Multi-motor coupling vibration intelligent diagnosis method and system

By generating a simulation dataset with fault labels and a deep reinforcement learning agent module, the problems of vibration signal coupling and data scarcity in multi-motor systems are solved, achieving highly accurate and adaptive fault diagnosis suitable for complex industrial environments.

CN121723115AActive Publication Date: 2026-03-24INST OF ENERGY HEFEI COMPREHENSIVE NAT SCI CENT (ANHUI ENERGY LAB)

Patent Information

Application Number
CN202610219886.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-02-24
Publication Date
2026-03-24
Estimated Expiration
2046-02-24

AI Technical Summary

Technical Problem

Vibration signals in multi-motor cooperative drive systems are severely coupled, making feature extraction and decoupling difficult. Fault sample data is scarce, and existing diagnostic methods have weak adaptive capabilities, making it difficult to achieve early warning and real-time decision-making.

Method used

The parameterized fault simulation module generates a coupled vibration simulation dataset with accurate fault labels, which is pre-trained using a deep reinforcement learning agent module. The model is then fine-tuned using an online policy optimization algorithm, and a self-iterative optimization diagnostic closed loop is formed through the simulation-model evolution module.

Benefits of technology

It enables cold starts under conditions of no historical fault samples, improves the accuracy of fault location and type identification, has early warning capabilities, and can quickly adapt to changes in equipment status, forming a diagnostic system that continuously optimizes itself.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121723115A_ABST
    Figure CN121723115A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of industrial equipment predictive maintenance and fault diagnosis, in particular to a multi-motor coupling vibration intelligent diagnosis method and system. According to the technical scheme, the method comprises the following steps: generating a coupled vibration simulation data set with an accurate fault tag through a parameterized fault simulation module based on collected vibration data of multiple motors in a normal operation state and physical parameters of motor equipment; according to the invention, cold start of the diagnosis system under the condition of no historical fault sample is realized; a deep reinforcement learning agent is used for processing a space-time coupling signal, and a layered reward mechanism is combined, so that the decoupling diagnosis precision and the early warning capability are remarkably improved; through the online self-adaptive learning module, the system can be finely adjusted in real time according to field feedback to quickly adapt to the state of individual equipment; and the simulation-model evolution module continuously optimizes the simulator and the diagnosis model by using the accumulated real data to form a self-iterative closed loop.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of predictive maintenance and fault diagnosis of industrial equipment, and in particular to a multi-motor coupled vibration intelligent diagnosis method and system. BACKGROUND

[0002] The multi-motor cooperative driving system is a core power component of modern industrial assembly lines, high-end equipment and automated production units. Through the precise coordination of multiple motors, the system can achieve complex motion control and high-load continuous operation, significantly improving production efficiency and process capacity. However, the complexity and coupling of the system also bring significant challenges to fault diagnosis, mainly in the following aspects:

[0003] Vibration signals are severely coupled, making feature extraction and decoupling difficult. In a multi-motor mechanical system, each driving unit is tightly connected through rigid or flexible components such as shaft couplings, sliders, guide rails, and lead screws, forming a complex dynamic coupling network. When a single motor or transmission component fails, the resulting vibration excitation will be transmitted, superimposed and modulated through the structure, resulting in a mixed strong coupling vibration at the sensor end. This coupling effect can severely mask the essential features of the fault, making traditional single-point signal analysis or simple frequency spectrum diagnosis methods significantly less accurate, with high false positive and false negative rates.

[0004] Fault sample data is scarce, and model training faces the cold start problem. Key equipment in industrial sites is usually expensive, precise in structure, and requires continuous and stable operation. It is often not feasible or economical to obtain sufficient and diverse fault vibration data by actively injecting faults (destructive testing). Therefore, the historical fault sample is extremely scarce, which makes it difficult for artificial intelligence diagnosis models (such as deep learning models) that rely on a large amount of labeled data to be effectively trained and initialized, i.e., facing the "cold start" dilemma. Data scarcity severely restricts the practical application and performance ceiling of data-driven methods in the field of industrial fault diagnosis.

