An industrial internet-based device whole-cycle management system and method

Abstract

This invention discloses a device lifecycle management system and method based on the Industrial Internet, relating to the field of Industrial Internet technology. This Industrial Internet-based device lifecycle management method trains a source domain fault diagnosis model by collecting source equipment data, extracts device fingerprints from target equipment baseline data, adds a domain adaptation module after fixing the source domain model parameters, and uses a gradient inversion layer and physical constraint regularization terms to achieve feature alignment and generate a target domain model. It also collects data in real time to output health status and remaining life predictions, and combines production plans and spare parts inventory to generate and execute maintenance decisions. This invention collects multi-source operating data from equipment through the Industrial Internet, extracts device fingerprints, and combines domain adaptive migration technology to generate a fault diagnosis model adapted to the target equipment. This achieves equipment health monitoring and remaining life prediction, relies on reinforcement learning to generate optimal maintenance decisions, adapts to individual equipment differences, ensures production continuity, and reduces operation and maintenance costs.

Description

Technical Field

[0001] This invention relates to the field of industrial internet technology, specifically to an industrial internet-based equipment lifecycle management system and method. Background Technology

[0002] With the rapid development of Industrial Internet technology, the manufacturing industry is undergoing a crucial transformation from digitalization to networking and intelligence. Industrial Internet platforms connect manufacturing elements such as equipment, production lines, and factories, enabling comprehensive data collection and cloud aggregation, thus providing the technological foundation for full-lifecycle equipment management. Full-lifecycle equipment management covers all stages from equipment procurement, installation, operation, maintenance to disposal. Its core objective is to extend equipment lifespan, reduce operating costs, and ensure production continuity through continuous monitoring and scientific maintenance of equipment status.

[0003] Current mainstream equipment management methods primarily rely on the layered architecture of industrial internet platforms. The equipment layer collects operating parameters such as vibration, temperature, and current from various sensors; the edge layer performs preliminary processing and caching of the collected data; the platform layer aggregates multi-source data and utilizes big data technology for analysis and mining; and the application layer provides users with functions such as status monitoring, alarm push notifications, and report generation. Some advanced solutions introduce machine learning modules at the platform layer, training fault diagnosis models using historical data to attempt to predict potential equipment failures. The application of digital twin technology enables real-time mapping between virtual models and physical equipment, providing new technical means for equipment health assessment. Decision optimization methods such as reinforcement learning are also being explored for generating maintenance strategies, aiming to find a balance between equipment health and production needs.

[0004] In practical industrial applications, equipment is often deployed in batches of the same model on production lines. Individual differences exist within the same model in terms of mechanical assembly, foundation installation, and operating environment, directly reflected in signal characteristics such as vibration response. Equipment lifecycle management faces the reality of a large number of devices, wide distribution, and frequent changes in operating conditions, placing higher demands on the generalization ability of fault diagnosis models. Existing solutions typically assume that training and deployment data follow the same distribution, and the model is directly deployed to other devices of the same model after training. However, as equipment operates over time, its signal characteristics slowly change due to aging and wear. Industrial sites accumulate a large amount of historical equipment operating data and maintenance records. This data contains patterns of equipment degradation and failure mechanisms. How to effectively utilize this data to improve the intelligence level of equipment management is a continuously researching direction in the field of industrial internet.

[0005] Existing technology, such as the invention patent application with publication number CN120124836A, discloses a method for equipment lifecycle management based on an industrial internet platform, including the following steps: S1: Collecting equipment operation data, maintenance records, and fault information through the equipment layer of the industrial internet platform, and transmitting the data to the edge layer in real time; S2: Performing preliminary processing on the received data through the edge layer, and forwarding it to the PaaS layer for big data analysis and modeling to identify the equipment's operating status, predict potential faults, and optimize maintenance plans; S3: Realizing remote monitoring and automated control of the equipment through the SaaS layer of the industrial internet platform; S4: Combining historical data, current status, and future predictions of the equipment, formulating and executing a full lifecycle management plan for the equipment. This invention achieves precise management and optimization of the entire equipment lifecycle by integrating functions such as equipment data acquisition, intelligent analysis, remote monitoring, and automated control.

[0006] Based on the above solutions, the limitations of existing technologies include at least the following problems: Although the method introduces a machine learning module at the PaaS layer for fault prediction modeling, it implicitly assumes that the training data and deployment data follow the same distribution. It does not consider the vibration response distribution shift caused by factors such as installation errors, differences in foundation stiffness, and fluctuations in environmental temperature and humidity between different individuals of the same model of equipment. In actual deployment, after the fault diagnosis model trained to 95% accuracy on the source device is directly transferred to the target device of the same model, the recognition accuracy often drops sharply to below 60%. The target device is mostly in a zero-sample scenario with no faults throughout its entire life cycle. The existing architecture lacks a cross-domain adaptive mechanism, making it difficult to complete the effective transfer and generalization of the model under the condition of unlabeled data on the target device. Ultimately, the fault prediction function claimed by the solution fails in the large-scale management of the entire life cycle of the equipment due to insufficient model adaptability. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a device lifecycle management system and method based on the Industrial Internet, which solves the problem that existing technologies fail in large-scale applications throughout the entire device lifecycle because they ignore individual differences among devices of the same model, leading to data distribution deviations and making it difficult for fault prediction models to be effectively generalized to unlabeled target devices.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for full-lifecycle management of equipment based on the Industrial Internet, comprising the following steps: collecting historical operating data of source equipment under multiple operating conditions and their corresponding fault labels; extracting multi-dimensional features by performing time-frequency domain transformation on the historical operating data; training a source domain fault diagnosis model based on a deep residual shrinkage network; collecting baseline operating data of the target equipment under healthy conditions during the stage when the target equipment is not carrying production load; extracting equipment fingerprints reflecting the differences between the mechanical assembly characteristics and foundation stiffness of the target equipment; fixing the feature extraction layer parameters of the source domain fault diagnosis model; and adding a feature extraction layer consisting of an adversarial domain classifier and a physical constraint regularization term after the feature extraction layer. The domain adaptation module minimizes the difference in high-level feature distribution between the source and target domains through a gradient inversion layer. Simultaneously, using device fingerprints as constraints, it forces the target domain features to align with similar features in the source domain through physical constraint regularization terms. This generates a target domain fault diagnosis model adapted to the individual characteristics of the target device without requiring fault label data for the target device. Real-time online operating data of the target device is collected and input into the target domain fault diagnosis model, which outputs the health status assessment results and remaining service life predictions of the target device. The remaining service life predictions are then input into a deep reinforcement learning model, combined with real-time production plans and spare parts inventory data, to generate maintenance decisions with the goal of minimizing overall costs and to issue and execute them.

[0009] Furthermore, the specific steps for extracting multidimensional features from historical operating data through time-frequency domain transformation are as follows: Perform variational mode decomposition on the historical operating data and determine the number of modes by observing the center frequency; sort the intrinsic mode function components according to their center frequencies and remove noise components with center frequencies lower than the power frequency; perform Hilbert transform on the retained components to obtain the instantaneous amplitude matrix and the instantaneous frequency matrix; concatenate the instantaneous amplitude matrix and the instantaneous frequency matrix on the time axis to form the time-frequency domain feature matrix.

[0010] Further, the specific steps for training the source domain fault diagnosis model are as follows: Construct a deep residual shrinkage network containing a residual module and a soft thresholding sub-network; input the time-frequency domain feature matrix into the network, and extract deep feature maps by the residual module; adaptively learn the threshold of each set of feature maps through the soft thresholding sub-network to perform soft thresholding processing on the deep feature maps; input the processed feature maps into a fully connected classification layer after dimensionality reduction by global average pooling, and complete the training with cross-entropy loss to obtain the source domain fault diagnosis model.

[0011] Further, the specific steps for extracting the equipment fingerprint are as follows: During the stage when the target equipment is not under production load, vibration acceleration signals are collected from three measuring points: machine feet, bearing housing, and machine casing; Fast Fourier Transform is performed on the signals from the three measuring points respectively to extract the peak amplitude and corresponding frequency of the preset frequency band, forming a spectrum peak vector; Transfer function analysis is performed on the signals from the three measuring points respectively to extract the natural frequency and damping ratio, forming a modal parameter vector; The spectrum peak vector and the modal parameter vector are normalized and then concatenated to form the equipment fingerprint vector.

