An incremental meta-learning online fault diagnosis method for industrial streaming data

Abstract

The application discloses an incremental meta-learning online fault diagnosis method for industrial streaming data, and comprises the following steps: S1: collecting vibration signals of different fault modes under different working conditions through a sensor, and dividing a plurality of diagnosis stage data sets; S2: building and training an initial fault diagnosis model; S3: introducing a Herding algorithm to select a small batch of representative samples from an old task training data set and store the samples in a data cache module; S4: jointly and randomly sampling a multi-task training set from the data cache module and a new stage training data set, and updating parameters of each sub-task model through meta-training; and S5: updating general model parameters through a dynamic balance meta-updating strategy. The method can mine general diagnosis knowledge among tasks of different stages, and through the design of a dynamic meta-updating strategy, learning of new knowledge and memory of old knowledge of the model can be balanced, so that the model can be quickly generalized to new diagnosis tasks, new fault data and new fault modes.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault diagnosis technology, and in particular to a novel online fault diagnosis method based on incremental meta-learning for streaming data. Background Technology

[0002] With the rapid development of modern information technology, computer technology, and artificial intelligence, complex equipment has been endowed with increasingly rich connotations and more complex functions. However, as various complex equipment continues to develop towards larger scale, higher speed, and greater precision, high load, high corrosion, and high operating rates have become the main characteristics, leading to unpredictable failures and even catastrophic events, causing huge economic losses, environmental pollution, and even casualties. Therefore, intelligent diagnosis and health status monitoring of complex equipment in operation are of great significance to ensure its safe and reliable operation.

[0003] In the era of industrial big data, the continuous improvement of sensor technology, IoT technology, and computing power has led to the widespread application of data-driven intelligent fault diagnosis technology in many industrial scenarios. This technology, with its ability to quickly learn diagnostic knowledge from massive amounts of historical data, has achieved high diagnostic accuracy and predictive performance. However, the operating status and performance of complex equipment typically change over time, and faults often develop and evolve gradually. The operation of equipment generates large volumes of real-time data streams, including sensor data, monitoring data, and log data. Industrial streaming data is characterized by continuity, high speed, large scale, diversity, real-time nature, and uncertainty. These characteristics require models to possess efficient data processing and rapid generalization capabilities, monitor the evolution of faults in real time, enable online updates, and achieve accurate diagnosis of emerging faults.

[0004] Shen Changqing et al. disclosed a "Lifelong Learning Method for Incremental Fault Diagnosis of Bearings Based on Generative Feature Replay" in Chinese Invention Patent CN116108346A. This method aims to address the catastrophic forgetting problem caused by the changing operating conditions and complex fault types in actual operation of existing technologies. However, the above invention only solves the problem of incremental fault types and lacks the ability to generalize to the distribution of unknown new fault data and the ability to update the model online. Therefore, it is necessary to study a fault diagnosis model with strong generalization and sustainability, which utilizes the general feature knowledge of equipment faults learned and discovered from historical data, and applies it to real-time changing data streams to achieve rapid and accurate identification of unknown new faults, in order to adapt to the challenges brought by the dynamic changes and uncertainties of industrial streaming data. Summary of the Invention

[0005] To overcome the shortcomings and deficiencies of existing technologies, this invention provides an incremental meta-learning online novel fault diagnosis method for streaming data. This method enables the model to quickly adapt to unknown new diagnostic tasks by learning common knowledge between tasks, and designs a dynamic meta-update strategy to balance the model's learning of new diagnostic knowledge and its memorization of old diagnostic knowledge, thereby realizing rapid online diagnosis of incremental fault patterns in industrial streaming data.

[0006] To achieve the objectives of this invention, the present invention provides an incremental meta-learning online fault diagnosis method for industrial streaming data, the specific steps of which are as follows:

[0007] S1. Collect vibration acceleration response signals of equipment under different working conditions and different fault modes, and divide them into multiple sets of fault datasets at different diagnostic stages.