[0005] Existing diagnosis methods have weak adaptability and are difficult to cope with changes in equipment state and working conditions. The running state of industrial equipment will change slowly due to wear, aging, load variation and environmental interference. Once the traditional diagnosis model is trained offline, its parameters are usually fixed, and it lacks the ability to learn and adjust online. When the equipment state deviates from the range covered by the training data, the model performance will degrade significantly, and it cannot achieve long-term reliable diagnosis. Early warning and real-time decision-making ability is insufficient. Many faults have progressive development characteristics, and their early features are weak and easily overwhelmed by noise. Existing methods focus more on classification and identification after the fault occurs, and have weak ability to provide early warning and real-time maintenance decision support in the early stages of fault development, missing the best opportunity for preventive maintenance. SUMMARY

[0006] The purpose of the present application is to solve the problem that most of the existing diagnostic systems in the background art are static models, which are difficult to adapt online once deployed according to equipment aging, working condition drift or new fault modes, and lack of continuous learning and evolution ability, and to propose a multi-motor coupled vibration intelligent diagnosis method and system.

[0007] In a first aspect, the present application provides a multi-motor coupled vibration intelligent diagnosis method, comprising the following steps:

[0008] Based on the collected vibration data of the multi-motor under normal operating conditions and the physical parameters of the motor equipment, a coupled vibration simulation data set with accurate fault labels is generated through a parameterized fault simulation module;

[0009] The coupled vibration simulation data set is input into a deep reinforcement learning agent module for pre-training, the state input of the deep reinforcement learning agent module is a time series three-dimensional tensor constructed based on multi-channel vibration signals, and the reward function is a composite reward function;

[0010] The multi-channel vibration signals in the real operating environment are input into the pre-trained agent model for real-time diagnosis, and the diagnosis suggestions are output and the feedback of the maintenance personnel based on the diagnosis results is received;

[0011] The feedback is converted into an immediate reward signal, and an online policy optimization algorithm is used to update the gradient of the policy network of the agent model, realizing online fine-tuning and adaptive optimization of the model;

[0012] The accumulated real data-label pairs verified by feedback are used to optimize the model parameters of the parameterized fault simulation module, generate a simulation data set closer to the real data distribution, and use the data set to perform incremental training and iterative updating of the agent model, forming a self-evolving diagnosis closed loop.

[0013] Optionally, the parameterized fault simulation module performs the following sub-steps:

[0014] Based on the physical parameters of the equipment, a parameterized physical fault model library including rotor imbalance, bearing local damage, misalignment and mechanical looseness is constructed;

[0015] The fault type, location and severity parameters are set, and the corresponding fault source vibration signals are generated through the physical fault model library;

[0016] Based on the transfer path model between the vibration source and the sensor, the fault source vibration signals are coupled and superimposed on the normal baseline vibration data to generate a multi-channel simulation vibration data set with structured fault labels.

[0017] Optionally, the dimensions of the time-series three-dimensional tensor include time steps, sensor channels, and feature dimensions; and the composite reward function includes at least a basic reward item based on diagnostic accuracy, a timeliness reward item related to how early or late a fault is diagnosed, and a safety reward item for punishing missed severe faults.

[0018] Optionally, the online policy optimization algorithm is a proximal policy optimization algorithm, which uses interaction data tuples including states, actions, rewards, and next states to update model parameters by maximizing a clipped policy objective function.

[0019] Optionally, the model parameters of the parameterized fault simulation module are optimized using accumulated feedback-verified real data-label pairs to generate a simulation data set, and the simulation data set is used to incrementally train and iteratively update the agent model, specifically including:

[0020] Data accumulation: store real vibration signals and their corresponding high-confidence fault labels confirmed by human feedback to build a real case database;

[0021] Distribution alignment: calculate the probability distribution difference between the simulation data generated by the parameterized fault simulation module and the data in the real case database in key features;

[0022] Parameter optimization: automatically adjust the fault model parameters and signal transmission path parameters in the parameterized fault simulation module to minimize the distribution difference;

[0023] Model iteration: use the optimized simulation module to generate a new simulation data set, incrementally train the agent model, and replace the online diagnosis model with the trained model with better performance.

[0024] In a second aspect, the present application provides a multi-motor coupled vibration intelligent diagnosis system for implementing the multi-motor coupled vibration intelligent diagnosis method of the first aspect, the system comprising:

[0025] A parameterized fault simulation module for generating labeled multi-motor coupled vibration simulation data based on physical mechanisms;

[0026] A deep reinforcement learning agent module containing a policy network and a value network for learning diagnostic strategies from simulation or real data;

[0027] An online diagnosis and interaction module for deploying the agent module, processing real-time vibration signals, outputting diagnostic recommendations, and collecting human feedback;

[0028] An online adaptive learning module for real-time reward calculation and updating of the policy network parameters of the agent module based on feedback;

[0029] A simulation-model evolution module is configured to optimize the parameterized fault simulation module according to real feedback data and drive iterative retraining of the intelligent agent module.

