Fault diagnosis method and system based on incremental learning model, and storage medium
By building a SCLIFD diagnostic system based on an incremental learning model, the problems of long diagnostic cycle and low hidden fault detection rate in traditional methods for high-end control valve diagnosis are solved, early fault warning and efficient diagnosis are achieved, and the accuracy and adaptability of fault detection are improved.
Patent Information
- Application Number
- CN202511039014.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-28
- Publication Date
- 2025-10-28
AI Technical Summary
Traditional fault diagnosis methods have long diagnostic cycles and low latent fault detection rates in scenarios with frequent operating condition fluctuations and multiple fault couplings, making it difficult to meet the needs of fast and accurate fault detection for high-end control valves.
A fault diagnosis method based on incremental learning model is adopted. By constructing the SCLIFD incremental learning model, including feature extraction module, supervised comparative knowledge distillation module, priority example selection module and classification module, the fault of high-end control valve is trained and diagnosed using simulated data sets.
It enables early and accurate warning of potential control valve failures, shortens the diagnosis cycle, reduces false alarm and missed alarm rates, improves the adaptability and generalization of fault diagnosis, and provides technical support for predictive maintenance.
Smart Images
Figure CN120850108A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fault diagnosis technology, and more specifically to a fault diagnosis method, system, and storage medium based on an incremental learning model. Background Technology
[0002] Currently, in the field of industrial automation, high-end control valves are core actuators in process industries, and their operating status directly determines the stability and safety of production systems. Faced with threats from a variety of complex faults, such as pressure reducing valve mismatch, gas pipeline damage, loose actuator bolts, external gas path damage, spring performance degradation, and diaphragm damage, traditional fault diagnosis methods mainly rely on expert experience and physical model analysis (such as transfer function modeling and vibration spectrum analysis). However, in real-world scenarios with frequent fluctuations in operating conditions and multiple coupled faults, these methods suffer from drawbacks such as long diagnostic cycles and low detection rates of latent faults.
[0003] Therefore, how to improve the efficiency and accuracy of fault detection is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0004] In view of this, the present invention provides a fault diagnosis method, system and storage medium based on an incremental learning model, which overcomes the above-mentioned defects.
[0005] In order to achieve the above object, the present invention adopts the following technical solutions:
[0006] A fault diagnosis method based on an incremental learning model, comprising the following steps:
[0007] A simulation dataset was constructed based on step signal test data of multiple fault modes;
[0008] An initial SCLIFD incremental learning model is constructed, and the initial SCLIFD incremental learning model is trained using the simulated dataset to generate an SCLIFD incremental learning model; the initial SCLIFD incremental learning model includes a feature extraction module, a supervised contrastive knowledge distillation module, a priority example selection module, and a classification module;
[0009] Acquire the data to be diagnosed, input the data to be diagnosed into the SCLIFD incremental learning model, and obtain the fault diagnosis result.
[0010] Optionally, the steps for obtaining the simulated dataset are as follows:
[0011] A control valve failure reproduction test bench was built, and multiple target failure modes were defined. For each of the target failure modes, the failure was reproduced on the control valve failure reproduction test bench using a controllable physical failure injection method, and the type and degree of the injected failure were recorded.
[0012] Under various operating parameters of the target fault modes, step control signals of different amplitudes and directions are applied to the target control valve, and valve stroke signals, drive data and air supply pressure signals are collected at the same time to form raw time series data.
[0013] The original time series data is preprocessed and segmented to obtain independent step response samples, and each step response sample is labeled with metadata tags.
[0014] The simulated dataset used to train and evaluate the SCLIFD incremental learning model is constructed based on the step response samples and their corresponding metadata tags.
[0015] Optionally, the test bench includes a target control valve, an intelligent valve positioner, an actuator, a gas supply system, a data acquisition system, and a control system.
[0016] Optionally, the target failure modes include pressure reducing valve mismatch, loose bolts, pipeline damage, pipeline flattening, gas path blockage, and diaphragm wear.
[0017] Optionally, the step response sample acquisition step is as follows:
[0018] The original time series data is format-converted, cleaned, and time-aligned to generate preprocessed data;
[0019] The preprocessed data is segmented into independent step response samples based on the step control signal.
[0020] Optionally, the initial SCLIFD incremental learning model includes:
[0021] The feature extraction module uses a ResNet18 network to extract deep features from the input data;
[0022] The supervised comparative knowledge distillation module is used to guide the student model to learn discriminative features based on the deep features through the teacher model and output feature samples.