[0008] S2. Build an initial fault diagnosis model and obtain the initial model parameters through training with a loss function;

[0009] S3. By introducing the Herding algorithm, a small batch of representative samples is selected from the old task training dataset and stored in the data cache module B. t-1 And based on the fault sample dataset D of the new task t And the old diagnostic task test data module B' t-1 Construct a joint test dataset {B' t-1 ∪D' t},D' t To obtain the fault sample dataset D for the new task t The dataset is constructed by extracting data from the middle and is based on the data caching module B. t-1 Construct a joint training dataset for the new stage using the fault sample dataset of the new task. Where, x i Represents the original sample, y i Indicates the fault category label, l i Indicates the task ID from which the sample comes;

[0010] S4, From the joint training dataset of the new phase A multi-task training set is formed by random sampling, and meta-training is used to update the model parameters of each sub-task.

[0011] S5. Based on the model parameters of each subtask, update the general model parameters through a dynamic balancing meta-update strategy, and test the model performance on each new and old diagnostic task.

[0012] Furthermore, in step S1, the collected samples are divided into multiple sets of fault diagnosis task datasets. Where t∈[1,T] represents different diagnostic stages, T represents the final diagnostic stage, and D tU represents the fault dataset for diagnostic phase t. t This represents the number of fault modes present during the diagnostic phase t.

[0013] Furthermore, in step S2, the initial model parameters are obtained by training using a label-smoothed cross-entropy loss function.

[0014] Furthermore, in step S2, the initial fault diagnosis model It consists of two networks: the feature mapping network f φ and fault classification network f θ Among them, the feature mapping network f φ It is designed based on ResNet and consists of one convolutional layer, one batch normalization layer, one max pooling layer, four ResNet modules, and one fully connected layer; the fault classification network f θ Composed of fully connected layers, this system is used to determine the fault category of input samples. Considering that with limited training data, neural networks tend to learn in the direction of maximizing the difference between correct and incorrect labels, and that insufficient training data cannot fully represent all sample features, this can lead to overfitting. Label smoothing is a common regularization strategy. By smoothing the one-hot class labels set during model training, noise is added to the labels, reducing the weight of the true sample label's class in the loss function calculation, which can suppress overfitting. Therefore, the model parameters are first initialized using the Kaiming initializer, and then the initial model is trained using the smoothed label cross-entropy loss function. The specific implementation steps are as follows:

[0015] S21. Initial stage dataset Smooth the labels of real samples:

[0016]

[0017] In the formula, ε represents the label smoothing hyperparameter. Indicates the original fault sample. This represents the fault category label corresponding to the sample, and n1 represents the number of samples in the initial stage of the dataset. This represents the smoothed sample label.

[0018] S22. Calculate the cross-entropy loss function L after label smoothing. ls Train and update the model parameters:

[0019]

[0020] Where, δ c (·) indicates that the code within the parentheses is processed using the Softmax function. This represents the average value of the output obtained after performing the operations within the parentheses on all fault samples in dataset D1. Represents the original fault sample After passing through the feature mapping network f φ The output after feature extraction Represents the original fault sample After passing through the feature mapping network f φ and fault classification network f θ The output after that.

[0021] Furthermore, in step S3, a small batch of samples is randomly drawn from the previous stage and placed into the data cache module B by introducing the Herding algorithm. t-1 and old diagnostic task test data module B' t-1 And from the fault sample dataset D of the new task t A portion of the data is extracted as the test dataset D' t , and the old diagnostic task test data module B' t-1 Together they form a joint test dataset {B' t-1 ∪D' t Then from data caching module B t-1 And the fault sample dataset D for the new task t K groups of samples are randomly selected from the dataset to form a joint training dataset for the new stage. Where, x i Represents the original sample, y i Describe the fault category label, l i Indicates the task ID from which the sample comes.

[0022] Furthermore, step S4 specifically includes the following:

[0023] S41. For the joint training dataset Multiple subtask datasets are obtained by grouping based on task ID. Where m∈[1,t], it represents different subtasks; J represents the number of samples in the dataset of the m-th subtask;

[0024] S42. Output the probability distribution of the model in each subtask dataset, and calculate the training loss function for each subtask:

[0025]

[0026] S43. Perform r rounds of iterative updates on each subtask. During the update, the feature mapping network f... φ and fault classification network f θ Isolated, feature mapping network f φThe parameters are updated on all subtasks to obtain new feature mapping network parameters φ, and the fault classification network f θ The parameters are updated individually on each subtask to obtain the fault classification network parameters {θ1,θ2,...,θ} for each old and new task. t}, where θ t This represents the fault classification network parameters for the t-th subtask.

