Adaptive fault diagnosis method and device based on active learning and related medium

Abstract

The application discloses an adaptive fault diagnosis method and device based on active learning and related medium, and the method comprises the following steps: obtaining historical fault data of equipment, wherein the historical fault data comprises a source domain and a target domain; sampling fault data in the target domain based on an energy-based sampling strategy to obtain a target sample set of the target domain; labeling the fault data in the target sample set, and merging the target sample set and the fault data in the source domain into a sample data set; training and learning sample data in the sample data set by using a supervised machine learning model to construct a fault diagnosis model; and using the fault diagnosis model to perform fault diagnosis classification on specified equipment state data. The valuable fault data is selected by using the energy-based sampling strategy, and is used for training and learning of the fault diagnosis model, so that the fault diagnosis accuracy and efficiency of the equipment can be improved.

Description

Technical Field

[0001] This invention relates to the field of equipment testing technology, and in particular to an adaptive fault diagnosis method, device, and related media based on active learning. Background Technology

[0002] Equipment operation and maintenance management is a crucial aspect of the stable operation of enterprises and factories. With the continuous advancement of the industrial revolution, traditional equipment management methods are gradually shifting towards automation. Previously, equipment operation and maintenance largely relied on manual experience. Due to the inability to obtain equipment fault information in a timely manner, planned maintenance and reactive maintenance were generally adopted. This approach could lead to over-maintenance, resource waste, or even safety accidents. In recent years, with the popularization of "smart factories," more and more factories are beginning to rely on technologies such as artificial intelligence, model mechanisms, and big data, employing intelligent algorithm models to effectively monitor and manage equipment status.

[0003] Equipment fault diagnosis has always been a core part of equipment management. Currently, commonly used methods for fault diagnosis include pure physical models based on equipment mechanisms, and spectral analysis of collected signal data, using frequency or amplitude magnitude combined with thresholds to determine equipment faults. In addition, popular models such as convolutional neural networks and adversarial transfer learning (ADLPR) provide automated fault diagnosis. However, traditional mechanistic models require understanding of the equipment's mechanisms and rely heavily on expert experience, making them difficult for ordinary personnel. Spectral analysis methods can make approximate judgments for common fault types, but their effectiveness diminishes significantly for complex signal data. While popular neural network models achieve automated diagnosis, they require a large amount of manually labeled fault data, resulting in high costs. However, in actual production environments, fault occurrences are infrequent, and fault data accumulation is limited. Therefore, how to build a highly accurate fault diagnosis model with limited fault data is a problem that those skilled in the art need to solve. Summary of the Invention

[0004] This invention provides an adaptive fault diagnosis method, apparatus, computer device, and storage medium based on active learning, aiming to improve the accuracy and efficiency of fault diagnosis for equipment.

[0005] In a first aspect, embodiments of the present invention provide an adaptive fault diagnosis method based on active learning, comprising:

[0006] Acquire historical fault data of the device, wherein the historical fault data includes a source domain and a target domain, the source domain is fault data with labels, and the target domain is fault data without labels;

[0007] Fault data in the target domain are sampled using an energy-based sampling strategy to obtain a target sample set for the target domain.

[0008] Label the fault data in the target sample set, and merge the target sample set with the fault data in the source domain into a sample dataset;

[0009] A fault diagnosis model is constructed by training the sample data in the sample dataset using a supervised machine learning model.

[0010] The fault diagnosis model is used to classify and diagnose faults in the specified equipment status data.

[0011] Secondly, embodiments of the present invention provide an adaptive fault diagnosis device based on active learning, comprising:

[0012] A data acquisition unit is used to acquire historical fault data of the device, wherein the historical fault data includes a source domain and a target domain, the source domain is fault data with tags, and the target domain is fault data without tags;

[0013] A data sampling unit is used to sample fault data in the target domain using an energy-based sampling strategy to obtain a target sample set for the target domain.

[0014] A data merging unit is used to label the fault data in the target sample set and merge the target sample set with the fault data in the source domain into a sample dataset.

[0015] The model building unit is used to train and learn from the sample data in the sample dataset using a supervised machine learning model in order to build a fault diagnosis model.