[0023] The priority example selection module is used to calculate and scale the sample difference based on the feature samples and cached samples, and update the cached samples in the memory cache.
[0024] The classification module uses a balanced random forest classifier to output fault classifications.
[0025] A fault diagnosis system based on an incremental learning model includes:
[0026] The dataset construction module is used to construct a simulated dataset based on step signal test data of multiple fault modes;
[0027] The model building module is used to build an initial SCLIFD incremental learning model, and train the initial SCLIFD incremental learning model using the simulated dataset to generate a new SCLIFD incremental learning model. The initial SCLIFD incremental learning model includes a feature extraction module, a supervised contrastive knowledge distillation module, a priority example selection module, and a classification module.
[0028] The fault diagnosis module is used to input the data to be diagnosed into the SCLIFD incremental learning model to obtain the fault diagnosis results.
[0029] A storage medium storing executable instructions, which, when executed by a processor, implement the aforementioned fault diagnosis method based on an incremental learning model.
[0030] As can be seen from the above technical solutions, the present invention provides a fault diagnosis method, system, and storage medium based on an incremental learning model, which has the following advantages compared with the prior art:
[0031] 1. By continuously learning and dynamically updating the model, fault characteristics and their evolution patterns can be captured more accurately, enabling early and accurate warnings of potential control valve faults;
[0032] 2. By introducing an incremental learning strategy and a balanced random forest classifier, the adaptability and generalization ability of the fault diagnosis model under small sample size conditions are improved;
[0033] 3. Combining incremental learning mechanisms with simulated datasets not only shortens the diagnostic cycle and reduces false alarm and false negative rates, but also provides strong technical support for predictive maintenance of control valves. Attached Figure Description
[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0035] Figure 1 This is a schematic diagram of the method flow provided by the present invention;
[0036] Figure 2(a) is a comprehensive simulation image of pressure reducing valve mismatch provided by the present invention; Figure 2(b) is a simulation image of actuator bolt loosening provided by the present invention; Figure 2(c) is a comprehensive simulation image of gas pipeline damage provided by the present invention; Figure 2(d) is a simulation image of gas pipeline flattening provided by the present invention; Figure 2(e) is a simulation image of external gas path blockage of positioner mechanical structure provided by the present invention; Figure 2(f) is a simulation image of actuator diaphragm wear provided by the present invention.
[0037] Figure 3 This is a schematic diagram of the training process of the SCLIFD incremental learning model provided by the present invention;
[0038] Figure 4(a) shows the classification confusion matrix of the SCLIFD incremental learning model under 5 fault modes provided by the present invention; Figure 4(b) shows the classification confusion matrix of the SCLIFD incremental learning model under 10 fault modes provided by the present invention. Detailed Implementation
[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0040] This invention discloses a fault diagnosis method based on an incremental learning model, such as... Figure 1 As shown, the specific steps are as follows:
[0041] Step 1: Construct a simulation dataset based on step signal test data of multiple fault modes;
[0042] Step 2: Construct an initial SCLIFD incremental learning model. Train the initial SCLIFD incremental learning model using a simulated dataset to generate a new SCLIFD incremental learning model. The initial SCLIFD incremental learning model includes a feature extraction module, a supervised contrastive knowledge distillation module, a priority example selection module, and a classification module.
[0043] Step 3: Obtain the data to be diagnosed and input it into the SCLIFD incremental learning model to obtain the fault diagnosis results.
[0044] In one embodiment, the steps for obtaining the simulated dataset are as follows:
[0045] A control valve failure reproduction test bench was built, and multiple target failure modes were defined. For each target failure mode, a controllable physical failure injection method was used to reproduce the failure on the control valve failure reproduction test bench, and the type and degree of the injected failure were recorded.
[0046] Under various operating parameters of different target fault modes, step control signals of different amplitudes and directions are applied to the target control valve, and valve stroke signals, drive data and air supply pressure signals are collected at the same time to form raw time series data.
[0047] The original time series data is preprocessed and segmented to obtain independent step response samples, and each step response sample is labeled with metadata tags.
[0048] A simulated dataset for training and evaluating the SCLIFD incremental learning model is constructed based on step response samples and corresponding metadata tags.
[0049] In one embodiment, the test bench includes a target control valve, an intelligent valve positioner, an actuator, an air supply system, a data acquisition system, and a control system.
[0050] In one embodiment, the target failure modes include pressure reducing valve mismatch, loose bolts, broken tubing, flattened tubing, blocked gas path, and diaphragm wear.
[0051] In one embodiment, the label includes at least the fault type, fault severity, step amplitude, step direction, initial position, and gas supply pressure.