[0030] Optionally, the parameterized fault simulation module comprises:

[0031] A fault model library containing a plurality of fault mathematical models defined by unbalance mass, eccentricity, damage size, offset, angular deviation and gap parameters;

[0032] A coupling simulation engine is configured to map vibration signals output by the fault model to each sensor channel according to a transfer path matrix and superimpose normal baseline signals.

[0033] Optionally, the deep reinforcement learning intelligent agent module comprises:

[0034] A feature extraction and state construction unit is configured to extract time domain, frequency domain and time-frequency domain features from original vibration signals and construct a time-series three-dimensional state tensor;

[0035] An intelligent agent core network adopts an architecture based on Actor-Critic, wherein an Actor network outputs a probability distribution of a diagnosis action and a Critic network evaluates a state value.

[0036] Optionally, the online adaptive learning module comprises:

[0037] A reward calculation unit is configured to calculate an instantaneous scalar reward according to a preset composite reward function and artificial feedback labels;

[0038] An experience replay pool is configured to buffer interactive experience data;

[0039] An online optimizer is configured to sample data from the experience replay pool and execute a proximal policy optimization algorithm to update network parameters.

[0040] Optionally, the simulation-model evolution module comprises:

[0041] A high-confidence real case database;

[0042] A distribution difference measurement unit is configured to quantify differences in feature distributions between simulation data and real data;

[0043] A simulation parameter optimizer is configured to automatically adjust fault simulation parameters by using a gradient descent method to minimize the distribution difference.

[0044] Compared with the prior art, the present application has at least one of the following beneficial technical effects:

[0045] Through the parameterized fault simulation module, the fault simulation data set with accurate labels is generated based on the physical parameters and normal operation data of the equipment, so that the diagnosis system can complete initial training and deployment without relying on any actual historical fault samples.

[0046] The deep reinforcement learning agent can effectively decouple the mixed vibration signal by directly processing the time sequence three-dimensional state tensor containing time and space information and combining a layered composite reward function containing accuracy, timeliness and safety, thereby improving the accuracy of fault positioning and type identification and realizing early warning.

[0047] The online adaptive learning module converts the feedback of maintenance personnel into a reward signal in real time, and fine-tunes the model by using an online policy optimization algorithm, so that the diagnosis system can quickly adapt to the individual state differences and slow performance drift of the specific equipment.

[0048] The simulation-model evolution module automatically optimizes the fault simulation parameters by comparing and minimizing the difference between the simulation data and the real feedback data in the feature distribution, thereby generating more realistic training data to drive the diagnosis model to perform iterative incremental training and form a closed loop of continuous self-optimization.

[0049] In summary, the present application generates labeled data through the parameterized fault simulation module, realizes cold start of the diagnosis system under the condition of no historical fault sample, uses the deep reinforcement learning agent to process the time and space coupled signal, combines the layered reward mechanism, significantly improves the decoupling diagnosis accuracy and early warning ability, through the online adaptive learning module, the system can be adjusted in real time according to the field feedback, and quickly adapt to the individual equipment state, and the simulation-model evolution module uses the accumulated real data to continuously optimize the simulator and the diagnosis model, forming a self-iterative closed loop. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 It is a flowchart of a multi-motor coupled vibration intelligent diagnosis method;

[0051] Figure 2 It is a working principle diagram of the parameterized fault simulation module;

[0052] Figure 3 It is a structural schematic diagram of a multi-motor driven sliding block assembly unit.

[0053] The drawings show that: 1, drive motor M3; 2, clamping jaw; 3, vibration sensor; 4, drive motor M1; 5, shaft coupling; 6, sliding block; 7, guide rail; 8, lead screw; 9, drive motor M2; 10, base; 11, first vibration sensor Ch1; 12, second vibration sensor Ch2; 13, third vibration sensor Ch3; 14, fourth vibration sensor Ch4; 15, fifth vibration sensor Ch5. DETAILED DESCRIPTION

[0054] The present application is described herein with reference to particular embodiments, and applications thereof will be readily apparent to those skilled in the art with the content of this description. The present application can be implemented in various ways, and can take on various alternative embodiments, as will be apparent to those skilled in the art. It is to be understood that the following examples and features of the examples are merely exemplary and can be combined with each other, in so far as they do not conflict.