[0027] Furthermore, step S5 specifically includes the following:

[0028] S51. Calculate the gradient of the general model update.

[0029]

[0030] in, This indicates that the general model parameters are obtained from the update in stage t-1, and φ represents the feature mapping network f. φ The parameter, θ j The fault classification network f represents θ The parameters.

[0031] S52. During training, the model is required to learn new tasks while retaining old knowledge. The model needs to adapt rapidly in the early learning stages and avoid drastic fluctuations in the later learning stages, maintaining the universal diagnostic knowledge learned between old and new tasks. Therefore, a momentum-based dynamic weight factor α is introduced to control the model to accelerate gradient updates in the early learning stages and decelerate learning in the later stages.

[0032]

[0033] In the formula, β is a decay rate that varies with the learning stage, representing the speed at which the model learns new knowledge.

[0034] S53. Update the general model parameters according to the dynamically changing weight factor α, so that the general model can learn in the average direction of all task-specific updates, ensuring that the model can simultaneously guarantee both new and old tasks. The model update method is as follows:

[0035]

[0036] S54, From the joint test dataset {B' of each new and old diagnostic task stage t-1 ∪D' t Random sampling was performed to test the overall diagnostic performance of the updated general diagnostic model for all diagnostic tasks.

[0037] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0038] 1. Existing data-driven fault diagnosis methods typically require large amounts of training data to help the model adapt to new diagnostic tasks, while also consuming significant time and computational resources. This does not meet the requirements of online real-time fault diagnosis based on streaming data in practical industrial applications. The present invention designs a meta-update strategy based on dynamic weight factors, enabling the model to maintain a dynamic balance between learning new diagnostic knowledge and retaining old diagnostic knowledge, minimizing catastrophic forgetting of learned diagnostic knowledge.

[0039] 2. This invention fully considers the problem of limited storage space during model updates. During training, the feature mapping network and the fault classification network are isolated to reduce interference. The feature mapping network parameters are updated on all sub-tasks, promoting the model's sharing of feature space across tasks and enhancing the learning of general features; the sample classification network parameters are optimized for each specific task, reducing the impact of the number of new task samples on model training and alleviating the problem of data imbalance between tasks.

[0040] 3. The incremental fault diagnosis model constructed in this invention introduces the idea of ​​"learning how to learn" in meta-learning, which enables the model to refine consistent diagnostic knowledge from old diagnostic knowledge. This effectively realizes the online rapid updating of the model and the accurate diagnosis of continuously generated unknown new faults. It provides a feasible and effective solution for scenarios in actual industrial applications where unexpected new faults occur, leading to incremental fault modes and complicated diagnostic tasks. It overcomes the challenges brought by the dynamic changes and uncertainties of industrial streaming data. Attached Figure Description

[0041] Figure 1 This is a flowchart of the incremental meta-learning online fault diagnosis method for industrial streaming data provided in this embodiment of the invention.

[0042] Figure 2 This is a framework diagram of the dynamic meta-update strategy for the incremental fault diagnosis model in the embodiments of this invention.

[0043] Figure 3 This is a comparison chart of the incremental fault classification accuracy of the method of this invention with other algorithms. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Please see Figure 1The present invention provides an incremental meta-learning online fault diagnosis method for industrial streaming data, which specifically includes the following steps:

[0046] S1. Vibration acceleration response signals of equipment under different operating conditions and different fault modes are collected using accelerometers, and the collected samples are divided into multiple fault diagnosis task datasets. Where t∈[1,T] represents different diagnostic stages, T represents the final diagnostic stage, and D t U represents the fault dataset for diagnostic phase t. t This represents the number of fault modes present during the diagnostic phase t.