[0012] Furthermore, the specific steps for minimizing the difference in high-level feature distribution between the source and target domains through the gradient inversion layer are as follows: Input the source domain data into the source domain fault diagnosis model, propagate forward to the output layer of the final residual module, and obtain the high-level feature vector of the source domain through global average pooling; input the target domain data into the network with fixed feature extraction layer parameters, propagate forward to the output layer of the final residual module, and obtain the high-level feature vector of the target domain through global average pooling; add an adversarial domain classifier to predict the domain labels of the two types of feature vectors, and calculate the domain classification loss using binary cross-entropy; insert a gradient inversion layer between the adversarial domain classifier and the feature extraction layer, backpropagating the gradient of the domain classification loss to the feature extraction layer, forcing the feature extraction layer to learn domain-independent features.

[0013] Furthermore, the specific steps forcing the target domain features to align with similar features in the source domain through physical constraint regularization are as follows: Select samples of the same model from the source domain, and calculate the class center vectors of each class on the high-level feature vectors of the source domain according to the fault category; calculate the maximum mean difference distance between the high-level feature vectors of the target domain samples and the class center vectors of each class, and use the class corresponding to the minimum distance as the pseudo-label; map the weight coefficients using the device fingerprint vector as input, and add the weighted maximum mean difference distance as the physical constraint regularization term to the total loss function; simultaneously minimize the fault classification loss, domain classification loss, and physical constraint regularization term to complete the feature alignment.

[0014] Further, the specific steps for outputting the remaining useful life prediction are as follows: Using the fault category probability output by the target domain fault diagnosis model as the observed value, construct a state-space model based on the double exponential degradation model; initialize the particle set, complete the particle state prediction according to the state transition equation, update the particle weights with the fault category probability at the current time and normalize them, and resample the updated particle set; use the weighted average of the resampled particle set as the current degradation state estimate, substitute it into the double exponential degradation model to obtain the remaining useful life estimate; perform kernel density estimation on the multi-time estimate sequence, and output the expected value and confidence interval of the remaining useful life.

[0015] Furthermore, the specific steps for generating and executing maintenance decisions are as follows: a state space is constructed using the remaining useful life prediction, health status, production plan urgency, and spare parts inventory status, and an action space is constructed using three types of maintenance actions; a deep Q-network is constructed; the deep Q-network is trained in a digital twin environment, and the action corresponding to the maximum Q value is taken as the optimal maintenance decision; the optimal maintenance decision is issued to the target equipment for execution and simultaneously pushed to the maintenance work order system.

[0016] Furthermore, the specific steps for real-time acquisition of online operating data of the target equipment and input into the target domain fault diagnosis model are as follows: real-time acquisition of vibration, temperature, and current operating data, followed by cleaning to remove outliers and null values; sliding slices of the cleaned data according to fixed time windows, with data within each time window constituting an input sample; inputting the input samples into the target domain fault diagnosis model, and calculating the probability of the health status category for the current time window through forward propagation.

[0017] An industrial internet-based equipment lifecycle management system includes: a source equipment data acquisition and training module, which collects historical operating data and fault labels of source equipment, performs time-frequency domain transformation on the historical operating data to extract multi-dimensional features, and trains a source domain fault diagnosis model based on a deep residual shrinkage network; a target equipment fingerprint extraction module, which collects baseline operating data of the target equipment during the non-production load stage and extracts equipment fingerprints reflecting the differences between mechanical assembly characteristics and foundation stiffness; and a domain adaptive migration module, which fixes the feature extraction layer parameters of the source domain fault diagnosis model and adds a domain adaptive module consisting of an adversarial domain classifier and a physical constraint regularization term thereafter, and uses a gradient inversion layer to optimize the model. The system minimizes the difference in high-level feature distribution between the source and target domains. Simultaneously, using device fingerprints as constraints, it forces the target domain features to align with similar features in the source domain through physical constraint regularization terms. This generates a target domain fault diagnosis model adapted to the individual characteristics of the target device without requiring fault label data for the target device. The real-time online prediction module collects online operating data of the target device in real time and inputs it into the target domain fault diagnosis model, outputting health status assessment results and remaining service life prediction values. The maintenance decision generation module inputs the remaining service life prediction values ​​into a deep reinforcement learning model, combines real-time production plans and spare parts inventory data, generates maintenance decisions with the goal of minimizing overall costs, and issues them for execution.

[0018] The present invention has the following beneficial effects:

[0019] (1) The equipment lifecycle management method based on the Industrial Internet collects historical operating data and fault labels under multiple working conditions on the source equipment, extracts multi-dimensional features through time-frequency domain transformation, and trains a source domain fault diagnosis model based on a deep residual shrinkage network. Then, it collects baseline operating data when the target equipment is not carrying production load, extracts equipment fingerprints reflecting the differences in mechanical assembly characteristics and foundation stiffness, fixes the parameters of the source domain model feature extraction layer, adds a domain adaptive module composed of an adversarial domain classifier and a physical constraint regularization term, completes model transfer without the need for target equipment fault labels, minimizes the difference in high-level feature distribution between the source domain and the target domain by using a gradient inversion layer, and forces the target domain features to align with the same type of features in the source domain by using equipment fingerprints as constraints, and finally generates a target domain fault diagnosis model that adapts to the individual features of the target equipment. In practical applications, this mechanism enables the high-precision model trained on the source equipment to be effectively transferred to other equipment of the same model, and can maintain high recognition accuracy even if the target equipment has never had a fault during its entire lifecycle.

[0020] (2) The equipment lifecycle management method based on the Industrial Internet adaptively determines the number of modes by the center frequency observation method when performing variational mode decomposition on historical operating data, removes the noise-dominant components whose center frequency is lower than the power frequency, and performs Hilbert transformation on the retained components to obtain the instantaneous amplitude matrix and instantaneous frequency matrix, which constitute the time-frequency domain feature matrix. After inputting the matrix into the deep residual shrinkage network, the residual module extracts the deep feature map. The soft thresholding subnetwork adaptively learns the threshold corresponding to each group of feature maps through the global average pooling layer and the fully connected layer. The learned threshold is used to perform soft thresholding on the deep feature map, setting the feature values ​​with absolute values ​​less than the threshold to zero and shrinking the feature values ​​with absolute values ​​greater than the threshold to zero. Thus, it can dynamically filter out noise components unrelated to the fault according to the distribution characteristics of the feature map itself, and retain weak fault features.

[0021] (3) In the equipment full life cycle management method based on industrial Internet, the vibration acceleration signals of three measuring points, namely machine feet, bearing seats and machine casing, are collected when the target equipment is not under production load during the equipment fingerprint extraction stage. The peak amplitude of the preset frequency band and the corresponding frequency are extracted by fast Fourier transform to form the spectrum peak vector. At the same time, the transfer function is analyzed to extract the natural frequency and damping ratio to form the modal parameter vector. The two types of vectors are normalized and spliced ​​to form the equipment fingerprint vector. In the domain adaptive training process, the same type of sample is selected from the source domain and the class center vector of each type is calculated according to the fault category on the high-level feature vector of the source domain. The maximum mean difference distance between the high-level feature vector of the target domain sample and the center vector of each type is calculated. The category corresponding to the minimum distance is used as the pseudo label. The weight coefficient is obtained by mapping through a fully connected network with the equipment fingerprint vector as input. The weighted maximum mean difference distance is added to the total loss function as a physical constraint regularization term. Thus, while the adversarial domain classifier forces the feature domain to remain unchanged, the physical constraint regularization term forces the target domain feature vector to shrink in the feature space to the region where the same type of feature vector of the source domain is located.

[0022] (4) This equipment lifecycle management method based on the Industrial Internet achieves closed-loop operation of equipment lifecycle management through the collaborative work between modules. The source equipment data acquisition and training module provides basic model support, the target equipment fingerprint extraction module captures individual characteristics of the equipment, the domain adaptive migration module realizes model adaptation, the real-time online prediction module monitors the equipment operating status, and the maintenance decision generation module implements operation and maintenance actions. The modules are closely connected, forming a complete management process from data acquisition, model training, status monitoring to maintenance execution, which can fully support the health management and operation and maintenance decision-making of the entire equipment lifecycle.

[0023] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0024] Figure 1 This is a flowchart of a device lifecycle management method based on the Industrial Internet of Things according to the present invention.

