal and conditional embedding , the predicted noise residual ; The network comprises: 4 layers of down-sampling convolution, each followed by a time embedding and conditional embedding fusion module; 4 layers of up-sampling deconvolution, which preserves high-frequency details through a skip connection.

5. The fault detection method based on digital twinning according to claim 3, characterized in that, The learning residual mechanism is: Transforming the signal generation into a progressive correction of the noise distribution, expressed as: ; the goal is to predict the noise residual for the current step ; In a data-scarce scenario, the generation process is decomposed into T steps, and each step only needs to predict a small correction amount of the current noise. In the key diagnostic features of bearing fault vibration signals, local changes in the high-frequency region are focused on through explicit modeling.

6. The fault detection method based on digital twinning according to claim 3, characterized in that, The method also comprises a diffusion step iteratively performing the following steps: ； wherein, is the standard noise difference, is a random perturbation term to increase the generation diversity.

7. The fault detection method based on digital twinning according to claim 3, characterized in that, The service model adopts a Transformer Encoder and a MCF Layer to realize the identification and diagnosis of faults.

8. A fault detection apparatus based on digital twinning, characterized by, The device comprises a processor and a memory, and the memory stores program instructions, and the processor invokes the program instructions stored in the memory to enable the device to execute the method of any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program comprises program instructions, and the program instructions are executed by the processor to enable the processor to execute the method of any one of claims 1-7.

Images