[0047] In some embodiments of the present invention, taking rolling bearings as an example, possible failure modes include inner ring failure, outer ring failure, rolling element failure, etc.

[0048] S2. Build an initial fault diagnosis model and obtain the initial model parameters through training with a loss function.

[0049] In some embodiments of the present invention, the loss function used is the label-smoothing cross-entropy loss function. In other embodiments, other loss functions may be used, as long as the model is trained.

[0050] Initial fault diagnosis model Including feature mapping network f φ and fault classification network f θ Among them, the feature mapping network f φ Designed based on ResNet, it includes a convolutional layer, a batch normalization layer, a max-pooling layer, four ResNet modules, and a fully connected layer. The convolutional layer extracts features at different scales and locations from the input data through convolution operations, achieving feature extraction and transformation. The batch normalization layer stabilizes the distribution of input data in each layer of the network, accelerating model learning while mitigating the gradient vanishing problem. The max-pooling layer's main function is to downsample the input feature map, reducing its size and extracting the main feature information. The ResNet module addresses the training difficulties caused by network depth, mitigating gradient vanishing and model degradation. The fully connected layer connects all neurons in the previous layer with all neurons in the current layer, achieving linear transformation and feature combination from input features to output. The fault classification network f... θThis includes fully connected layers to determine the fault category of the input samples. Considering that with limited training data, neural networks tend to learn in the direction of maximizing the difference between correct and incorrect labels, and that insufficient training data cannot fully represent all sample features, potentially leading to overfitting, the model parameters are first initialized using a Kaiming initializer, and then trained using a smoothed label-based cross-entropy loss function. Noise is added to the labels by using predefined one-hot class labels during training, reducing the weight of the true sample label's class in the loss function calculation and thus suppressing overfitting. The specific implementation steps are as follows:

[0051] S21. Initial stage dataset Smooth the labels of real samples:

[0052]

[0053] In the formula, ε represents the label smoothing hyperparameter. Indicates the original fault sample. This represents the fault category label corresponding to the sample, and n1 represents the number of samples in the initial stage of the dataset. U1 represents the number of fault modes in the first diagnostic stage.

[0054] S22. Calculate the cross-entropy loss function after label smoothing, train and update the initial model parameters:

[0055]

[0056] Where, δ c (·) indicates that the code within the parentheses is processed using the Softmax function. This represents the average value of the output obtained after performing the operations within the parentheses on all fault samples in dataset D1. Represents the original fault sample After passing through the feature mapping network f φ The output after feature extraction Represents the original fault sample After passing through the feature mapping network f φ and fault classification network f θ The output after that.

[0057] S3. By introducing the Herding algorithm, the old task training dataset D is used. t-1 A small batch of representative samples is selected and stored in data cache module B. t-1 .

[0058] In some embodiments of the present invention, in step S3, a small batch of samples is randomly drawn from the previous stage and placed into the data cache module B by introducing the Herding algorithm. t-1 and old diagnostic task test data module B' t-1 And from the fault sample dataset D of the new task t A portion of the data is extracted as the test dataset D' t , and the old diagnostic task test data module B' t-1 Together they form a joint test dataset {B' t-1 ∪D' t Then from data caching module B t-1 And the fault sample dataset D for the new task t K groups of samples are randomly selected from the dataset to form a joint training dataset for the new stage. Where, x i Represents the original sample, y i Describe the fault category label, l i Indicates the task ID from which the sample comes.

[0059] S4, From the joint training dataset of the new phase The training set is composed of multiple tasks through random sampling, and the parameters of each sub-task model are updated through meta-training.

[0060] Step S4 specifically includes the following:

[0061] S41. For the joint training dataset Grouping by task ID yields multiple subtask datasets. Where m∈[1,t], representing different subtasks; J represents the number of samples in the dataset of the m-th subtask; l j Indicates sample x j The task ID from which it originated.