[0055] Embodiment: To make the purpose, technical solutions and advantages of the present application clearer, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. The present application provides a multi-motor coupling vibration fault diagnosis method and system based on deep reinforcement learning and simulation generation, aiming to solve the problems of cold start difficulty, low accuracy and poor adaptability of the diagnosis model caused by serious vibration signal coupling and lack of historical fault data in the prior art. The present embodiment takes the multi-motor driven slider assembly unit shown in the accompanying drawings as the application scenario for illustration, but the application of the present application is not limited thereto.

[0056] I. Overall system implementation framework

[0057] Reference Figure 1 The diagnosis system of the present application mainly follows the closed-loop process of "simulation generation pre-training, online interactive diagnosis, feedback driven evolution" when implemented. The system can be started without any historical fault data in the initial stage, only relying on normal operation data and physical parameters of the equipment. The core implementation modules include:

[0058] Parameterized fault simulation module: responsible for generating simulation fault data set with accurate labels.

[0059] Deep reinforcement learning agent module: as the core diagnosis decision unit, pre-training and online inference are performed.

[0060] Online diagnosis and interaction module: responsible for on-site data acquisition, diagnosis result display and human-computer interaction.

[0061] Online adaptive learning module: real-time fine-tuning of the agent is performed using on-site feedback.

[0062] Simulation-model evolution module: the simulator is optimized using accumulated real data to drive model iteration and upgrade.

[0063] II. Specific implementation of the parameterized fault simulation module

[0064] The purpose of this module is to generate realistic coupled vibration data with multiple fault labels, providing sufficient pre-training samples for deep reinforcement learning agents. The implementation steps are as follows, and the working principle can be found in [reference needed]. Figure 2 :

[0065] Baseline data and physical parameter acquisition: at the target device (e.g., Figure 3 The multi-motor driven slider assembly unit shown operates stably for a sufficient period of time under healthy conditions, synchronously acquiring multi-channel normal vibration data through multiple installed vibration sensors (such as Ch1-Ch5). ,in, Indicates the first One sample, Number of sensor channels The number of sampling points. It represents the real number field. Simultaneously, it acquires the physical parameters of the equipment, including the rated speed of each motor, bearing type, and the relative position coordinates of the sensors and each motor.

[0066] Constructing a parametric physical fault model library: Based on common mechanical fault mechanisms, establish fault models that can be controlled by key parameters. Specifically, this includes:

[0067] Rotor imbalance model: composed of unbalanced mass and eccentricity Parameterization. The generated simulated fault signal is... ,in, Indicates the first Taiwan Electric in time The generated simulated vibration signal For the transmission coefficient, This represents the time-domain function of the centrifugal force caused by rotor imbalance, where m represents the equivalent unbalanced mass on the rotor, and r is the eccentricity. It is the first The rotational angular frequency of the trolley motor, It is the initial phase angle.

[0068] Bearing local damage model: based on damage size The position is parameterized, and the simulation is a periodic impact sequence, with the waveform of a single impact being a decaying sinusoidal model. ,in, It is a vibration signal caused by a single impact response due to localized damage to the bearing. It is the impact amplitude coefficient, which is positively correlated with the damage size s. The attenuation coefficient is... This is the structure's natural frequency.

[0069] Misalignment model: determined by offset and angle deviation Parameterization, the simulated signal is ,in The simulated vibration signal is caused by misalignment. It is the second harmonic amplitude coefficient. ,in The system sensitivity coefficient, This is twice the rotational speed frequency, i.e., the characteristic frequency of misalignment. This indicates the phase angle.