[0016] The diagnostic classification unit is used to perform fault diagnosis classification on the specified equipment status data using the fault diagnosis model.

[0017] Thirdly, embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the adaptive fault diagnosis method based on active learning as described in the first aspect.

[0018] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the adaptive fault diagnosis method based on active learning as described in the first aspect.

[0019] This invention provides an adaptive fault diagnosis method, apparatus, computer device, and storage medium based on active learning. The method includes: acquiring historical fault data of a device, wherein the historical fault data includes a source domain and a target domain, the source domain being labeled fault data and the target domain being unlabeled fault data; sampling the fault data in the target domain using an energy-based sampling strategy to obtain a target sample set for the target domain; labeling the fault data in the target sample set and merging the target sample set with the fault data in the source domain to form a sample dataset; training the sample data in the sample dataset using a supervised machine learning model to construct a fault diagnosis model; and using the fault diagnosis model to perform fault diagnosis classification on specified device state data. This invention selects valuable fault data using an energy-based sampling strategy and, after labeling, uses it together with the fault data in the source domain for training the fault diagnosis model, thus improving the accuracy and efficiency of fault diagnosis for devices. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating an adaptive fault diagnosis method based on active learning, provided in an embodiment of the present invention.

[0022] Figure 2 A schematic diagram of a sub-process of an adaptive fault diagnosis method based on active learning provided in an embodiment of the present invention;

[0023] Figure 3 A schematic block diagram of an adaptive fault diagnosis device based on active learning provided in an embodiment of the present invention;

[0024] Figure 4 This is a schematic block diagram of an adaptive fault diagnosis device based on active learning, provided as an embodiment of the present invention. Detailed Implementation

[0025] 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, not all, of the embodiments of the present invention. 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.

[0026] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0027] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0028] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0029] Please see below. Figure 1 , Figure 1 The flowchart of an adaptive fault diagnosis method based on active learning provided in an embodiment of the present invention specifically includes steps S101 to S105.

[0030] S101. Obtain historical fault data of the device, wherein the historical fault data includes a source domain and a target domain, the source domain is fault data with tags, and the target domain is fault data without tags;

[0031] S102. The fault data in the target domain is sampled using an energy-based sampling strategy to obtain a target sample set for the target domain.

[0032] S103. Label the fault data in the target sample set, and merge the target sample set with the fault data in the source domain into a sample dataset;

[0033] S104. The sample data in the sample dataset is trained and learned through a supervised machine learning model to construct a fault diagnosis model.

[0034] S105. Use the fault diagnosis model to perform fault diagnosis and classification on the specified equipment status data.

[0035] In this embodiment, historical fault data is first acquired, which includes labeled fault data (i.e., the source domain) and unlabeled fault data (i.e., the target domain). Valuable fault data is selected from the target domain based on an energy sampling strategy as a target sample set, which is then combined with the fault data in the source domain to form a sample dataset. This sample dataset is then used for model training to construct the fault diagnosis model, which can then be used to classify and diagnose faults in specified devices.

[0036] This embodiment selects valuable fault data through an energy-based sampling strategy, and after labeling, uses it together with fault data in the source domain for training the fault diagnosis model, thereby improving the accuracy and efficiency of fault diagnosis for equipment.

[0037] Factory equipment generates a large amount of data during daily operation, which is collected by sensors and returned to the front end for analysis by staff. This data reflects the equipment status from various angles (such as temperature, speed, vibration, etc.) and is untagged. Maintenance personnel need to monitor the equipment status based on this untagged data, detect malfunctions in real time, and, if a malfunction occurs, promptly determine the type of fault so that an effective maintenance plan can be developed. This allows for real-time control of the equipment status and reduces losses caused by delayed responses to sudden equipment failures.