[0062] S42. Output the probability distribution of the model in each subtask dataset, and calculate the training loss function for each subtask:

[0063]

[0064] x j Let y represent the j-th fault sample in the m-th subtask dataset. j U represents the label of the j-th fault sample in the m-th subtask dataset. m f represents the number of failure modes in the m-th diagnostic stage. φ (x j ) represents the original fault sample x j After passing through the feature mapping network f φ The output after feature extraction, f θ(f φ (x j )) represents the original fault sample x j After passing through the feature mapping network f φ and fault classification network f θ The output after that.

[0065] S43. Perform r rounds of iterative updates on each subtask. During the update, the feature mapping network f... φ and fault classification network f θ Isolated, feature mapping network f φ The parameters are updated on all subtasks to obtain new feature mapping network parameters φ, and the fault classification network f θ The parameters are updated individually on each subtask to obtain the fault classification network parameters {θ1,θ2,...,θ} for each old and new task. t}, where θ t This represents the fault classification network parameters for the t-th subtask.

[0066] S5. Based on the model parameters of each subtask, update the general model parameters through a dynamic balancing meta-update strategy, and test the model performance on each new and old diagnostic task.

[0067] Step S5 specifically includes the following:

[0068] S51. Calculate the gradient of the general model update:

[0069]

[0070] in, This indicates that the general model parameters are obtained from the update in stage t-1, and φ represents the feature mapping network f. φ The parameter, θ j The fault classification network f represents θ The parameters.

[0071] S52. During training, the model is required to learn new tasks while retaining old knowledge. The model needs to adapt rapidly in the early learning stages and avoid drastic fluctuations in the later learning stages, maintaining the universal diagnostic knowledge learned between old and new tasks. Therefore, a momentum-based dynamic weight factor α is introduced to control the model to accelerate gradient updates in the early learning stages and decelerate learning in the later stages.

[0072]

[0073] In the formula, β is a decay rate that varies with the learning stage, representing the speed at which the model learns new knowledge.

[0074] S53. Update the general model parameters according to the dynamically changing weight factor α, so that the general model can learn in the average direction of all task-specific updates, ensuring that the model can simultaneously guarantee both new and old tasks. The model update method is as follows:

[0075]

[0076] in, This represents the updated general model parameters.

[0077] S54, From the joint test dataset {B' of each new and old diagnostic task stage t-1 ∪D' t Random sampling was performed to test the overall diagnostic performance of the updated general diagnostic model for all diagnostic tasks.

[0078] In some embodiments of the present invention, the model training and parameter update processing strategies in the above method are as follows: Figure 2 As shown, the specific steps include the following:

[0079] Step 1: Dataset Reconstruction. This involves reconstructing the general diagnostic model obtained in the previous diagnostic phase. Uploaded to the new diagnostic phase and from data cache library B t-1 And the fault sample dataset D for the new task t K groups of samples are randomly selected from the dataset to form a joint training dataset for the new stage. Where, x i Represents the original sample, y i Describe the fault category label, l i This indicates the task ID from which the sample originates. Then, the data is reassembled to obtain the joint training dataset. Multiple subtask datasets are obtained by grouping based on task ID. Where m∈[1,t], represents different subtasks.

[0080] Step 2: Subtask Model Parameter Update. First, input the fault samples from each subtask dataset into the model to obtain its predicted class probability distribution. Then, calculate the training loss function for each subtask based on the sample's true label:

[0081]

[0082] Perform r iterations on each subtask to obtain the updated model parameters {φ,θ1,θ2,...,θ} for both the old and new tasks. t}

[0083] Step 3: Update the general diagnostic model. First, based on the model parameters {φ,θ1,θ2,...,θ} obtained in Step 2 for each subtask and the new task,... tCalculate the update gradient of the general model:

[0084]

[0085] in, This indicates that the general model parameters are obtained from the update in stage t-1, and φ represents the feature mapping network f. φ The parameter, θ j The fault classification network f represents θ The parameters.

[0086] During training, the model is required to learn new tasks while retaining old knowledge. The model needs to adapt rapidly in the early learning stages and avoid drastic fluctuations in the later stages, maintaining the universal diagnostic knowledge learned between old and new tasks. Therefore, a momentum-based dynamic weight factor α is introduced to control the model's gradient updates to accelerate in the early learning stages and decelerate in the later stages.