[0052] Furthermore, by building a control valve fault reproduction test bench and injecting faults into the test control valve using a reasonable reproduction method, and by using the step signal test module in the ValveLink software that comes with the Fisher positioner, fault datasets simulating various actual working conditions were collected.
[0053] The step signal test module simulates the control valve movement process under different fault modes, generating a simulation dataset containing multiple fault types. Initial parameters include time (seconds), stroke (%), input (%), drive (%), and supply pressure (psi), as well as secondary parameters such as limit point, gain, dead time, overload, error, and stroke time. Here, % refers to the valve opening degree. In this embodiment, the collected status categories include 10 fault categories (F1, F2, F3, F4, F6, F6, F7, F8, F9, F10) and one fault-free category. Only the initial parameters are used as data features. Specific information is shown in Table 1. The comprehensive simulation image of pressure reducing valve mismatch is shown in Figure 2(a). In this embodiment, four states of 0.15MPa, 0.2MPa, 0.25MPa, and 0.3MPa are selected. The simulation image of loose actuator bolts is shown in Figure 2(b). The comprehensive simulation image of gas pipeline damage is shown in Figure 2(c). The simulation image of gas pipeline flattening is shown in Figure 2(d). The simulation image of external gas path blockage of the positioner mechanical structure is shown in Figure 2(e). The simulation image of actuator diaphragm wear is shown in Figure 2(f).
[0054] Table 1
[0055]
[0056]
[0057]
[0058] In one embodiment, the step response sample acquisition step is as follows:
[0059] The original time series data is format-converted, cleaned, and time-aligned to generate preprocessed data.
[0060] The preprocessed data is segmented into independent step response samples based on the step control signal.
[0061] Furthermore, time alignment involves: performing precise time synchronization processing on all data channels based on the timestamps recorded during data acquisition to ensure that all data points have a consistent time reference; and conducting multiple repeated tests for each combination of (fault mode + operating condition + step parameter) during testing.
[0062] The sample segmentation is as follows: using the step trigger signal in the acquired data stream as the segmentation point, the continuous time-aligned data is segmented into independent step response samples; each step response sample contains: a pre-trigger window of data of a predetermined length before the step trigger moment, which is used to provide a steady-state reference before the step occurs; and response window data sufficient to capture the complete dynamic response process after the step trigger moment.
[0063] Furthermore, the step amplitude of the step control signal is multiple different values with absolute values within the travel range of the preset interval; the step direction includes the valve opening direction and the valve closing direction; before the step, it is ensured that the valve reaches a steady state in the initial position; the test duration ensures that the dynamic response is fully recorded until a new steady state is reached.
[0064] In one embodiment, training and test datasets are constructed based on a simulated dataset, according to the input characteristics of the SCLIFD incremental learning model. The training dataset consists of 50% step signal test data simulating fault modes. Here, 5 or 10 types of fault data are placed sequentially in the same table. Since this model accepts .npy files, the table needs to be appropriately converted, while ensuring that the fault categories remain unchanged. The test dataset also contains step signal test data, in the same format as the training dataset, but different from the training data.
[0065] In one embodiment, the initial SCLIFD incremental learning model includes:
[0066] The feature extraction module uses a ResNet18 network to extract deep features from the input data;
[0067] The supervised comparative knowledge distillation module is used to guide the student model to learn discriminative features based on deep features through the teacher model and output feature samples.
[0068] The priority example selection module is used to calculate and scale the sample dissimilarity based on the feature samples and cached samples, and update the cached samples in the memory cache.
[0069] The classification module uses a balanced random forest classifier to output fault classifications.
[0070] Furthermore, the SCLIFD incremental learning model employs ResNet18 as the feature extractor. Its residual structure alleviates the challenges of training deep networks, accelerating training and improving performance. The overall structure is simple and easy to deploy. Supervised Comparative Knowledge Distillation (SCKD) addresses the challenges of learning discriminative features and mitigating catastrophic forgetting. A novel sample replay method, Method Priority Example Selection (MES), is used to mitigate catastrophic forgetting by storing the most representative samples in a memory cache. During the selection of the most representative samples, the sample dissimilarity is appropriately scaled to reduce computation time and improve accuracy. Finally, a balanced random forest (RF) classifier is used for classification due to its high performance in balancing class distribution and reducing classification bias during overall learning.