[0038] Based on the above, this invention proposes an adaptive fault diagnosis method based on active learning for diagnosing faults in real-time acquired unlabeled equipment status data. First, based on accumulated historical fault data, a small portion of labeled fault data is obtained, containing several common equipment fault types. The domain studied by this fault data is called the source domain. The unlabeled fault diagnosis problem studied in this embodiment is the target domain. The problem to be solved is that, given a set of equipment operating data, the fault diagnosis model needs to match the "correct" fault type (label). Energy-based modeling (EBM) seeks an appropriate energy function so that the energy of correct inputs and outputs in the sample is lower than the energy of incorrect inputs and outputs. Based on this idea, this invention proposes an active learning fault diagnosis method based on an "energy" strategy. First, valuable samples are selected and labeled in the unlabeled target domain according to the "energy" strategy. The newly obtained labeled samples are added to the existing labeled fault data for training. The final trained model assigns the lowest energy to correct labels and higher energy to other incorrect labels. This process continues through continuous model iteration and updates, resulting in an increasing number of labeled samples. Once a sufficient number of labeled samples have been accumulated, a machine learning model is used to complete fault diagnosis. This method significantly reduces the cost of manual labeling and enables automated fault diagnosis based on a small number of samples.

[0039] In one embodiment, such as Figure 2 As shown, S102 includes steps S201 to S203.

[0040] S201. Calculate the free energy for each fault data in the target domain, and sort the calculation results to select the top a% of fault data with the largest free energy as the first candidate sample set.

[0041] S202. Calculate the uncertainty for each fault data in the first candidate sample set, and sort the calculation results to select the top b% of fault data with the greatest uncertainty as the second candidate sample set.

[0042] S203. Repeatedly select the top a% of fault data with the largest free energy and the top b% of fault data with the largest uncertainty until the fault data in the second candidate sample set meets the preset quantity requirement, and use the second candidate sample set as the target sample set.

[0043] In this embodiment, when sampling the target domain, the source domain is denoted as S and the target domain is denoted as T, where T = T l ∪T u T l T is the set of samples selected from the target domain.u For the set of unselected samples, initialize And the sampling process is completed based on the uncertainty of sample selection. Specifically, T is first calculated. u The free energy of each fault data point is calculated, and the free energies are arranged from largest to smallest. The top a% of fault data are then selected as the first candidate sample set. Next, the uncertainty of each fault data point in the first candidate sample set is calculated, and the results are similarly arranged from largest to smallest. The top b% of fault data are then selected as the second candidate sample set, denoted as [the second candidate sample set]. at this time Repeat the above steps until a T is selected. l The sample size meets the predetermined requirements.

[0044] In one embodiment, S201 includes:

[0045] The probability p(x) of each fault data in the target domain is calculated according to the following formula:

[0046] p(x)=∑ y p(x, y) = ∑ y exp(-E(x,y)) / Z

[0047] In the formula, x represents the input fault data, y represents the output label, E(x, y) represents the energy function, p(x, y) represents the joint probability density function, and Z = ∑ x ∑ y exp(-E(x, y)) represents the partition function;

[0048] And, calculate the probability p(x) of each fault data in the target domain according to the following formula:

[0049]

[0050] In the formula, F(x) represents the free energy function;

[0051] Establish the relationship between the energy function E(x, y) and the free energy function F(x):

[0052] F(x) = -logΣ y exp(-E(x,y))

[0053] The fault data in the target domain are sampled using a linear network according to the following formula:

[0054]

[0055]

[0056] In the formula, E(x, y, W) represents the energy function under the linear network, F(x, W) represents the free energy function under the linear network, and W = (w1, w2…w…) m )∈R m×n Let m represent the parameter weight matrix of the linear network, m represent the number of fault data, n represent the dimension, T represent the transpose operation, and j represent the j-th fault category.

[0057] Furthermore, S202 includes:

[0058] The uncertainty U(x) is calculated for each fault data point according to the following formula:

[0059]

[0060] In the formula, E(x, y) * W) represents the energy function for the minimum energy value, and E(x, y′, W) represents the energy function for the minimum energy value. * =argmin y E(x, y, W) represents the fault tag with the lowest energy value. A fault label indicating the second lowest energy value.

[0061] In this embodiment, considering that equipment fault diagnosis is essentially a multi-classification problem—that is, determining whether a problem has occurred in the equipment operation based on operational data, or what type of fault it is—the input to the fault diagnosis model is the equipment operational data. Sensors are installed at measuring points on important parts of the equipment, such as the input shaft, output shaft, and various transmission shafts, and the vibration signal data collected from each measuring point is the data to be analyzed.