[0087]

[0088] In the formula, β is a decay rate that varies with the learning stage, representing the speed at which the model learns new knowledge.

[0089] The general model parameters are updated based on dynamically changing weight factors α, allowing the general model to learn in the average direction of all task-specific updates, ensuring that the model can simultaneously handle both new and old tasks. The model update method is as follows:

[0090]

[0091] Step 4: Iterative Update and Performance Testing of the General Diagnostic Model. If the model update iteration count has not reached the maximum iteration test e, repeat steps 1-3; when the model update iteration count reaches the maximum iteration test e, complete the training and update of the general diagnostic model, obtaining the final stable feature mapping network f. φ and fault classification network f θ The parameters. During model performance testing, the joint test dataset {B' from each new and old diagnostic task stage is used. t-1 ∪D' t Random sampling was performed to test the overall diagnostic performance of the updated general diagnostic model for all diagnostic tasks.

[0092] In some embodiments of this invention, incremental fault diagnosis experiments were designed on the publicly available PU bearing fault dataset (a rolling bearing dataset provided by the University of Paderborn, Germany) to verify the effectiveness of the proposed method. The PU bearing dataset is one of the most commonly used validation datasets in the field of mechanical fault diagnosis. Its experimental testing apparatus includes a test motor, shaft, bearing module, flywheel, and load motor. The experimental bearing is a 6203 rolling bearing, and the experimental data was simulated by installing ball bearings with different damage types in the test module of the testing apparatus. In the experimental operating conditions, spindle speed, load torque, and bearing radial force are the three main variables. The spindle operates at two speeds: 1500 rpm and 900 rpm; the load torque is 0.1 Nm and 0.7 Nm; and the bearing radial force is 400 N and 1000 N, respectively. In this embodiment, eight different types of artificially damaged bearings with different fault locations (inner and outer rings), different damage degrees (level 1 and level 2), and different fault injection methods (electro-etching, drilling, and EDM) were used to test and evaluate the effectiveness of the proposed method.

[0093] Based on the aforementioned automotive transmission fault dataset, multiple incremental fault diagnosis tasks across operating conditions were set up in the experimental cases, specifically including different speeds (T1-T2), different loads (T3-T4), different bearing radial forces (T5-T6), different speeds and loads (T7-T8), and different speeds and different bearing radial forces (T9-T10). Detailed information on the incremental fault diagnosis experiments is shown in Table 1, where "N15_M07_F10" represents an operating condition with a speed of 1500 rpm, a load torque of 0.7 Nm, and a bearing radial force of 1000 N. Each fault type has 500 samples, and each sample contains 2048 data points. Each incremental fault diagnosis task includes an initial training phase (training the initial model using source domain fault data) and five incremental update phases (training and updating the model using multiple sets of target domain fault data with different fault categories). After the initial phase, each known fault category retains 25 samples in the next phase. The number of fault categories in the initial phase is set to 3, and in each incremental phase, one new fault type is added in each phase.

[0094] Table 1. Incremental fault diagnosis experiment designed based on automotive transmission fault dataset.

[0095]

[0096] In model training and updates, the RAdam optimizer with an initial learning rate of 0.01 was used to optimize model parameters; the maximum number of training iterations, e, was 50; the learning rate was multiplied by 1 / 5 and updated once at training iterations of 20, 40, and 60. Specific experimental parameters in the embodiments are shown in Table 2. All algorithms were compiled and tested using the PyTorch framework, with an NVIDIA GeForce GTX TITAN X GPU.

[0097] Table 2 Experimental parameter settings

[0098]

[0099] Table 3 shows the model accuracy results of the proposed method in different incremental transfer diagnostic tasks. As can be seen from the table, the proposed method achieved good experimental performance in all incremental transfer diagnostic tasks. Specifically, as the number of stages increases, the model's diagnostic accuracy slowly decreases. This indicates that with the increasing number of stages, the number of newly emerging fault categories also increases, thus increasing the difficulty of the diagnostic task. Even so, the proposed diagnostic method still achieved excellent performance, meaning that it can effectively retain the diagnostic knowledge learned from a small number of remaining samples. Even after multiple stages, the diagnostic accuracy of the updated model obtained by the proposed method in each task is generally above 90%, demonstrating the effectiveness and robustness of the proposed method. Furthermore, the training data used in the initial model and the training data used in the incremental stages were collected from different devices, resulting in some data distribution differences. The high diagnostic accuracy achieved by the proposed method also indicates that the data distribution differences between the source and target domain samples have been effectively reduced.