[0025] Figure 2 This is a block diagram of a device lifecycle management system based on the Industrial Internet of Things according to the present invention. Detailed Implementation

[0026] Please see Figure 1This invention provides a technical solution: a method for full-lifecycle management of equipment based on the Industrial Internet, comprising the following steps: collecting historical operating data of source equipment under multiple operating conditions and their corresponding fault labels; extracting multi-dimensional features by performing time-frequency domain transformation on the historical operating data; training a source domain fault diagnosis model based on a deep residual shrinkage network; collecting baseline operating data of the target equipment under healthy conditions during the stage when the target equipment is not carrying production load; extracting equipment fingerprints reflecting the differences between the mechanical assembly characteristics and foundation stiffness of the target equipment; fixing the feature extraction layer parameters of the source domain fault diagnosis model; and adding a domain adaptive function composed of an adversarial domain classifier and a physical constraint regularization term after the feature extraction layer. The module minimizes the difference in high-level feature distribution between the source and target domains through a gradient inversion layer. Simultaneously, using device fingerprints as constraints, it forces target domain features to align with similar features in the source domain through physical constraint regularization terms. This generates a target domain fault diagnosis model adapted to the individual characteristics of the target device without requiring fault label data. Real-time online operating data of the target device is collected and input into the target domain fault diagnosis model, outputting the target device's health status assessment results and remaining service life predictions. The remaining service life predictions are then input into a deep reinforcement learning model, combined with real-time production plans and spare parts inventory data, to generate and execute maintenance decisions aimed at minimizing overall cost.

[0027] Among them, the fault labels are the identification information of various faults during the operation of the source equipment. Specifically, they correspond to four types of typical industrial equipment faults: bearing wear, rotor imbalance, foundation loosening, and seal leakage. Each fault label corresponds to a unique numerical identifier, that is, bearing wear corresponds to label 1, rotor imbalance corresponds to label 2, foundation loosening corresponds to label 3, seal leakage corresponds to label 4, and normal operation corresponds to label 0.

[0028] The multiple operating conditions specifically include the rated load condition of the source equipment, the 50% load condition, the light load condition (20% load), and the start-stop condition. The duration of the historical operating data collected under each operating condition is no less than 100 hours. The sampling frequency is determined according to the data type. The sampling frequency for vibration data is 500Hz, the sampling frequency for temperature data is 10Hz, and the sampling frequency for current data is 50Hz.

[0029] The comprehensive cost is the sum of downtime losses, spare parts procurement costs, and maintenance labor costs, and its calculation formula is as follows:

[0030] ;

[0031] in, The downtime loss is expressed in yuan per hour and is calculated using the following formula:

[0032] ;

[0033] in, The output value produced by the equipment per unit time. This refers to the downtime.

[0034] The cost of spare parts procurement is determined based on the unit price of the spare parts corresponding to the type of failure.

[0035] For example, the cost of spare parts for worn bearings is 800 yuan per set, and the cost of correcting rotor imbalance is 500 yuan.

[0036] To cover maintenance labor costs, the cost is calculated based on the hourly rate of maintenance personnel, using the following formula:

[0037] ;

[0038] in, To maintain staff hourly wages, To maintain working hours;

[0039] For example: Under rated load conditions, vibration data is sampled at a frequency of 500Hz for 100 hours, and a total of [data missing] data were collected. There are 1 data point, corresponding to fault label 1 (bearing wear). In its comprehensive cost calculation, the equipment's output value per unit time P = 200 yuan / hour, and the downtime is... Hourly spare parts cost Yuan, maintenance personnel hourly wage W=50 yuan, maintenance hours Hours, Yuan, Yuan, comprehensive cost Yuan.

[0040] Specifically, the steps for extracting multidimensional features from historical operational data through time-frequency domain transformation are as follows:

[0041] Variational mode decomposition was performed on historical operating data, and the number of modes was determined by the center frequency observation method. Specifically:

[0042] Set a penalty factor for variational mode decomposition. Number of iterations Convergence threshold By inputting historical operating data into the variational mode decomposition model, several intrinsic mode function components are obtained. ( (where the number of modes is 0), perform a Fourier transform on each component to obtain the center frequency of each component. If the difference between any two center frequencies is not less than 1 Hz, then the number of current modes is determined. The final number of modes;

[0043] For example: performing variational mode decomposition on a vibration acceleration signal (sampling frequency 500Hz, sampling duration 1s), initially setting the number of temporary modes. After decomposition, the center frequencies of each component are 50Hz, 150Hz, 250Hz, 350Hz, 450Hz, and 550Hz, respectively. The difference between each center frequency is greater than 1Hz, thus determining the number of modes. ;

[0044] The intrinsic mode function components are sorted by their center frequencies, and noise components with center frequencies lower than the fundamental power frequency are removed. Specifically:

[0045] The power frequency of the computing source device The power frequency fundamental frequency is the fundamental frequency corresponding to the rated speed of the source equipment, and the calculation formula is:

[0046] ;

[0047] in, The rated speed of the source equipment is expressed in r / min.

[0048] All intrinsic mode function components are sorted by center frequency Sort by size from smallest to largest, then remove center frequency. The components of the effective intrinsic mode function are the remaining components.

[0049] For example: A certain source device is a motor, with a rated speed of... The power frequency fundamental frequency is calculated. The center frequencies of the intrinsic mode function components obtained by variational mode decomposition are 10Hz, 20Hz, 30Hz, 50Hz and 80Hz. After removing the components with center frequencies of 10Hz and 20Hz (both less than 25Hz), the remaining three components corresponding to 30Hz, 50Hz and 80Hz are the effective components.

[0050] Performing a Hilbert transform on the retained components yields the instantaneous amplitude matrix and the instantaneous frequency matrix, specifically:

[0051] For each retained eigenmode function component ( The Hilbert transform is performed on the number of modes retained. The Hilbert transform formula is:

[0052] ;

[0053] in, for The Hilbert transform result;

[0054] Depend on and Constitutes analytical signal ,in The imaginary unit;

[0055] Analyzing the instantaneous amplitude of the signal Instantaneous frequency ;

[0056] Arrange the instantaneous amplitude of each retained component in time series to form an instantaneous amplitude matrix. ,in The number of sampling points for historical running data, and the element in the k-th row and i-th column of the matrix. This represents the instantaneous amplitude of the k-th retained component at the i-th sampling point;

[0057] Similarly, the instantaneous frequencies of each retained component are arranged in time series to form an instantaneous frequency matrix. The element in the k-th row and i-th column of the matrix This represents the instantaneous frequency of the k-th retained component at the i-th sampling point;

[0058] For example: retain 3 intrinsic mode function components ( Number of sampling points After Hilbert transformation, the instantaneous amplitude matrix is ​​obtained. and instantaneous frequency matrix ,in This indicates that the instantaneous amplitude of the first retained component at the 100th sampling point is 0.85 m / s². This indicates that the instantaneous frequency of the second retained component at the 200th sampling point is 52Hz;

[0059] The instantaneous amplitude matrix and the instantaneous frequency matrix are concatenated along the time axis to form the time-frequency domain feature matrix, which is as follows:

[0060] Instantaneous amplitude matrix With instantaneous frequency matrix The matrix is ​​concatenated column-wise, meaning the first N columns are the N columns of the instantaneous amplitude matrix, and the last N columns are the N columns of the instantaneous frequency matrix. The concatenated matrix yields the time-frequency domain feature matrix. ,in The number of intrinsic mode function components to be retained is N, where N is the number of sampling points. This time-frequency domain feature matrix serves as the input to the subsequent source domain fault diagnosis model.

[0061] For example: Instantaneous amplitude matrix Instantaneous frequency matrix The spliced ​​matrix yields the time-frequency domain feature matrix. In the matrix, the first 500 columns of the first row are the instantaneous amplitude sequence of the first retained component, the last 500 columns of the first row are the instantaneous frequency sequence of the first retained component, and so on, to ensure that the time-frequency characteristics of each retained component are completely preserved.

[0062] The specific steps for training the source domain fault diagnosis model are as follows:

[0063] A deep residual shrinking network containing residual modules and a soft-thresholding subnetwork is constructed, specifically as follows:

[0064] The deep residual shrinking network consists of an input layer, a feature extraction layer, a soft thresholding sub-network, a global average pooling layer, and a fully connected classification layer;

[0065] The input dimension of the input layer and the time-frequency domain feature matrix Consistent, meaning the number of neurons in the input layer is ;

[0066] The feature extraction layer contains 6 residual modules. Each residual module consists of 2 convolutional layers (3×3 kernel size, stride 1, and same padding), a batch normalization layer (BN layer), and a ReLU activation function. The input and output dimensions of each residual module are consistent. Residual propagation is achieved through shortcut connections to avoid gradient vanishing.