[0062] The entire model is primarily built using active learning methods. In actual production, equipment often operates normally with very few failures, resulting in a small amount of accumulated fault data. Building a model based on this limited sample size would be extremely difficult. Active learning, however, can select valuable samples from the unlabeled target domain for manual annotation, significantly reducing the cost of manual annotation while expanding the number of labeled samples. There are many sampling strategies for active learning. Unlike previous methods, this embodiment uses an energy-based model for sampling. During sampling, it minimizes the difference in energy distribution between the source and target domains to select the "most valuable samples." In an energy-based model, the essence is to find an appropriate energy function so that the energy of correct inputs and outputs in the sample is lower than that of incorrect inputs and outputs. The equipment fault diagnosis task essentially involves matching the model's input with a correct fault diagnosis result (label). Here, the collected vibration signal is denoted as x as the input to the model, the output label is denoted as y, y∈(1,2…m), and the energy function is denoted as E(x,y). Then, the probability of each target sample appearing in the target domain is:

[0063] p(x)=∑ y p(x, y) = ∑ y exp(-E(x,y)) / Z (1)

[0064] Where p(x, y) is the joint probability density function, Z = ∑ x ∑ y exp(-E(x,y)) is the partition function, a concept frequently used in equilibrium statistical physics. The partition function is equivalent to the free energy.

[0065] According to the definition of the energy function, the energy distribution of the model trained only with labeled source domain data is smaller than the energy distribution of the unlabeled target domain. The higher the energy of the sample in the target domain, the better it reflects the information of the target domain. Therefore, the valuable sample in the target domain is defined as the sample with a higher energy function. In (1), the partition function Z is fixed. The higher the energy function E(x, y), the smaller the probability p. That is, the sample with a lower probability is the "most valuable sample". Since Z is difficult to calculate, the free energy function F(x) is a function hidden in the energy model. It can also reflect the probability p(x) of the sample. The formula is as follows:

[0066]

[0067] Combining equations (1) and (2), the relationship between the energy function E(x, y) and the free energy function F(x) can be written as follows:

[0068] F(x) = -log∑ y exp(-E(x,y)) (3)

[0069] The core of active learning involves using a strategy to sample unlabeled samples in the target domain. These representative samples are then manually labeled, increasing the number of labeled samples. Once a sufficient number of samples has been accumulated, a supervised method can be used to build a fault classification model. In this step, the energy function and free energy function are defined, and the subsequent energy-based sampling process is based on these definitions.

[0070] To simplify the sampling process, a single-layer linear network is used here to sample the samples in the target domain. The input to the linear network is the source domain sample x∈R. n The dimension is n, and the fault type is y, y∈(1, 2…m). The parameter weight matrix of the linear layer is W=(w1, w2…w…m). m )∈R m×n The essence of the energy function is to establish the relationship between the input x and the output y. This relationship assigns lower "energy" to samples with correct labels and higher "energy" to samples with incorrect labels. Applying this idea to a single-layer linear network, the energy function and free energy function can be defined as follows:

[0071]

[0072]

[0073] In most cases, there is a discrepancy in the sample distribution between the source and target domains. From an energy perspective, this means there is a discrepancy in the free energy distribution. Therefore, free energy can be used as a measure to reflect the difference between the domains. Furthermore, according to the definition of the energy model, the energy function assigns the lowest energy to the correct input label. Since the data in the source domain consists of accumulated data with correct labels, the free energy of the source domain is lower than that of the target domain. To improve the accuracy of fault diagnosis tasks, it is necessary to minimize the difference in distribution between the source and target domains, that is, to make the free energy distribution regions between the domains as consistent as possible. The energy-based sampling strategy in this embodiment selects samples with higher free energies in the target domain, because these samples better reflect the "features" of the target domain and are more representative typical samples. During the sampling process, such samples are fully ensured by combining domain features and instance uncertainty. Here, the difference between the lowest and second-lowest energies is used to measure the uncertainty U(x) of sample selection, specifically defined as follows:

[0074]

[0075] Here y * =argminyE(x, y, W) is the fault label corresponding to the lowest energy value. This is the fault tag corresponding to the second lowest energy value.