[0070] Mechanical loosening model: based on gaps Parameterization; the simulated signal includes subharmonics and amplified noise. . This represents a simulated vibration signal caused by mechanical loosening. , The amplitudes are the 1 / 2 harmonic and 1 / 3 harmonic amplitudes, respectively. It is the rotational angular frequency. These represent the initial phases of the 1 / 2 and 1 / 3 harmonic components, respectively. express, It is a Gaussian white noise sequence. This is the noise amplification factor. This is the noise gain factor;

[0071] Coupled simulation data generation: Define a transfer path matrix Where K is the number of sensor channels and M is the number of potential fault sources. The initial value of matrix element T[k][m] can be estimated based on the physical distance between the sensor and the fault source using an attenuation model. Estimate, of which Here, is the attenuation coefficient of the vibration signal during transmission, and is the distance between the vibration source and the sensor. By setting different fault types, locations (such as motors M1, M2, and M3) and severity parameters, the corresponding fault model is invoked to generate the fault source signal. , The specific fault model mentioned above (m=1, representing) m=2 represents m=3 represents m=4 represents Then follow the formula ,in, It is the final generated number Simulated vibration data of each channel at time t. This is the first vibration sensor that synchronously collects data. Normal vibration data for each channel at time t, where M is the total number of fault types and m is the fault type number, m = 1, 2, ..., M. For matrix elements, For the simulation of fault signals in a specific fault model, For the k-th channel, all relevant fault signals are weighted and superimposed. This is then coupled to the baseline data to generate the final signal. Each data sample is labeled with a structured tag such as {Fault Source: Drive Motor M2 9, Type: Bearing Outer Ring Damage}, forming a simulation dataset. .

[0072] III. Specific Implementation of the Deep Reinforcement Learning Agent Module

[0073] This module is the core of the system's intelligent decision-making. It receives vibration signals, analyzes them, and outputs diagnostic actions.

[0074] State space construction: The input multi-channel vibration signal is segmented into non-overlapping time windows of fixed length. For each time window sample, the following features are extracted in parallel:

[0075] Temporal characteristics: such as root mean square (RMS), kurtosis, and crease factor.

[0076] Frequency domain characteristics: The frequency spectrum is calculated by FFT to extract the energy proportion of the frequency band where the rotation frequency and its harmonics of each motor and the bearing fault characteristic frequency are located.

[0077] Time-frequency domain features: The energy of the node with the highest energy distribution entropy value is selected as the feature through wavelet packet transform.

[0078] The features of the most recent T consecutive time windows are stacked in the format of [time step T, number of sensor channels K, feature dimension F] to construct a three-dimensional temporal state tensor. , as input to the intelligent agent.

[0079] Network architecture and training: The agent adopts an Actor-Critic architecture.

[0080] Policy Network (Actor): Input is the state tensor The algorithm sequentially passes through a one-dimensional convolutional layer (for extracting local spatial feature patterns) and a long short-term memory network layer (for capturing long temporal dependencies), ultimately outputting the probability distribution of all predefined diagnostic actions. ,in The discrete motion space includes fault types such as normal, rotor imbalance, misalignment, local bearing damage, and mechanical loosening.

[0081] Value Network (Critic): Has a feature extraction layer similar to the policy network, outputting the long-term value of the current state. Scalar estimation.

[0082] Pre-training: using a simulated dataset The agent is pre-trained using a proximal strategy optimization algorithm.

[0083] Action space and reward function design:

[0084] Action space: defined as a composite action, including fault location (specific motor number), fault type (such as imbalance, bearing damage, etc.), confidence level, and suggested operation (such as no action, warning, shutdown for inspection).

[0085] Hierarchical composite reward function It guides agent optimization during the training and online learning phases.

[0086] Accuracy Rewards When the diagnosis is correct, administer Rewards are given for mistakes. punish, This represents the base reward, which is a positive value.

[0087] Timely rewards To encourage early warning, if the fault characteristics are fully manifested within the reference time window... If the diagnosis is correct, an additional timely reward will be given. ,in As a timeliness reward coefficient, This indicates the actual duration from the occurrence of the fault to the current moment. The reference time window indicating when the fault characteristics are fully manifested, and This reward is only activated at that time to encourage early warnings and improve the timeliness of system warnings.

[0088] Security rewards Severe penalties will be imposed for failure to report serious faults to ensure system reliability. To ensure system reliability It indicates the severity of the punishment.

[0089] IV. Specific Implementation of the Online Diagnosis and Interaction Module

[0090] This module is responsible for deploying the trained agent to the actual industrial site for real-time diagnostics.

[0091] Hardware deployment and signal acquisition: Figure 3 The multi-motor driven slider assembly unit shown is an example. The multi-motor driven slider assembly unit will be described in detail below.

[0092] 1. Overall structure and installation foundation

[0093] The base 10 serves as the rigid foundation and mounting platform for the entire assembly unit. All core moving parts and sensors are directly or indirectly fixed to it, and the fifth vibration sensor Ch5 15 is installed at its geometric center.