[0067] The soft thresholding subnetwork consists of one fully connected layer (with the same number of neurons as the number of feature channels output by the feature extraction layer) and a sigmoid activation function, used to adaptively learn the soft threshold for each feature map; a global average pooling layer is used to reduce the dimensionality of the feature map to a feature vector;

[0068] The fully connected classification layer contains 5 neurons (corresponding to 4 types of faults + 1 type of normal state), and the output layer uses the Softmax activation function to output the probability of each type of healthy state;

[0069] For example: Time-frequency domain feature matrix The input layer has 3000 neurons, the feature extraction layer has 6 residual modules, each residual module outputs 64 feature channels, the soft thresholding subnetwork has 64 neurons in the fully connected layer, the global average pooling layer reduces the 64-channel feature map to a 64-dimensional feature vector, and the fully connected classification layer has 5 neurons, outputting the probability of 5 health states.

[0070] The time-frequency domain feature matrix is ​​input into the network, and the residual module extracts the deep feature map, specifically as follows:

[0071] The obtained time-frequency domain feature matrix Remodeling The dimension of the input layer of the deep residual shrinking network is mapped to the input layer as follows: Feature map ( (Initial number of feature channels).

[0072] The feature map is sequentially input into 6 residual modules. Each residual module extracts features through a convolutional layer, normalizes through a BN layer, performs non-linear mapping through a ReLU activation function, and then adds the input features and output features through a shortcut connection to obtain a deep feature map.

[0073] The dimension of the deep feature map is ,in For feature map height, The width of the feature map. This represents the final number of feature channels;

[0074] For example: Time-frequency domain feature matrix After being reshaped, it became The input layer is mapped to After processing by 6 residual modules, the deep feature map with dimension is obtained. That is, the number of feature channels increased from 32 to 64, and feature details were fully extracted;

[0075] The deep feature maps are then subjected to soft thresholding by adaptively learning the threshold for each set of feature maps through a soft thresholding sub-network. Specifically:

[0076] The deep feature map output by the residual module (H is the feature map height, W is the feature map width, and C is the number of feature channels) Input the soft-thresholding subnetwork,

[0077] The feature map is mapped to C threshold parameters through a fully connected layer. ( Then, the threshold parameter is mapped to the Sigmoid activation function. The interval is used to obtain the final soft threshold. ;

[0078] The soft thresholding formula is:

[0079] ;

[0080] in, To process the feature value of the c-th channel, h-th row, and w-th column, This is a sign function; it outputs 1 when the input is positive, -1 when the input is negative, and 0 when the input is 0.

[0081] For example: a channel in a deep feature map eigenvalues The soft-thresholding subnetwork learns the threshold for this channel. Then the processed eigenvalues ;

[0082] If a certain characteristic value After processing This achieves the suppression of noise characteristics;

[0083] The processed feature map is reduced in dimensionality by global average pooling and then input into a fully connected classification layer. Training is completed using cross-entropy loss to obtain the source domain fault diagnosis model, which is as follows:

[0084] Feature map after soft thresholding Global average pooling is performed. The formula for global average pooling is:

[0085] ;

[0086] in, For the first The pooling results of each channel yield a C-dimensional feature vector. ;

[0087] The feature vector is input into a fully connected classification layer, which then performs a linear mapping. ( This is the weight matrix. For bias vectors, (Number of categories) to obtain the output vector The probabilities of each category are obtained through the Softmax activation function:

[0088] ;

[0089] The training process uses the cross-entropy loss function to calculate the error between the predicted value and the true fault label. The cross-entropy loss formula is as follows:

[0090] ;

[0091] in, The number of training samples, Let be the true label of the i-th training sample (one-hot encoded; if the sample belongs to the m-th class, then...). (The rest are 0). Predict the probability that the i-th sample belongs to the m-th class;

[0092] The Adam optimizer is used to minimize the cross-entropy loss, and the optimizer parameter is set to the learning rate. attenuation coefficient Number of iterations In each iteration, the accuracy of the validation set is calculated. When the accuracy of the validation set does not improve for 10 consecutive iterations, training is stopped. The network model obtained at this time is the source domain fault diagnosis model.

[0093] For example: number of training samples Feature vector dimension Fully connected classification layer weight matrix Bias vector A training sample has a true label of 1 (bearing wear) and a one-hot encoding of [model name missing]. The predicted probability of type is Then the cross-entropy loss of this sample is The average loss of all training samples is the cross-entropy loss of the current round.

[0094] In this implementation scheme, variational mode decomposition is performed on the operating data, and mode selection is completed by combining the center frequency. Low-frequency interference components are then removed. With the help of Hilbert transform, time-frequency features are extracted and spliced ​​together. This can completely preserve the key information of equipment operation, making the features of the input model more consistent with the actual operating performance of the equipment. The deep residual shrinkage network is built to stabilize the feature extraction process with the residual structure. Irrelevant interference is filtered through adaptive threshold processing. Combined with reasonable training methods, the model's judgment ability is optimized, enabling the model to accurately capture abnormal performance of equipment operation, thereby effectively reducing the impact of external interference on the diagnostic results.

[0095] Specifically, the steps for extracting device fingerprints are as follows:

[0096] During the period when the target equipment is not under production load, vibration acceleration signals are collected from three measuring points: machine feet, bearing housing, and machine casing. Specifically:

[0097] The target equipment is in the no-load operation stage when it is not carrying a production load. At this time, the equipment is only running on its own and is not connected to a production load. The operating speed is 95%-100% of the rated speed.

[0098] Vibration acceleration sensors (model PCB 352C33, measurement range ±50g, sensitivity 100mV / g) were deployed at five measurement points: the machine feet (two feet at the front and rear, two at the rear), the bearing housing (one at the drive end and one at the non-drive end, two at the bearing housing), and the housing (one at the center). The sensor sampling frequency was set to 500Hz, and the sampling duration was set to 10s. Data was collected at each measurement point. One vibration acceleration signal data point, with the data unit being m / s²;

[0099] For example: A target device is a centrifugal pump with a rated speed of 1480 r / min and a no-load operating speed of 1475 r / min. One PCB 352C33 sensor is deployed at each of the following measurement points: front and rear of the machine feet, drive end and non-drive end of the bearing housing, and middle of the casing. The sampling frequency is 500 Hz, the sampling time is 10 s, and 5000 data points are collected at each measurement point. Among them, a data point at the middle of the casing has a speed of 1.23 m / s², and a data point at the drive end of the bearing housing has a speed of 0.89 m / s².

[0100] Perform Fast Fourier Transform on the signals from the three measurement points respectively, extract the peak amplitude and corresponding frequency of the preset frequency band, and construct the spectral peak vector, which is as follows:

[0101] The preset frequency band is set to 10-1000Hz (this band covers the main characteristic frequencies of mechanical faults in the equipment), and the vibration acceleration signal at each measuring point is... ( Perform a Fast Fourier Transform (FFT) on the sampled points (where the sampled points are the number of points). The FFT formula is as follows:

[0102] ;

[0103] in, The spectral amplitude corresponding to frequency f. The sampling frequency;

[0104] Calculate the absolute value of the spectral amplitude Within the preset frequency band of 10-1000Hz, extract the top 5 largest amplitude peaks. And the frequency corresponding to each amplitude peak. ;

[0105] By sequentially concatenating the five amplitude peaks with their five corresponding frequencies, a 10-dimensional spectral peak vector is formed. ;

[0106] For example: After FFT, the first five peak amplitude values ​​of the vibration acceleration signal at a certain measuring point within the 10-1000Hz frequency band are 3.2 m / s² (corresponding to 50Hz), 2.8 m / s² (corresponding to 100Hz), 2.5 m / s² (corresponding to 150Hz), 2.1 m / s² (corresponding to 200Hz), and 1.8 m / s² (corresponding to 250Hz). Then, the peak vector of the spectrum at this measuring point is: ;

[0107] Transfer function analysis was performed on the signals from the three measurement points to extract the natural frequencies and damping ratios, forming a modal parameter vector, which is as follows:

[0108] The transfer function analysis uses the frequency response function method to analyze the vibration acceleration signal at each measuring point. Excitation signal of equipment base (The excitation signal is generated by striking with a hammer, and the sampling frequency is the same as the vibration signal, which is 500Hz.) Perform FFT to obtain the vibration signal spectrum. and excitation signal spectrum transfer function ;

[0109] For transfer function Curve fitting was performed, and the first three natural frequencies of the device were extracted using a polynomial fitting method. (Unit: rad / s) and the corresponding order of damping ratio (Unitless, value range is 0-1);

[0110] By sequentially concatenating the third-order natural frequency and the third-order damping ratio, a 6-dimensional modal parameter vector is formed. ;

[0111] For example: After transfer function analysis, the first three natural frequencies extracted from a certain measuring point are as follows: (Corresponding frequency 50Hz) (Corresponding frequency 100Hz) (Corresponding frequency 150Hz), the corresponding damping ratios are respectively Then the modal parameter vector of the measuring point is ;

[0112] The device fingerprint vector is formed by concatenating the normalized peak spectral vector and the modal parameter vector. Specifically:

[0113] For the peak vector of the spectrum and modal parameter vector Min-max normalization is performed separately, and the normalization formula is as follows:

[0114] ;

[0115] in, Let be any element in the vector. The minimum value in the vector. The maximum value in the vector is the element that, after normalization, maps all elements to the interval [0, 1].