[0076] In one embodiment, the adaptive fault diagnosis method based on active learning further includes:

[0077] According to the following formula, the loss function L is used. null (x, y; W) optimizes the difference in free energy distribution between the source and target domains:

[0078] L null (x,y;W)=E(x,y;W)-F(x;W)

[0079] And, according to the following formula, the regularization term L is applied. fea The cross-domain distribution between the source and target domains is calibrated:

[0080] L fea (x, W)=max(0, F(x, W)-Δ)

[0081] In the formula, Δ=E x～S F(x, W) represents the average value;

[0082] According to the following formula, combined with the loss function L null (x, y; W) and regularization term L fea Construct a sampling loss function L to optimize the sampling process:

[0083]

[0084] In the formula, τ represents the initial hyperparameters, and S∪T l This represents the set of the source domain and the target sample set.

[0085] In this embodiment, to reduce the differences in free energy distribution between domains and better achieve the inter-domain adaptation problem, the loss function is defined as L. null (x, y; W). Meanwhile, to further reduce distribution bias, a regularization term L is added. fea This can, to some extent, calibrate cross-domain distributions. Therefore, combining the loss function L... null (x, y; W) and regularization term L fea The sampling loss function can then be obtained, and the sampling loss function can be used to optimize the sampling process.

[0086] In one embodiment, S104 includes:

[0087] A deep residual network is used to extract features from the sample data to obtain the corresponding feature vectors.

[0088] The feature vector is input into the XGboost model, which outputs the classification result corresponding to the feature vector, and the classification result is used as the fault diagnosis classification result corresponding to the sample data.

[0089] In this embodiment, after the preceding steps, valuable samples can be selected from the original unlabeled target domain for labeling, thus expanding the number of labeled samples. When the number of labeled samples accumulates to a certain extent, supervised learning methods can be used to diagnose and classify equipment faults.

[0090] The vibration and other signal data collected by equipment are generally quite complex, containing both time and frequency domain information. Furthermore, the influence of external noise or other factors further complicates the data, making feature extraction using ordinary signal decomposition somewhat inadequate. This embodiment employs a deep residual network (ResNet) from the field of image processing to perform deep feature extraction on labeled sample data. ResNet is more stable than previous convolutional networks and is better suited for extracting deeper features. The feature vectors are then used as input to the XGBoost model, which performs the multi-class fault diagnosis task.

[0091] Therefore, the fault classification task is completed by ResNet and XGboost. ResNet is responsible for deep feature extraction of the input sample data to obtain feature vectors, and the subsequent fault diagnosis results are completed by the XGboost model. The XGboost model has good performance in multi-classification tasks and can achieve the goal of equipment fault diagnosis.

[0092] This embodiment proposes an adaptive fault diagnosis method based on active learning for equipment fault diagnosis tasks. Addressing the problem of limited fault data in practice, which prevents the direct establishment of fault classification models using supervised learning methods, this method proposes an active learning approach based on an "energy" concept sampling strategy. Building upon a linear network, energy and free energy functions are defined. The energy-based sampling process reduces the difference in free energy distribution between the source and target domains by selecting valuable target samples, achieving inter-domain adaptation. Regularization terms further calibrate these inter-domain differences. Valuable samples are labeled, and when a sufficient number of samples are accumulated, ResNet+XGboost is used to achieve the final fault diagnosis classification. This method, while maintaining fault diagnosis accuracy, is more scientific and effective than previous methods of randomly selecting data for labeling, and also significantly reduces the cost of large-scale manual labeling.

[0093] In one embodiment, the step of using a deep residual network to extract features from the sample data to obtain corresponding feature vectors includes:

[0094] Feature extraction of the sample data is performed using cascaded residual units according to the following formula:

[0095] y l =h(x l )+F(x l W l )

[0096] x l+1 =f(y l )

[0097]

[0098] In the formula, x L x represents the final extracted feature vector. l and x l+1 Let W1 represent the input and output of the l-th residual unit, respectively, where W1 is the weight matrix, F is the residual function, and h(x) is the weight function. l )=x l Let f denote the identity mapping, and f be the activation function.