[0094] 2. Horizontal drive and motion module

[0095] Drive motors M1 4 and M2 9 serve as the power source for horizontal movement, and are arranged in parallel and fixed on the base 10.

[0096] Transmission and guiding components: The output shafts of the two horizontal motors are connected to parallel lead screws 8 via couplings 5. Guide rails 7 and lead screws 8 are installed parallel to each other, together forming a motion guiding system.

[0097] The slider 6 is the core load-bearing component for horizontal movement. It engages with the lead screw 8 through a nut, and its bottom cooperates with the groove of the guide rail 7, thereby converting the rotational motion of the motor into precise linear reciprocating motion.

[0098] 3. Vertical Drive and Execution Module

[0099] The drive motor M31 is mounted on the slider 6, and its output shaft drives vertical movement.

[0100] The gripper 2, as the end effector, is driven directly by the drive motor M31 or through a transmission mechanism to realize the gripping and releasing of the workpiece.

[0101] 4. Vibration monitoring sensor network layout

[0102] Five vibration sensors 3 (general schematic) constitute a monitoring network designed to decouple multi-source vibration signals, and their arrangement strategy is as follows:

[0103] Direct monitoring points:

[0104] The first vibration sensor Ch1 11 is installed on the non-drive end housing of the horizontal drive motor M1 4 to directly monitor its body vibration.

[0105] The second vibration sensor Ch2 12 is installed on the non-drive end housing of the horizontal drive motor M2 9 to directly monitor its vibration.

[0106] The fourth vibration sensor Ch4 14 is installed at the mounting flange of the vertical drive motor M3 1 to directly monitor its vibration status.

[0107] Key points of coupling monitoring:

[0108] The third vibration sensor, Ch3 13, is installed at the center of slider 6. This is a key node where the motion of the three motors is coupled, and it can comprehensively reflect the vibration generated by the interaction of horizontal and vertical motion.

[0109] Environmental reference points:

[0110] The fifth vibration sensor Ch5 15 is mounted at the geometric center of the base 10 to monitor background vibration and environmental noise, providing a reference for the system vibration signal and helping to improve the signal-to-noise ratio.

[0111] A monitoring network is formed by three drive motors (M1, M2, M3) and five vibration sensors (Ch1-Ch5). Sensor signals are input to the industrial control computer in real time via a synchronous acquisition card.

[0112] Real-time diagnostic process: The acquired raw vibration signal is first preprocessed (denoising and normalization). Then, the state tensor for the current moment is constructed following the same process as in the pre-training phase. The data is then input into the deployed agent model. After inference, the model outputs an action probability distribution. The system selects the action with the highest probability as a diagnostic suggestion. For example, location: M2, fault type: bearing wear, confidence level: 85%, suggestion: warning.

[0113] Human-Computer Interaction: Diagnostic suggestions are presented to maintenance personnel through a graphical human-computer interface. Based on the actual situation during on-site inspections, maintenance personnel provide feedback by clicking the "Correct," "False Alarm," or "Missing Alarm" buttons. This feedback is a key data source driving subsequent online learning and model evolution.

[0114] V. Specific Implementation of the Online Adaptive Learning Module

[0115] This module enables the system to quickly adapt to the unique characteristics of specific equipment based on on-site feedback.

[0116] Feedback Conversion: Convert a single piece of feedback from maintenance personnel (such as clicking "Correct") into an immediate scalar reward that conforms to the definition of the reward function. (For example, to obtain) ).

[0117] Online policy update: The system uses the online PPO algorithm for policy fine-tuning. Each interaction generates a quadruple (...). , , , The data is stored in a fixed-capacity experience replay pool. When the amount of data in the pool reaches a small batch size, a batch of data is randomly sampled, and the dominance function is calculated. And perform a PPO gradient update to optimize the policy network parameters. Its objective function is:

[0118]

[0119] in, For loss function, For the current policy network parameters, For old strategy parameters, For the first The action to be performed step by step For the first The state tensor of the step, Let the probability of the action be given by the current policy. The probability of an action under the old strategy. This is the hyperparameter pruning coefficient (usually set to 0.2), used to constrain the step size of policy updates and ensure training stability. This process continues, causing the model policy to converge towards the optimal diagnostic policy for the current device.

[0120] VI. Specific Implementation of the Simulation-Model Evolution Module

[0121] This module is the core of the system's lifelong learning and continuous performance improvement, indirectly enhancing the diagnostic model by optimizing the simulator.