[0116] The normalized spectral peak vector (10-dimensional) and normalized modal parameter vector (6-dimensional) sequentially concatenated to obtain a 16-dimensional device fingerprint vector. This vector uniquely represents the differences in mechanical assembly characteristics and foundation stiffness of the target equipment, and there are significant differences in fingerprint vectors between different equipment.

[0117] For example: the normalized peak vector of the spectrum at a certain measuring point is The normalized modal parameter vector is The concatenated device fingerprint vector is then... .

[0118] In this implementation plan, vibration signals from multiple key measuring points are collected during the no-load phase of the target equipment. This accurately captures the inherent characteristics of the equipment's operation, avoiding interference from production loads. The peak values ​​and corresponding frequencies of the spectrum are extracted through fast Fourier transform, and the natural frequency and damping ratio are obtained by combining transfer function analysis. This comprehensively captures information related to the mechanical assembly and foundation stiffness of the equipment. After normalization processing, the data are stitched together to form an equipment fingerprint, which can uniquely distinguish the individual differences of different equipment, thereby reducing diagnostic bias caused by individual differences.

[0119] Specifically, the steps to minimize the difference in high-level feature distributions between the source and target domains using a gradient inversion layer are as follows:

[0120] The source domain data is input into the source domain fault diagnosis model, propagated forward to the output layer of the final residual module, and then subjected to global average pooling to obtain the high-level feature vector of the source domain, which is as follows:

[0121] The source domain data is the time-frequency domain feature matrix obtained by time-frequency domain transformation of the historical operating data of the source device. ( The number of modes, (number of sampling points), The source domain fault diagnosis model, trained by inputting it, is forward-propagated to the last residual module of the feature extraction layer, and the output is a deep feature map. ( (Number of feature channels);

[0122] Global average pooling is performed on the deep feature map. The pooling formula is as follows:

[0123] ;

[0124] Obtain the high-level feature vector of the source domain ,in The first high-level feature vector of the source domain One element;

[0125] For example: source domain time-frequency domain feature matrix After forward propagation through the source domain model, the residual module finally outputs a deep feature map. Global average pooling yields the high-level feature vectors of the source domain. The first element The second element And so on, for a total of 64 elements;

[0126] The target domain data is input into the network with fixed feature extraction layer parameters, and then propagated forward to the output layer of the final residual module. Global average pooling is then used to obtain the high-level feature vector of the target domain, which is as follows:

[0127] The target domain data is the time-frequency domain feature matrix obtained by time-frequency domain transformation of the collected baseline operating data of the target device. (Consistent with the data dimensions of the source domain);

[0128] The parameters of the feature extraction layer of the source domain fault diagnosis model are fixed (i.e., the parameters of the 6 residual modules of the feature extraction layer are not updated). The network with the fixed feature extraction layer is input, propagated forward to the last residual module, and outputs a deep feature map. (Same dimension as the deep feature map of the source domain);

[0129] Global average pooling is performed on the deep feature map using the same pooling formula as the source domain to obtain the high-level feature vector of the target domain. , and the source domain high-level feature vector The dimensions are exactly the same, ensuring that the distributions of the two are comparable;

[0130] For example: target domain time-frequency domain feature matrix After inputting a fixed feature extraction layer, the output is a deep feature map. Global average pooling yields high-level feature vectors of the target domain. The first element The second element The dimension is consistent with the high-level feature vector dimension of the source domain;

[0131] An adversarial domain classifier is added to predict domain labels for the two types of feature vectors. The domain classification loss is calculated using the binary cross-entropy, specifically as follows:

[0132] The adversarial domain classifier consists of two fully connected layers and a sigmoid activation function. The first fully connected layer has 10 neurons. ( (For high-level feature vector dimensions), the second fully connected layer has 1 neuron, the output layer uses the Sigmoid activation function, and the output domain label prediction probability. ,in Corresponding source domain (domain label) ), Corresponding target domain (domain label) );

[0133] High-level feature vectors of the source domain and target domain high-level feature vector Each input is fed into an adversarial domain classifier to obtain its respective domain label prediction probability. and ;

[0134] The domain classification loss is calculated using binary cross-entropy. The loss formula is:

[0135] ;

[0136] in, The number of samples in the source domain. The number of samples in the target domain. For source domain tags, For target domain tags, For the first Domain prediction probability of a source domain sample For the first Domain prediction probability of a target domain sample;

[0137] For example: Number of samples in the source domain Number of samples in the target domain High-level feature vector dimension The adversarial domain classifier has 32 neurons in the first fully connected layer and 1 in the second layer;

[0138] A certain source domain sample After input, predict the probability. (Corresponding source domain tag) ), loss item is ;

[0139] Sample of a certain target domain After input, predict the probability. (Corresponding target domain tag) ), loss item is The average loss of all samples is the domain classification loss. ;

[0140] A gradient reversal layer is inserted between the adversarial domain classifier and the feature extraction layer. This layer backpropagates the gradient of the domain classification loss to the feature extraction layer, forcing the feature extraction layer to learn domain-independent features. Specifically:

[0141] The role of the gradient inversion layer (GRL) is to reverse the domain classification loss output by the adversarial domain classifier during backpropagation. The gradient is multiplied by a fixed negative coefficient. The inverted gradient is then backpropagated to the feature extraction layer, as shown in the formula:

[0142] ;

[0143] in, The parameters of the feature extraction layer, These are the parameters of the gradient inversion layer;

[0144] Since the parameters of the feature extraction layer are fixed, the gradient inversion layer updates the parameters of the adversarial domain classifier through backpropagation gradient, which indirectly pushes the high-level features output by the feature extraction layer to optimize in the domain-independent direction. This forces the adversarial domain classifier to have difficulty accurately distinguishing the features of the source domain and the target domain, thereby allowing the feature extraction layer to learn the domain-independent features common to the source domain and the target domain, and minimizing the difference in the distribution of high-level features between the two.

[0145] For example: during backpropagation, the domain classification loss The gradient of the gradient reversal layer parameters is Gradient inversion layer coefficients The gradient backpropagated to the feature extraction layer is: By updating the parameters of the adversarial domain classifier through this gradient, the alignment of the feature distributions of the source domain and the target domain is gradually achieved.

[0146] The specific steps to force the target domain features to align with similar features in the source domain using physical constraint regularization terms are as follows:

[0147] Samples of the same model are selected from the source domain, and the class center vectors of each type on the high-level feature vectors of the source domain are calculated according to the fault category. Specifically:

[0148] Source device samples of the same model and specifications as the target device are selected from the source domain samples. The fault labels of these samples include 4 types of faults (labels 1-4) and normal status (label 0).

[0149] For each type of label Collect all source domain high-level feature vectors corresponding to this type of label. ( Given the number of source domain samples in class m, calculate the class center vector of that class. The calculation formula is:

[0150] ;

[0151] Class center vector (C represents the dimension of the high-level feature vector), consistent with the dimension of the high-level feature vectors of the source and target domains;

[0152] For example: In the sample of the same model of source equipment, the number of samples labeled 1 (bearing wear) High-level feature vector of each sample The class center vector of this class is calculated. The first element The second element Similarly, each fault category and normal state corresponds to a class center vector;

[0153] Calculate the maximum mean difference distance between the high-level feature vectors of the target domain samples and the center vectors of each class, and use the class corresponding to the minimum distance as the pseudo-label. Specifically:

[0154] Take the high-level feature vector of the target domain Calculate the class center vector between each source domain and the class center vector of each class. Maximum mean difference (MMD) distance The MMD distance formula is:

[0155] ;

[0156] in, The average value of the high-level feature vectors of the target domain samples. The square of the L2 norm;

[0157] The distances of the five MMDs were calculated. Select the smallest distance among them. The category corresponding to the minimum distance This refers to the pseudo-label of the target domain sample;

[0158] For example: the average value of high-level feature vectors of samples in the target domain The MMD distances between the vector and the five class center vectors are calculated as follows: minimum distance Corresponding category If so, the pseudo-label of the target domain sample is 1 (bearing wear).