[0071] The model training process is as follows Figure 3 As shown, specifically: the fault data under the corresponding path is input into the SCLIFD incremental learning model and divided into 5 and 10 fault categories; the memory buffers for the 5 and 10 fault categories are set to 300 and 400 respectively, BATCH_SIZE to 32, RANDOM_SEED to 66, NUM_EPOCHS to 100, LR to 0.01, WEIGHT_DECAY to 1e-5, and MOMENTUM to 0.9; the Adam optimizer is used for model optimization; the training steps are as follows:
[0072] The newly acquired data and the selected priority examples are combined to form the current training set;
[0073] Input the current training set into two networks simultaneously: the frozen old feature network (trained after the previous incremental session) and the current feature network to be updated (initialized with the old network parameters). Extract the old features (output of the old network) and the new features (output of the current network) respectively.
[0074] After normalizing the extracted old and new features, the distillation loss is calculated. Specifically, the distillation loss is calculated by inputting the normalized old and new features into the classification head to generate corresponding logits. The difference in distribution between the current network's logits and the old network's logits is measured using KL divergence or cross-entropy to obtain the distillation loss. The calculation formula is as follows:
[0075]
[0076] In the formula, N represents the total number of features; I is the index set; A(i) is the set of all available samples related to a given feature i; The distribution loss for feature (i, a); P(z) i ;z a ) T The probability distribution between sample i and sample a calculated from the old network; P(z) i ;z a ) S The probability distribution between sample i and sample a is calculated for the new network.
[0077] The SCL loss is calculated based on the new features output by the current feature network. The calculation formula is as follows:
[0078]
[0079] In the formula, P(i) is the set of positive samples associated with a given feature i; A(i) is the set of all available samples associated with a given feature i; z i z p z a are the feature vectors of samples i, p, and a, respectively; τ is the temperature parameter.
[0080] Joint optimization and network update are performed based on distillation loss and SCL loss, and the parameters of the current feature network are updated using backpropagation with the total loss; the expression for the total loss function is:
[0081] L = L scl +λL dis ;
[0082] In the formula, λ is the balancing weight.
[0083] After training is complete, the current feature network is frozen and used as the old feature network for the next incremental session.
[0084] It also includes prioritizing the example set update, specifically: selecting representative samples from the new class data and adding them to the MES according to a preset strategy (such as based on sample importance or class balance); it is necessary to ensure that the updated MES covers all seen categories (old class + new class) to provide data support for subsequent incremental sessions.
[0085] Fault diagnosis steps:
[0086] Real-time collected data is input into the latest feature network, which outputs latent features. The latent features are then input into a pre-trained RF classifier, which outputs a fault category prediction. The RF classifier needs to be retrained (or fine-tuned) with the updated features after each incremental session to be compatible with the new category. When a sample with a confidence level below a threshold (unknown fault mode) appears in the real-time collected data, a new class label is triggered. After accumulating enough new class samples, a new round of training is started (returning to the training step).
[0087] In one embodiment, specific data is used for illustration:
[0088] Step 1: First, prepare the data:
[0089] Nineteen fault datasets were collected using the step signal test module in ValveLink software included with the Fisher locator, with 1050 samples collected for each category.
[0090] The collected data were filtered according to category conditions. The filtered samples were visualized and subjected to correlation analysis. In addition, based on the actual operating environment in the factory, common failure modes were summarized, and a total of 10 common and no-fault faults were screened out. Based on the parameter types that the intelligent positioner can monitor, its data characteristics include time, stroke, input, drive, and air supply pressure. Partial data of the simulated dataset is shown in Table 2.
[0091] Table 2 Partial Simulation Dataset
[0092] Time (seconds) journey(%) enter(%) drive(%) Gas supply pressure (psi) 0 44.4 0 70.5 183.97 0.4 48.91 0 38.7 184.04 0.8 33.11 0 55.52 184.04 1.2 18.12 0 72.69 184.04 1.6 9.42 0 72.53 184.04 2 4.49 0 67.94 184.04 2.4 2.14 0 67.5 184.04 2.8 1.02 0 68.09 184.04 3.2 0.92 0 66.67 184.04 3.6 0.78 0 66.77 184.04 4 0.78 0 68.38 184.04 4.4 0.78 0 67.7 184.04 4.8 0.78 0 67.55 184.04 5.2 0.78 0 66.38 184.04 5.6 0.73 0 67.94 184.04 6 0.73 0 67.7 184.04 6.4 0.73 0 67.11 184.04 6.8 0.68 0 65.64 184.04 7.2 0.68 0 67.16 184.04
[0093] Step 2: Construct the dataset:
[0094] The dataset is constructed according to the input format of the SCLIFD incremental learning model and divided into training and testing datasets. To meet the requirements of the incremental learning model, concise and high-quality training data is needed. The step signal data in the fault dataset is diluted and divided into two parts, resulting in a training set of 525 and a testing set of 525 for each class. The total training set for 5 fault classes is 3150, and the total testing set is 3150; the total training set for 10 fault classes is 5775, and the total testing set is 5775. The fault-free scenario is treated as the original data (class 0), and the remaining fault modes are divided into classes 1 to 10.