[0122] Construction of a high-confidence case library: Real-time vibration data and their actual fault labels corresponding to all manually confirmed "correct" diagnoses are stored as high-confidence data-label pairs in a dedicated database.

[0123] Simulator parameter optimization:

[0124] Distribution difference assessment: Periodically sample a set of real data from the case library and use the current parametric fault simulation module to generate simulation data with the same fault parameters.

[0125] Difference Measurement and Optimization: Calculate the difference in probability distributions between simulated data and real data on key features. The calculation formula is as follows:

[0126]

[0127] Where P(x) is the empirical probability of feature x appearing in the real data, and Q(x) is the model probability of the same feature x appearing in the simulation data. It is the set of all possible eigenvalues, using KL divergence. Measurement is performed. Among them... For the true data distribution, This is the distribution of simulation data.

[0128] Parameter tuning: to minimize With the goal of improving the simulation model, optimization algorithms such as gradient descent are used to automatically adjust the parameters of the fault model (such as m, r, s, d, etc.) and the parameters of the transfer path matrix T in the fault simulation module, so that the distribution of the simulation data continuously approximates the distribution of the real data.

[0129] Model Iterative Update: Using the optimized fault simulation module, a new batch of simulation datasets that more closely resemble real-world scenarios is generated. Using simulation datasets Incremental training is performed on the current online diagnostic model (which can incorporate historical real-world data). Once training is complete, a more powerful next-generation diagnostic model is obtained, replacing the original online model and completing one iterative evolution. This closed-loop process can be executed periodically or triggered automatically, allowing the system's diagnostic capabilities to continuously improve with device uptime.

[0130] It is worth noting that, through the specific implementation methods described above, the system implemented by this invention requires no historical fault data. Training is initiated by generating data through physical mechanism simulation, thus solving the initial data bottleneck in industrial scenarios. This invention utilizes deep reinforcement learning to directly process coupled spatiotemporal state tensors, combined with a targeted reward mechanism, effectively identifying and locating faults from mixed signals. Through an online learning module, it rapidly adapts to individual equipment differences, and through a simulation-model evolution closed loop, it uses continuously generated field data to drive a spiral improvement in the overall system performance, possessing true intelligent evolutionary characteristics. This invention features a clear modular design, forming a complete technical closed loop from simulation, training, deployment to evolution, making it suitable for complex industrial environments with high requirements for reliability, accuracy, and long-term maintenance costs.

[0131] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A method for intelligent diagnosis of multi-motor coupled vibration, characterized in that, Includes the following steps: Based on the vibration data collected under normal operating conditions of multiple motors and the physical parameters of the motor equipment, a coupled vibration simulation dataset with accurate fault labels is generated through the parameterized fault simulation module. The coupled vibration simulation dataset is input into the deep reinforcement learning agent module for pre-training. The state input of the deep reinforcement learning agent module is a time-series three-dimensional tensor constructed based on multi-channel vibration signals, and the reward function is a composite reward function. The multi-channel vibration signals from the real operating environment are input into the pre-trained intelligent agent model for real-time diagnosis, outputting diagnostic suggestions and receiving feedback from maintenance personnel based on the diagnostic results. The feedback is transformed into an instant reward signal, and an online policy optimization algorithm is used to update the policy network of the agent model using gradients. By utilizing accumulated, feedback-verified real-data-label pairs, the model parameters of the parameterized fault simulation module are optimized, a simulation dataset is generated, and the agent model is incrementally trained and iteratively updated using the simulation dataset.

2. The intelligent diagnostic method for multi-motor coupled vibration according to claim 1, characterized in that, The parameterized fault simulation module performs the following sub-steps: Based on the physical parameters of the equipment, a parametric physical fault model library is constructed, including rotor imbalance, local bearing damage, misalignment, and mechanical loosening. Set the fault type, location, and severity parameters, and generate the corresponding fault source vibration signal through the physical fault model library; Based on the transmission path model between the vibration source and the sensor, the vibration signal of the fault source is coupled and superimposed onto the normal baseline vibration data to generate a multi-channel simulated vibration dataset with structured fault labels.

3. The intelligent diagnostic method for multi-motor coupled vibration according to claim 1, characterized in that, The dimensions of the temporal three-dimensional tensor include time step, number of sensor channels, and feature dimension; The composite reward function includes at least: a basic reward based on diagnostic accuracy, a timeliness reward related to the timing of fault diagnosis, and a security reward that penalizes missed reports of serious faults.