[0159] The weight coefficients are obtained by mapping the device fingerprint vector as input, and the weighted maximum mean difference distance is added to the total loss function as a physical constraint regularization term. Specifically:

[0160] Extracted device fingerprint vector Input a mapping network with two fully connected layers, where the first fully connected layer has 8 neurons and the second fully connected layer has 1 neuron. The output layer uses the sigmoid activation function to obtain the weight coefficients. The greater the difference between the device fingerprint vector and the fingerprint vector of the same model sample in the source domain, The larger the value, the stronger the physical constraint;

[0161] Physical constraint regularization term The calculation formula is:

[0162] ;

[0163] in, The minimum MMD distance between the target domain sample and the source domain class center vector is calculated; the total loss function is the fault classification loss. Domain classification loss With physical constraint regularization terms The sum, the formula is:

[0164] ;

[0165] in, These are loss weights used to balance the impact of the three types of losses.

[0166] For example: Device fingerprint vector mapping yields weight coefficients. Minimum MMD distance Then the physical constraint regularization term ;

[0167] If the fault classification loss Domain classification loss The total loss

[0168] ;

[0169] Simultaneously minimize the fault classification loss, domain classification loss, and physical constraint regularization term to achieve feature alignment, specifically as follows:

[0170] Using the total loss function To optimize the objective, the total loss is minimized by combining the pseudo-labels of the target domain samples, the class center vectors of the source domain, and the physical constraint regularization term. This forces the high-level feature vectors of the target domain to move closer to the class center vectors of the source domain corresponding to their pseudo-labels, reducing the MMD distance between the two. This achieves accurate alignment between the target domain features and similar features of the source domain, ensuring that the target domain fault diagnosis model can achieve accurate health status assessment without the need for fault labels on the target equipment.

[0171] In this implementation scheme, by using a gradient inversion layer to adjust the feature propagation direction, the feature distribution gap between the source domain and the target domain can be narrowed, allowing the model to better identify the common operational features in the two types of data. By combining the device's own fingerprint to set corresponding constraints, the feature alignment strength can be adjusted according to the individual device situation. Then, by calculating the distribution distance of different types of features to determine pseudo-labels, the target domain data can also complete effective learning even in the absence of labels. The overall training process takes into account fault differentiation, domain distribution uniformity, and individual device differences, so that the transferred model can stably adapt to the target device and make accurate judgments based on the knowledge accumulated by the source device.

[0172] Specifically, the steps for outputting the predicted remaining useful life are as follows:

[0173] Using the fault category probabilities output by the target domain fault diagnosis model as observations, a state-space model based on the double exponential degradation model is constructed, specifically as follows:

[0174] The state-space model consists of state transition equations and observation equations. The state transition equations are as follows:

[0175] ;

[0176] in, Let t be the degradation state value at time t, and a, b, c, and d be the parameters of the double exponential degradation model (obtained by fitting the historical degradation data of the source equipment, which is consistent with the degradation pattern of the same model and operating conditions of the source equipment). The noise is state noise, which follows a mean of 0 and a variance of . Gaussian distribution;

[0177] The observation equation is:

[0178] ;

[0179] in, The observed values ​​are the maximum probabilities of various health states output by the target domain fault diagnosis model. To observe the true value, To observe the noise, it follows a mean of 0 and a variance of . Gaussian distribution;

[0180] For example: the target domain model outputs the probabilities of various states at a certain moment as follows: normal 0.05, bearing wear 0.92, rotor imbalance 0.02, foundation loosening 0.01, and seal leakage 0.00. (Observed values...) Substitute these values ​​into the observation equation to establish the correlation between the observed values ​​and the state values.

[0181] Initialize the particle set, predict the particle state according to the state transition equation, update the particle weights with the fault category probability at the current time and normalize them, and resample the updated particle set. Specifically:

[0182] Initialize particle set ,in Let be the degenerate state value of the i-th initial particle. The initial particle weights, and ;

[0183] According to the state transition equation Predict the particle state at time t ;

[0184] Calculate particle likelihood based on observation equations Update particle weights The weights are then normalized. ;

[0185] When the particle weight variance is greater than a preset threshold, the system resampling method is used to extract particles according to their weight proportions to ensure the diversity of the particle set.

[0186] For example: Initialize the number of particles N=1000, and the initial particle weights are all 0.001. After predicting the particle state at time t, based on the observed values... Calculate the likelihood, update the weights and normalize. Some particles with too low weights are removed, and particles are re-extracted to keep the particle count unchanged.

[0187] Using the weighted average of the resampled particle set as the estimate of the current degradation state, and substituting it into the bi-exponential degradation model, we obtain the estimate of the remaining useful life, which is as follows:

[0188] The particle set after resampling is Current degradation state estimate ;

[0189] Will Substitute into the bi-exponential degradation model Combined with preset fault thresholds Solve Obtain the time of failure Estimated remaining useful life ,in This is the current running time;

[0190] For example: particle weighted average obtained after resampling The parameters of the double exponential model are a=5.2, b=0.0015, c=1.8, d=0.0008, and the fault threshold is... Current running time Solving for the given information, we can obtain the following results: Estimated remaining useful life .

[0191] The multi-time-time estimated value sequence is subjected to kernel density estimation to output the expected value and confidence interval of the remaining lifetime, as follows:

[0192] Collect remaining useful life estimates over M consecutive time points. Using Gaussian kernel function Kernel density estimation is performed to obtain the probability density function of the remaining lifetime:

[0193] ;

[0194] Where h is the kernel function bandwidth (determined by Silverman's rule, formula: , (where M is the standard deviation of the multi-time estimate series and M is the number of time points).

[0195] The mean of the probability density function is used as the expected value of the remaining useful life. The values ​​corresponding to the cumulative probabilities of 2.5% and 97.5% were selected as confidence intervals;

[0196] For example: Five time-time estimates are collected at 852h, 858h, 865h, 862h, and 860h. After kernel density estimation, the expected value is... The confidence interval is [855h, 864h], which is used as the final output.

[0197] The specific steps for generating maintenance decisions and issuing them for execution are as follows:

[0198] The state space is constructed using the remaining useful life prediction, health status, production plan urgency, and spare parts inventory status. The action space is constructed using three types of maintenance actions. A deep Q-network (the core network of the deep reinforcement learning model) is then built, specifically as follows:

[0199] state space RUL represents the predicted remaining useful life, H represents the equipment health status (normal, bearing wear, rotor imbalance, foundation looseness, seal leakage), P represents the urgency of the production plan (value 0-1, the larger the value, the higher the urgency), and I represents the spare parts inventory status (sufficient, shortage, no inventory). The action space includes three maintenance actions: immediate shutdown maintenance, maintenance delayed by 72 hours, and maintenance delayed by 168 hours.

[0200] The deep Q-network consists of an input layer, two hidden layers, and an output layer. The number of neurons in the input layer is consistent with the dimension of the state space, the number of neurons in the hidden layer is set according to the state dimension, and the number of neurons in the output layer is 3, which correspond to the Q-values ​​of three different operational actions and are used to evaluate the benefits of different actions. Its core function is to realize the action value function learning of the deep reinforcement learning model.

[0201] In a digital twin environment, a deep Q-network is trained, and the action corresponding to the maximum Q-value is taken as the optimal maintenance decision. Specifically:

[0202] Build a digital twin model of the target equipment to replicate the equipment's operating conditions, production plans, and spare parts inventory scenarios. Input the state space data into a deep Q-network and initialize the Q-value function.

[0203] During training, the temporal difference learning algorithm (the core implementation of the Q-learning algorithm) is used to update the Q-value, with the following formula:

[0204] ;

[0205] in, Here, r is the discount factor, and r is the action reward value. The new state after the action is performed. Actions in the new state;

[0206] The reward value is based on minimizing the overall cost; the lower the cost, the higher the reward value.