[0099] In this embodiment, the deep residual network (ResNet) consists of two residual units connected in series, and the kernel size is 3×3. The feature vector representation is obtained after passing through two residual units. Specifically,

[0100] The expanded sample obtained after the previous steps is used as the input x of the model, and the fault label is denoted as y. The entire deep residual network consists of two residual units. The calculation process of a single residual unit is as follows:

[0101] y l =h(x l )+F(x l W l (7)

[0102] x l+1 =f(y l (8)

[0103] Where x l and x l+1 W represents the input and output of the l-th residual unit, respectively. l Here, F is the weight matrix, F is the residual function, representing the learned residuals, and h(x) is the weight matrix. l )=x lLet f represent the identity mapping, and f be the activation function (e.g., ReLU). Based on the above equation, the learned features from the shallow layer l to the deep layer L can be obtained as follows:

[0104]

[0105] Figure 3 This is a schematic block diagram of an adaptive fault diagnosis device 300 based on active learning, provided in an embodiment of the present invention. The device 300 includes:

[0106] The data acquisition unit 301 is used to acquire historical fault data of the device, wherein the historical fault data includes a source domain and a target domain, the source domain is fault data with tags, and the target domain is fault data without tags.

[0107] The data sampling unit 302 is used to sample fault data in the target domain using an energy-based sampling strategy to obtain a target sample set for the target domain.

[0108] The data merging unit 303 is used to label the fault data in the target sample set and merge the target sample set with the fault data in the source domain into a sample dataset.

[0109] The model building unit 304 is used to train and learn the sample data in the sample dataset through a supervised machine learning model in order to build a fault diagnosis model.

[0110] The diagnostic classification unit 305 is used to perform fault diagnosis classification on the specified equipment status data using the fault diagnosis model.

[0111] In one embodiment, such as Figure 4 As shown, the data sampling unit 302 includes:

[0112] The first sorting unit 401 is used to calculate the free energy of each fault data in the target domain and sort the calculation results to select the top a% of fault data with the largest free energy as the first candidate sample set.

[0113] The second sorting unit 402 is used to calculate the uncertainty of each fault data in the first candidate sample set and sort the calculation results to select the top b% of fault data with the greatest uncertainty as the second candidate sample set.

[0114] The repeated execution unit 403 is used to repeatedly execute the selection of the top a% of fault data with the largest free energy and the top b% of fault data with the largest uncertainty until the fault data in the second candidate sample set meets the preset quantity requirement, and the second candidate sample set is used as the target sample set.

[0115] In one embodiment, the first sorting unit 401 includes:

[0116] The first probability calculation unit is used to calculate the probability p(x) of the occurrence of each fault data in the target domain according to the following formula:

[0117] p(x)=∑ y p(x, y) = ∑ y exp(-E(x,y)) / Z

[0118] In the formula, x represents the input fault data, y represents the output label, E(x, y) represents the energy function, p(x, y) represents the joint probability density function, and Z = ∑ x ∑ y exp(-E(x, y)) represents the partition function;

[0119] The second probability calculation unit is used to calculate the probability p(x) of occurrence of each fault data in the target domain according to the following formula:

[0120]

[0121] In the formula, F(x) represents the free energy function;

[0122] The relation-establishing unit is used to establish the relationship between the energy function E(x, y) and the free energy function F(x):

[0123] F(x) = -log∑ y exp(-E(x,y))

[0124] The data sampling unit is used to sample fault data in the target domain using a linear network according to the following formula:

[0125]

[0126]

[0127] In the formula, E(x, y, W) represents the energy function under the linear network, F(x, W) represents the free energy function under the linear network, and W = (w1, w2…w…) m )∈R m×n Let m represent the parameter weight matrix of the linear network, m represent the number of fault data, n represent the dimension, T represent the transpose operation, and j represent the j-th fault category.

[0128] In one embodiment, the second sorting unit 402 includes:

[0129] An uncertainty calculation unit is used to calculate the uncertainty U(x) for each fault data according to the following formula:

[0130]

[0131] In the formula, E(x, y) * W) represents the energy function for the minimum energy value, and E(x, y′, W) represents the energy function for the minimum energy value. * =argmin y E(x, y, W) represents the fault tag with the lowest energy value. A fault label indicating the second lowest energy value.