4. The intelligent diagnostic method for multi-motor coupled vibration according to claim 1, characterized in that, The online policy optimization algorithm is a proximal policy optimization algorithm. It uses an interactive data tuple containing state, action, reward and next state to update the model parameters by maximizing the pruned policy objective function.

5. The intelligent diagnostic method for multi-motor coupled vibration according to claim 1, characterized in that, Using accumulated, feedback-verified real-world data-label pairs, the model parameters of the parameterized fault simulation module are optimized, a simulation dataset is generated, and the agent model is incrementally trained and iteratively updated using the simulation dataset. Specifically, this includes: Data accumulation involves storing real vibration signals confirmed by human feedback and their corresponding high-confidence fault labels to build a database of real cases. Distribution alignment is performed, and the probability distribution differences of key features between the simulation data generated by the parameterized fault simulation module and the data in the real case database are calculated. Parameter optimization aims to minimize the distribution differences by automatically adjusting the fault model parameters and signal transmission path parameters in the parameterized fault simulation module. The model is iterated by generating a new simulation dataset using the optimized simulation module, incrementally training the agent model, and replacing the online diagnostic model with a more powerful trained model.

6. A multi-motor coupled vibration intelligent diagnostic system, characterized in that, The system for implementing the intelligent diagnostic method for multi-motor coupled vibration as described in any one of claims 1-5 includes: The parameterized fault simulation module is used to generate labeled multi-motor coupled vibration simulation data based on physical mechanisms. The deep reinforcement learning agent module, which includes a policy network and a value network, is used to learn diagnostic policies from simulated or real data. The online diagnosis and interaction module is used to deploy the intelligent agent module, process real-time vibration signals, output diagnostic suggestions, and collect human feedback. An online adaptive learning module is used to calculate rewards in real time based on feedback and update the policy network parameters of the agent module; The simulation-model evolution module is used to optimize the parameterized fault simulation module based on real feedback data and drive the iterative retraining of the agent module.

7. The multi-motor coupled vibration intelligent diagnostic system according to claim 6, characterized in that, The parameterized fault simulation module includes: The fault model library contains various mathematical fault models defined by parameters such as unbalanced mass, eccentricity, damage size, offset, angular deviation, and clearance. The coupled simulation engine is used to map the vibration signal output by the fault model to each sensor channel according to the transfer path matrix, and superimpose it with the normal baseline signal.

8. The multi-motor coupled vibration intelligent diagnostic system according to claim 6, characterized in that, The deep reinforcement learning agent module includes: The feature extraction and state construction unit is used to extract time-domain, frequency-domain, and time-frequency-domain features from the original vibration signal and construct a time-series three-dimensional state tensor. The core network of the intelligent agent adopts an Actor-Critic-based architecture, in which the Actor network outputs the probability distribution of diagnostic actions, and the Critic network evaluates the state value.

9. The multi-motor coupled vibration intelligent diagnostic system according to claim 6, characterized in that, The online adaptive learning module includes: The reward calculation unit calculates the real-time scalar reward based on the preset composite reward function and human feedback labels; An experience replay pool is used to cache interactive experience data. An online optimizer is used to sample data from the experience replay pool and execute a near-end policy optimization algorithm to update network parameters.

10. A multi-motor coupled vibration intelligent diagnostic system according to claim 6, characterized in that, The simulation-model evolution module includes: A high-confidence database of real-world cases; Distribution difference measurement unit, used to quantify the difference in feature distribution between simulated data and real data; The simulation parameter optimizer uses the gradient descent method to automatically adjust the fault simulation parameters with the goal of minimizing the distribution differences.

Citation Information

Patent Citations

  • Multi-mode fusion motor intelligent maintenance system

    CN114580666A

  • Induction motor few-label fault diagnosis method based on fine tuning generative adversarial network

    CN118275882A

  • Motor composite fault diagnosis method based on empirical knowledge and data fusion driving

    CN119669833A

  • Engineering machinery fault prediction and intelligent maintenance method based on deep learning

    CN121145027A

  • Knowledge and twin model driven actuator key fault injection and diagnosis method

    WO2024113574A1

Cited By

  • An AMT motor fault diagnosis method and system based on machine learning

    CN122286497A