[0207] After training until the Q-value converges, for the current input state, the action with the largest Q-value in the output layer is selected as the optimal maintenance decision;

[0208] For example, in a digital twin environment, the input state is RUL=860h, the health state is bearing wear, the production plan urgency is 0.8, and the spare parts inventory is sufficient. The Q values ​​of the three actions output by the deep Q network are -135, -98, and -112, respectively. The maintenance with the largest Q value and a delay of 72h is selected as the optimal decision.

[0209] The optimal maintenance decision is issued to the target equipment for execution and simultaneously pushed to the maintenance work order system. Specifically:

[0210] The optimal maintenance action is converted into an executable instruction and sent to the target equipment control system. The control equipment executes the maintenance action according to the decision requirements. If it is delayed maintenance, the system monitors the equipment operating status in real time and automatically triggers a maintenance reminder after the delay time is reached.

[0211] At the same time, maintenance decision-related information (maintenance time, maintenance type, required spare parts, maintenance personnel) is pushed to the maintenance work order system to generate standardized maintenance work orders, which are then assigned to the corresponding maintenance personnel. Production plan adjustment instructions are updated synchronously to ensure that maintenance actions are coordinated with production progress and to avoid affecting normal production.

[0212] For example, the optimal decision is to delay maintenance for 72 hours. The system issues a delayed maintenance instruction to the equipment control system and pushes a work order to the maintenance work order system, specifying that bearing wear maintenance will be carried out after 72 hours. The production plan is adjusted accordingly to reserve maintenance time.

[0213] In this implementation plan, a degradation sequence is constructed by combining the fault diagnosis results of the target domain and high-level features. The degradation law of the equipment is fitted by an appropriate model, and then smoothing is performed to reduce prediction fluctuations, ensuring that the remaining service life prediction results are consistent with the actual operation of the equipment. When generating maintenance decisions, the optimal maintenance action is selected by combining the prediction results, production plan and spare parts inventory through reinforcement learning, taking into account both maintenance costs and production schedule, avoiding problems caused by premature or late maintenance, achieving the lowest maintenance cost, and ensuring the orderly operation of production.

[0214] Specifically, the steps for real-time acquisition of online operating data of the target device and input into the target domain fault diagnosis model are as follows:

[0215] The system collects three types of operational data—vibration, temperature, and current—in real time using sensors deployed on the target equipment. Specifically:

[0216] Vibration acceleration sensors, temperature sensors, and current sensors are deployed at key measuring points such as the machine feet, bearing housings, and casing of the target equipment.

[0217] The vibration sensor sampling frequency is consistent with the baseline data, at 500Hz, and one data point is collected every 2ms; the temperature sensor sampling frequency is 10Hz, and one data point is collected every 100ms.

[0218] The current sensor has a sampling frequency of 50Hz and collects one data point every 20ms;

[0219] The sensor collects data in real time and transmits it to the data processing module to ensure the real-time performance and integrity of the data. The data units are m / s², ℃, and A.

[0220] For example, when a target device is running, the vibration sensor collects one vibration acceleration data every 2ms, the temperature sensor collects one operating temperature data every 100ms, and the current sensor collects one operating current data every 20ms, which are transmitted to the background data processing module in real time.

[0221] The real-time collected data is cleaned to remove outliers and null values. Specifically:

[0222] Outliers were removed using the 3σ criterion, and the mean of each data category was calculated. and standard deviation Values ​​deviating from the mean by ±3σ are considered outliers.

[0223] For missing values ​​that occur during the data collection process, linear interpolation is used to fill them in. The interpolation formula is as follows:

[0224] ;

[0225] in, This represents the value after filling in the missing information at time t. These are the valid data adjacent to each other before and after time t;

[0226] For example, if the vibration data collected at a certain moment is 3.2 m / s², and the mean of this type of data is 1.5 m / s² and the standard deviation is 0.5 m / s², the deviation from the mean exceeds 3σ, so it is judged as an outlier and removed.

[0227] The temperature data at a certain moment is empty, and the data before and after it are 38℃ and 40℃ respectively. It is filled with 39℃ by linear interpolation.

[0228] The cleaned data is sliced ​​into fixed time windows, with each time window containing a single input sample. Specifically:

[0229] Set a fixed time window duration (1s), and the vibration data, temperature data, and current data within each time window constitute a subsample; set the sliding step size to 1 / 2 of the time window duration, and extract data sequentially according to the sliding step size to ensure that the data are continuous and non-repetitive.

[0230] The three types of data within each time window are concatenated sequentially to form a complete input sample.

[0231] For example: Set the time window duration to 1 second and the sliding step size to 0.5 seconds. Within each window, there are 500 vibration data points, 10 temperature data points, and 25 current data points. After splicing, they form an input sample containing 535 data points. The data points are then extracted sequentially according to the sliding step size to obtain a continuous input sample sequence.

[0232] The input samples are fed into the target domain fault diagnosis model, and the probability of the health status category in the current time window is calculated through forward propagation, specifically as follows:

[0233] For each input sample, variational mode decomposition, Hilbert transform, and matrix concatenation are performed using the time-frequency domain transformation method to extract the time-frequency domain feature vector;

[0234] The feature vector is input into the target domain fault diagnosis model, and after passing through the feature extraction layer, soft thresholding sub-network, global average pooling layer and fully connected classification layer, the probability of various health states is calculated through forward propagation.

[0235] Output the probability of the health status category of the device within the current time window. The category corresponding to the maximum probability is the current health status of the device, and it also provides input for the prediction of the remaining service life.

[0236] For example: After extracting features from an input sample through time-frequency domain transformation, it is input into the target domain model. After forward propagation, the output probabilities of various states are: normal 0.05, bearing wear 0.96, rotor imbalance 0.02, foundation loose 0.01, and seal leakage 0.00. The current health state is determined to be bearing wear.

[0237] In this implementation plan, multiple sensors are deployed at key measurement points of the target equipment to collect various operational data in real time. This enables comprehensive capture of various status information during equipment operation, ensuring comprehensive and timely data collection. The collected data is preprocessed to remove abnormal data and fill in missing values, reducing interference from invalid data on model judgment and ensuring that the input data to the model is clean and reliable. Features are extracted according to a unified time-frequency domain transformation method to ensure a coherent and unified data processing flow. This allows the target domain fault diagnosis model to stably receive valid input and accurately output the equipment health status and remaining service life, ensuring the orderly progress of equipment operation and maintenance and avoiding diagnostic deviations due to data issues.

[0238] Please see Figure 2This invention provides a technical solution: a device lifecycle management system based on the Industrial Internet, comprising: a source device data acquisition and training module, which collects historical operating data and fault labels of source devices, extracts multi-dimensional features by performing time-frequency domain transformation on the historical operating data, and trains a source domain fault diagnosis model based on a deep residual shrinkage network; a target device fingerprint extraction module, which collects baseline operating data of the target device when it is not under production load, and extracts device fingerprints reflecting the differences between mechanical assembly characteristics and foundation stiffness; and a domain adaptive migration module, which fixes the feature extraction layer parameters of the source domain fault diagnosis model, and adds a domain adaptive module consisting of an adversarial domain classifier and a physical constraint regularization term thereafter, which minimizes the difference in high-level feature distribution between the source and target domains through a gradient inversion layer, and simultaneously uses the device fingerprint as a constraint condition through a physical constraint regularization term. The system forces the target domain features to align with similar features in the source domain, generating a target domain fault diagnosis model adapted to the individual characteristics of the target equipment without requiring fault label data for the target equipment. A real-time online prediction module collects online operating data from the target equipment in real time and inputs it into the target domain fault diagnosis model, outputting health status assessment results and predicted remaining service life. A maintenance decision generation module inputs the predicted remaining service life into a deep reinforcement learning model, combining real-time production plans and spare parts inventory data to generate maintenance decisions aimed at minimizing overall cost, and then issues these decisions for execution. A data storage module stores various types of data generated during the implementation of the method, employing an industrial-grade database for real-time data reading and writing and long-term storage, supporting data retrieval, and featuring a distributed storage architecture and data backup mechanism to prevent data loss and ensure data security and traceability.

[0239] The modules interact with each other via an industrial internet bus to ensure real-time data transmission and sharing.

[0240] The real-time online prediction module establishes real-time communication with the target device's sensors, synchronously receives the operating data collected by the sensors, and simultaneously feeds back the health status assessment results and remaining service life prediction values ​​to the maintenance decision generation module in real time.

[0241] The maintenance decision generation module works in conjunction with the production planning system, spare parts inventory system, equipment control system, and maintenance work order system to enable the rapid issuance and execution of maintenance decisions.