[0100] Table 3. Model fault diagnosis accuracy under different incremental fault diagnosis tasks.

[0101]

[0102] To further verify the superiority of the method in this embodiment, several classic fault diagnosis algorithms were compared with the method of this invention in various incremental fault diagnosis tasks:

[0103] Comparison Method 1: The baseline method uses only a feature extractor and a fault classifier for fault classification. The feature extractor structure is consistent with the proposed method. The loss function of the baseline method is consistent with the loss function in the initial training stage, and the model structure in each incremental stage is consistent with the model in the initial stage.

[0104] Comparison Method 2: DACNN, the network consists of a feature extractor, a label predictor and a domain classifier. The feature extractor structure is consistent with the proposed method. It extracts domain-general features of the source domain and the target domain through an adversarial training strategy.

[0105] The average accuracy of the method of this invention and the above-described comparative method at each stage in various incremental diagnostic tasks is as follows: Figure 3 As shown (each task is represented from left to right as the inventive method, the comparative method, and the comparative method 2), it can be seen that the method of the present invention has achieved the highest average diagnostic accuracy. The method of the present invention can effectively solve the catastrophic forgetting problem in the data-driven model and realize the rapid identification and diagnosis of emerging faults in industrial scenarios.

[0106] In summary, this invention addresses the challenges of identifying new faults and catastrophic knowledge loss when performing incremental fault diagnosis on massive amounts of data from different devices. It proposes an incremental meta-learning online fault diagnosis method for industrial streaming data, enabling online model updates and accurate fault diagnosis across multiple diagnostic tasks. This method has significant application value in actual industrial production.

[0107] Finally, it should be noted that although the implementation of the present invention has been described in detail with reference to examples, those skilled in the art will readily understand that any modifications, substitutions, and improvements made without departing from the spirit and principles of the invention as set forth in the appended claims should be included within the scope of protection of the present invention.

Claims

1. An incremental meta-learning online fault diagnosis method for industrial streaming data, characterized in that, The method comprises the following steps: S1, collecting vibration acceleration response signals of different fault modes of equipment under different working conditions, and dividing them into multiple groups of fault data sets in different diagnosis stages; S2, building an initial fault diagnosis model and obtaining initial model parameters through loss function training; S3, select a small batch of representative samples from the old task training dataset by introducing the Herding algorithm and store them in the data cache module , and based on the new task fault sample dataset and the old diagnostic task test data module Construct a joint test dataset , The dataset constructed by extracting data from the fault sample dataset of the new task, and based on the data cache module and the fault sample dataset of the new task to construct the joint training dataset of the new stage , represents the original sample, represents the fault class label, represents the task ID from which the sample comes S4, from the new stage of joint training data set randomly sample from the meta-training set to form a multi-task training set, and update the parameters of each sub-task model; S5, updating the general model parameters according to the parameters of each sub-task model through a dynamic balance meta-update strategy; In step S2, the initial fault diagnosis model comprises a feature mapping network and a fault classification network , wherein the feature mapping network comprises a convolutional layer, a batch normalization layer, a max-pooling layer, four ResNet modules, and a fully connected layer; the fault classification network comprises a fully connected layer for determining the fault category of the input sample; During training, the parameters of the model are first initialized, and then the initial model is trained through a cross-entropy loss function with label smoothing. The specific steps include: S21, smoothing the labels of the initial phase dataset the labels of the real samples are smoothed: ； wherein, denotes the smoothed sample label, denotes the label smoothing hyperparameter, denotes the original failure sample, denotes the sample corresponding failure class label, denotes the initial phase dataset sample number, denotes the number of failure modes in the first diagnostic phase; S22、calculating the cross-entropy loss function after label smoothing training and updating the model parameters: ； wherein, represents a function processing on the inside of the parentheses, Softmax represents the average value of the outputs of all the fault samples in the dataset after the operation in the parentheses, represents the output of the original fault sample after the feature extraction including the feature mapping network represents the output of the original fault sample after the feature extraction including the feature mapping network and the fault classification network .​​ 2. The incremental meta-learning online fault diagnosis method for industrial streaming data according to claim 1, characterized in that, In step S1, the collected samples are divided into multiple groups of fault diagnosis task datasets wherein represent different diagnosis stages, represent the final diagnosis stage, represent diagnosis stages of the fault dataset, represent the number of fault patterns existing in diagnosis stages .