[0132] In one embodiment, the active learning-based adaptive fault diagnosis device 300 further includes:

[0133] The first optimization unit is used to apply the loss function L according to the following formula. null (x, y; W) optimizes the difference in free energy distribution between the source and target domains:

[0134] L null (x,y;W)=E(x,y;W)-F(x;W)

[0135] The calibration unit is used for, and, according to the following formula, employs the regularization term L. fea The cross-domain distribution between the source and target domains is calibrated:

[0136] L fea (x, W)=max(0, F(x, W)-Δ)

[0137] In the formula, Δ=E x～S F(x, W) represents the average value;

[0138] The second optimization unit is used to combine the loss function L according to the following formula. null (x, y; W) and regularization term L fea Construct a sampling loss function L to optimize the sampling process:

[0139]

[0140] In the formula, τ represents the initial hyperparameters, and S∪T l This represents the set of the source domain and the target sample set.

[0141] In one embodiment, the model building unit 304 includes:

[0142] The first feature extraction unit is used to extract features from the sample data using a deep residual network to obtain the corresponding feature vector.

[0143] The result output unit is used to input the feature vector into the XGboost model, and the XGboost model outputs the classification result corresponding to the feature vector, and uses the classification result as the fault diagnosis classification result corresponding to the sample data.

[0144] In one embodiment, the first feature extraction unit includes:

[0145] The second feature extraction unit is used to extract features from the sample data using a series of residual units according to the following formula:

[0146] y l =h(x l )+F(x l W l )

[0147] x l+1 =f(y l )

[0148]

[0149] In the formula, x L x represents the final extracted feature vector. l and x l+1 W represents the input and output of the first residual unit, respectively. l Let F be the weight matrix, F be the residual function, and h(x) be the weight matrix. l )=x l Let f denote the identity mapping, and f be the activation function.

[0150] Since the embodiments of the apparatus and the embodiments of the method correspond to each other, please refer to the description of the embodiments of the method for the embodiments of the apparatus, which will not be repeated here.

[0151] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed, can perform the steps provided in the above embodiments. The storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0152] This invention also provides a computer device, which may include a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it can implement the steps provided in the above embodiments. Of course, the computer device may also include various network interfaces, power supplies, and other components.

[0153] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0154] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. An adaptive fault diagnosis method based on active learning, characterized in that, include: Acquire historical fault data of the device, wherein the historical fault data includes a source domain and a target domain, the source domain is fault data with labels, and the target domain is fault data without labels; Fault data in the target domain are sampled using an energy-based sampling strategy to obtain a target sample set for the target domain. Label the fault data in the target sample set, and merge the target sample set with the fault data in the source domain into a sample dataset; A fault diagnosis model is constructed by training the sample data in the sample dataset using a supervised machine learning model. The fault diagnosis model is used to classify and diagnose faults in the specified equipment status data. The step of sampling fault data in the target domain using an energy-based sampling strategy to obtain a target sample set for the target domain includes: The free energy is calculated for each fault data in the target domain, and the calculation results are sorted to select the top a% of fault data with the largest free energy as the first candidate sample set. The uncertainty is calculated for each fault data in the first candidate sample set, and the calculation results are sorted to select the top b% of fault data with the greatest uncertainty as the second candidate sample set. Repeatedly select the top a% of fault data with the largest free energy and the top b% of fault data with the largest uncertainty until the fault data in the second candidate sample set meets the preset quantity requirement, and then use the second candidate sample set as the target sample set.

2. The adaptive fault diagnosis method based on active learning according to claim 1, characterized in that, The step of calculating the free energy for each fault data point in the target domain and sorting the calculation results to select the top a% of fault data points with the highest free energy as the first candidate sample set includes: The probability p(x) of each fault data in the target domain is calculated according to the following formula: p(x)=∑ y p(x,y)=∑ y exp(-E(x,y)) / Z In the formula, x represents the input fault data, y represents the output label, E(x,y) represents the energy function, p(x,y) represents the joint probability density function, and Z = ∑ x ∑ y exp(-E(x, y)) represents the partition function; And, calculate the probability p(x) of each fault data in the target domain according to the following formula: In the formula, F(x) represents the free energy function; Establish the relationship between the energy function E(x,y) and the free energy function F(x): F(x)=-log∑ y exp(-E(x,y)) The fault data in the target domain are sampled using a linear network according to the following formula: In the formula, E(x,y,W) represents the energy function under the linear network, F(x,W) represents the free energy function under the linear network, and W=(w1,w2…w m )∈R m×n Let m represent the parameter weight matrix of the linear network, m represent the number of fault data, n represent the dimension, T represent the transpose operation, and j represent the j-th fault category.