[0242] The data storage module synchronously stores all data generated by each module, including historical data of the source domain, baseline data of the target domain, model training parameters, prediction results, and maintenance decision records;

[0243] Each module has a data caching function to prevent management processes from stalling due to data transmission interruptions.

[0244] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0245] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for full lifecycle management of equipment based on the Industrial Internet, characterized in that, Includes the following steps: Historical operating data of source equipment under multiple operating conditions and their corresponding fault labels are collected. Multidimensional features are extracted by time-frequency domain transformation of the historical operating data, and a source domain fault diagnosis model based on a deep residual shrinkage network is trained. During the period when the target equipment is not under production load, baseline operating data of the target equipment under healthy conditions are collected, and equipment fingerprints reflecting the differences between the mechanical assembly characteristics and foundation stiffness of the target equipment are extracted. The feature extraction layer parameters of the fixed source domain fault diagnosis model are added after the feature extraction layer, and a domain adaptation module consisting of an adversarial domain classifier and a physical constraint regularization term is added. The difference in high-level feature distribution between the source domain and the target domain is minimized through the gradient inversion layer. At the same time, the device fingerprint is used as a constraint condition, and the physical constraint regularization term forces the target domain features to align with the same type of features in the source domain. A target domain fault diagnosis model adapted to the individual features of the target device is generated without the need for target device fault label data. Real-time acquisition of online operating data of the target equipment and input into the target domain fault diagnosis model; output of health status assessment results and remaining service life prediction values ​​of the target equipment. The remaining useful life prediction is input into a deep reinforcement learning model, which combines real-time production plans and spare parts inventory data to generate maintenance decisions with the goal of minimizing overall cost and then issues them for execution.

2. The method for full lifecycle management of equipment based on the Industrial Internet according to claim 1, characterized in that, The specific steps for extracting multidimensional features from historical operational data by performing time-frequency domain transformation are as follows: Variational mode decomposition was performed on historical operating data, and the number of modes was determined by the center frequency observation method; The intrinsic mode function components are sorted by center frequency, and noise components with center frequencies lower than the power frequency are removed. Perform a Hilbert transform on the retained components to obtain the instantaneous amplitude matrix and the instantaneous frequency matrix; The instantaneous amplitude matrix and the instantaneous frequency matrix are concatenated on the time axis to form a time-frequency domain feature matrix.

3. The method for full lifecycle management of equipment based on the Industrial Internet according to claim 2, characterized in that, The specific steps for training the source domain fault diagnosis model are as follows: Construct a deep residual shrinkage network containing residual modules and soft thresholding subnetworks; The time-frequency domain feature matrix is ​​input into the network, and the residual module extracts the deep feature map. The deep feature maps are processed by adaptively learning the threshold of each group of feature maps through a soft thresholding sub-network; The processed feature map is then reduced in dimensionality by global average pooling and input into a fully connected classification layer. Training is completed using cross-entropy loss to obtain the source domain fault diagnosis model.

4. The method for full lifecycle management of equipment based on the Industrial Internet according to claim 1, characterized in that, The specific steps for extracting device fingerprints are as follows: During the stage when the target equipment is not under production load, vibration acceleration signals are collected from three measuring points: machine feet, bearing housing, and machine casing. Perform Fast Fourier Transform on the signals from the three measurement points respectively, extract the peak amplitude and corresponding frequency of the preset frequency band, and form a spectrum peak vector; Transfer function analysis was performed on the signals from the three measurement points to extract the natural frequency and damping ratio, thus constructing a modal parameter vector. The device fingerprint vector is formed by concatenating the normalized peak vector of the spectrum and the modal parameter vector.

5. The method for full lifecycle management of equipment based on the Industrial Internet according to claim 3, characterized in that, The specific steps for minimizing the difference in high-level feature distributions between the source and target domains using a gradient inversion layer are as follows: The source domain data is input into the source domain fault diagnosis model, propagated forward to the output layer of the final residual module, and then obtained by global average pooling to obtain the high-level feature vector of the source domain. The target domain data is input into the network with fixed feature extraction layer parameters, and then propagated forward to the output layer of the final residual module. The high-level feature vector of the target domain is obtained by global average pooling. An adversarial domain classifier is added to predict domain labels for the two types of feature vectors, and the domain classification loss is calculated using the binary cross-entropy. A gradient reversal layer is inserted between the adversarial domain classifier and the feature extraction layer to backpropagate the gradient of the domain classification loss to the feature extraction layer, forcing the feature extraction layer to learn domain-independent features.

6. The method for full lifecycle management of equipment based on the Industrial Internet according to claim 5, characterized in that, The specific steps to force the target domain features to align with similar features in the source domain using physical constraint regularization terms are as follows: Samples of the same model are selected from the source domain, and the class center vectors of each type on the high-level feature vectors of the source domain are calculated according to the fault category. Calculate the maximum mean difference distance between the high-level feature vectors of the target domain samples and the center vectors of each class, and use the class corresponding to the minimum distance as the pseudo label; The weight coefficients are mapped using the device fingerprint vector as input, and the weighted maximum mean difference distance is added as a physical constraint regularization term to the total loss function. Simultaneously minimize the fault classification loss, domain classification loss, and physical constraint regularization term to complete feature alignment.

7. The method for full lifecycle management of equipment based on the Industrial Internet according to claim 1, characterized in that, The specific steps for outputting the predicted remaining useful life are as follows: Using the fault category probability output by the target domain fault diagnosis model as the observation value, a state-space model based on the double exponential degradation model is constructed. Initialize the particle set, predict the particle state according to the state transition equation, update the particle weights with the fault category probability at the current time and normalize them, and resample the updated particle set. The weighted average of the resampled particle set is used as the current degradation state estimate, which is then substituted into the double exponential degradation model to obtain the remaining lifetime estimate. The multi-time estimate sequence is subjected to kernel density estimation, and the expected value of the remaining lifetime and the confidence interval are output.

8. The method for full lifecycle management of equipment based on the Industrial Internet according to claim 7, characterized in that, The specific steps for generating maintenance decisions and issuing them for execution are as follows: The state space is composed of the remaining useful life prediction, health status, production plan urgency, and spare parts inventory status, and the action space is composed of three types of operation and maintenance actions. A deep Q network is constructed. Deep Q-network training is completed in a digital twin environment, and the action corresponding to the maximum Q value is the optimal maintenance decision. The optimal maintenance decision is issued to the target equipment for execution and simultaneously pushed to the maintenance work order system.

9. The method for full lifecycle management of equipment based on the Industrial Internet according to claim 1, characterized in that, The specific steps for real-time acquisition of online operational data of the target device and input into the target domain fault diagnosis model are as follows: Real-time acquisition of vibration, temperature, and current operating data, followed by cleaning to remove outliers and null values; The cleaned data is sliced ​​by sliding the slices according to fixed time windows, and the data in each time window constitutes an input sample. Input samples into the target domain fault diagnosis model, and calculate the health status category probability of the current time window through forward propagation.

10. A device lifecycle management system based on the Industrial Internet, employing the device lifecycle management method based on the Industrial Internet as described in any one of claims 1-9, characterized in that, include: The source device data acquisition and training module collects historical operating data and fault labels of the source device, performs time-frequency domain transformation on the historical operating data to extract multi-dimensional features, and trains a source domain fault diagnosis model based on a deep residual shrinkage network. The target equipment fingerprint extraction module collects baseline operating data of the target equipment when it is not under production load, and extracts equipment fingerprints that reflect the differences between mechanical assembly characteristics and foundation stiffness. The domain adaptive migration module fixes the feature extraction layer parameters of the source domain fault diagnosis model and adds a domain adaptive module consisting of an adversarial domain classifier and a physical constraint regularization term. It minimizes the difference in high-level feature distribution between the source domain and the target domain through a gradient reversal layer. At the same time, it uses the device fingerprint as a constraint condition and forces the target domain features to align with the same type of features in the source domain through the physical constraint regularization term. It generates a target domain fault diagnosis model that adapts to the individual features of the target device without the need for target device fault label data. The real-time online prediction module collects online operating data of the target equipment in real time and inputs it into the target domain fault diagnosis model, and outputs health status assessment results and remaining service life prediction values. The maintenance decision generation module inputs the remaining service life prediction value into the deep reinforcement learning model, combines it with real-time production plan and spare parts inventory data, generates maintenance decisions with the goal of minimizing overall cost, and issues them for execution.

Images