3. The incremental meta-learning online fault diagnosis method for industrial streaming data according to claim 1, characterized in that, In step S2, the initial model parameters are obtained by training through a cross-entropy loss function with label smoothing.

4. The incremental meta-learning online fault diagnosis method for industrial streaming data according to claim 1, characterized in that, In step S3, the Herding algorithm is introduced to randomly select a small batch of samples from the previous phase and put them into the data cache module and the old diagnosis task test data module , and a part of the data from the new task fault sample data set is selected as the test data set , and the old diagnosis task test data module together constitute the joint test data set , and then randomly select a group of samples from the data cache module and the new task fault sample data set to form the joint training data set of the new phase , wherein, represents the original sample, represents the fault category label, represents the task ID from which the sample comes.

5. The incremental meta-learning online fault diagnosis method for industrial streaming data according to claim 4, characterized in that, Step S4 specifically includes the following contents: S41, the joint training data set of the new stage According to the task ID, a plurality of sub-task data sets are obtained , wherein , represents different sub-tasks; represents the number of samples in the th sub-task data set; S42, output the probability distribution of the model on each sub-task data set sample, and calculate the training loss function of each sub-task: S43, performing on each subtask wheel iteration update, update will feature mapping network and fault classification network isolation, feature mapping network The parameters of the feature mapping network are updated on all subtasks to obtain new feature mapping network parameters , fault classification network The parameters of the fault classification network are updated separately on each subtask to obtain fault classification network parameters for each old task and new task , wherein represents the fault classification network parameters of the thsubtask.

6. The incremental meta-learning online fault diagnosis method for industrial streaming data according to claim 5, characterized in that, In step S42, the training loss function is ； In the formula, represents the first subtask data set the first fault sample, represents the first label of the fault sample in the first subtask data set, represents the number of fault modes of the first diagnosis stage, represents the output of the original fault sample after feature mapping network feature extraction, represents the output of the original fault sample after feature mapping network and fault classification network.

7. The incremental meta-learning online fault diagnosis method for industrial streaming data according to any one of claims 1-6, characterized in that, Step S5 includes the following steps: S51, calculate the gradient of the general model update: S52, in the training process, introduce a momentum-based dynamic weight factor The control model accelerates the gradient update in the early learning stage and slows down the learning in the later learning stage: ； wherein, is a decay rate that varies with the learning phase, representing the speed at which the model learns new knowledge, represents a different diagnostic phase, represents the final diagnostic phase; S53、 according to the dynamically changed weight factor updating the general model parameters, so that the general model learns in the average update direction of all tasks, ensuring that the model can guarantee both new tasks and old tasks; S54, jointly test data sets from each new and old diagnosis task stage The comprehensive diagnosis performance of the updated general diagnosis model on all diagnosis tasks is tested by random sampling.

8. The incremental meta-learning online fault diagnosis method for industrial streaming data according to claim 7, characterized in that, In step S51, the gradient of the general model update is ； wherein, denotes the stage update obtains general model parameters, denotes the parameters of the feature mapping network , denotes the parameters of the fault classification network .

9. The incremental meta-learning online fault diagnosis method for industrial streaming data according to claim 7, characterized in that, The model updating method in step S53 is as follows: ； in, This represents the updated general model parameters. Indicates the first The phased update yields the general model parameters. Representation Feature Mapping Network The parameters, Fault classification network The parameters.

Images