3. The adaptive fault diagnosis method based on active learning according to claim 2, characterized in that, The step of calculating the uncertainty for each fault data in the first candidate sample set and sorting the calculation results to select the top b% of fault data with the greatest uncertainty as the second candidate sample set includes: The uncertainty U(x) is calculated for each fault data point according to the following formula: In the formula, E(x,y) * E(x,y',W) represents the energy function with the lowest energy value, and E(x,y',W) represents the energy function with the second lowest energy value. * =argmin y E(x,y,W) represents the fault label with the lowest energy value, y' = A fault label indicating the second lowest energy value.

4. The adaptive fault diagnosis method based on active learning according to claim 3, characterized in that, Also includes: According to the following formula, the loss function L is used. null (x,y,W) optimizes the difference in free energy distribution between the source and target domains: L null (x,y,W)=E(x,y,W)-F(x,W) And, according to the following formula, the regularization term L is applied. fea The cross-domain distribution between the source and target domains is calibrated: L fea (x,W)=max(0,F(x,W)-Δ) In the formula, Δ=E x～s F(x,W) represents the average value; According to the following formula, combined with the loss function L null (x,y,W) and regularization term L fea Construct a sampling loss function L to optimize the sampling process: In the formula, τ represents the initial hyperparameters, and S∪T l This represents the set of the source domain and the target sample set.

5. The adaptive fault diagnosis method based on active learning according to claim 1, characterized in that, The step of training and learning the sample data in the sample dataset using a supervised machine learning model to construct a fault diagnosis model includes: A deep residual network is used to extract features from the sample data to obtain the corresponding feature vectors. The feature vector is input into the XGboost model, which outputs the classification result corresponding to the feature vector, and the classification result is used as the fault diagnosis classification result corresponding to the sample data.

6. The adaptive fault diagnosis method based on active learning according to claim 5, characterized in that, The step of using a deep residual network to extract features from the sample data to obtain corresponding feature vectors includes: Feature extraction of the sample data is performed using cascaded residual units according to the following formula: y l =h(x l )+F'(x l ,W l ) x l+1 =f(y l ) In the formula, x L’ x represents the final extracted feature vector. l and x l+1 W represents the input and output of the l-th residual unit, respectively. l Let F' be the weight matrix, h(x) be the residual function, and F' be the residual function. l )=x l Let f denote the identity mapping, and f be the activation function.

7. An adaptive fault diagnosis device based on active learning, characterized in that, include: A data acquisition unit is used to acquire historical fault data of the device, wherein the historical fault data includes a source domain and a target domain, the source domain is fault data with tags, and the target domain is fault data without tags; A data sampling unit is used to sample fault data in the target domain using an energy-based sampling strategy to obtain a target sample set for the target domain. A data merging unit is used to label the fault data in the target sample set and merge the target sample set with the fault data in the source domain into a sample dataset. The model building unit is used to train and learn from the sample data in the sample dataset using a supervised machine learning model in order to build a fault diagnosis model. The diagnostic classification unit is used to perform fault diagnosis classification on the specified equipment status data using the fault diagnosis model. The data sampling unit includes: The first sorting unit is used to calculate the free energy of each fault data in the target domain and sort the calculation results to select the top a% of fault data with the largest free energy as the first candidate sample set. The second sorting unit is used to calculate the uncertainty of each fault data in the first candidate sample set and sort the calculation results to select the top b% of fault data with the greatest uncertainty as the second candidate sample set. The repeated execution unit is used to repeatedly execute the selection of the top a% of fault data with the largest free energy and the top b% of fault data with the largest uncertainty until the fault data in the second candidate sample set meets the preset quantity requirement, and the second candidate sample set is used as the target sample set.

8. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the adaptive fault diagnosis method based on active learning as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the adaptive fault diagnosis method based on active learning as described in any one of claims 1 to 6.

Images