[0095] Step 3: Constructing the SCLIFD incremental learning model:
[0096] A high-end control valve fault diagnosis model was constructed using the SCLIFD incremental learning model, with preprocessed data loaded into the model. During each fault mode classification, a portion of the data was reserved in a memory cache, and the cached data was merged with the new data for feature extraction. Supervised contrastive learning loss and the Adam optimizer were used during training. The model was trained iteratively until it achieved high accuracy on the dataset.
[0097] Step 4: Model Evaluation: Evaluate the model using the test dataset, calculating various performance metrics such as accuracy, training time, testing time, and confusion matrix, as follows: Figure 4(a) and 4(b) As shown, among the four classifiers, the best prediction accuracy for 5 types of faults is 93.4%, and the best prediction accuracy for 10 types of faults is 96.5%. The model will be further adjusted and optimized based on the evaluation results.
[0098] The trained SCLIFD incremental learning model is deployed into the actual automatic fault diagnosis system, adapting to data changes and trends in real time. Based on real-time monitoring data, it accurately diagnoses faults in high-end control valves. The SCLIFD incremental learning model can precisely identify various fault categories in control valves. Combined with historical maintenance experience and guidance manuals, the system outputs relevant fault handling strategies and suggestions. This is particularly valuable for newly hired maintenance personnel, providing significant learning opportunities and enhancing their control valve maintenance capabilities, ensuring the safe and stable operation of control valves.
[0099] This embodiment also discloses a fault diagnosis system based on an incremental learning model, including:
[0100] The dataset construction module is used to construct a simulated dataset based on step signal test data of multiple fault modes;
[0101] The model building module is used to build an initial SCLIFD incremental learning model. The initial SCLIFD incremental learning model is trained using a simulated dataset to generate a new SCLIFD incremental learning model. The initial SCLIFD incremental learning model includes a feature extraction module, a supervised contrastive knowledge distillation module, a priority example selection module, and a classification module.
[0102] The fault diagnosis module is used to input the data to be diagnosed into the SCLIFD incremental learning model to obtain fault diagnosis results.
[0103] This embodiment also discloses a storage medium storing executable instructions, which implements a fault diagnosis method based on an incremental learning model when the processor executes the instructions.
[0104] 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 apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0105] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A fault diagnosis method based on an incremental learning model, characterized in that, The specific steps are as follows: A simulation dataset was constructed based on step signal test data of multiple fault modes; An initial SCLIFD incremental learning model is constructed, and the initial SCLIFD incremental learning model is trained using the simulated dataset to generate an SCLIFD incremental learning model; the initial SCLIFD incremental learning model includes a feature extraction module, a supervised contrastive knowledge distillation module, a priority example selection module, and a classification module; Acquire the data to be diagnosed, input the data to be diagnosed into the SCLIFD incremental learning model, and obtain the fault diagnosis result.
2. The fault diagnosis method based on an incremental learning model according to claim 1, characterized in that, The steps for obtaining the simulated dataset are as follows: A control valve failure reproduction test bench was built, and multiple target failure modes were defined. For each of the target failure modes, the failure was reproduced on the control valve failure reproduction test bench using a controllable physical failure injection method, and the type and degree of the injected failure were recorded. Under various operating parameters of the target fault modes, step control signals of different amplitudes and directions are applied to the target control valve, and valve stroke signals, drive data and air supply pressure signals are collected at the same time to form raw time series data. The original time series data is preprocessed and segmented to obtain independent step response samples, and each step response sample is labeled with metadata tags. The simulated dataset used to train and evaluate the SCLIFD incremental learning model is constructed based on the step response samples and their corresponding metadata tags.
3. The fault diagnosis method based on an incremental learning model according to claim 2, characterized in that, The test bench includes a target control valve, an intelligent valve positioner, an actuator, a gas supply system, a data acquisition system, and a control system.
4. The fault diagnosis method based on an incremental learning model according to claim 2, characterized in that, The target failure modes include pressure reducing valve mismatch, loose bolts, pipeline damage, pipeline flattening, gas path blockage, and diaphragm wear.
5. The fault diagnosis method based on an incremental learning model according to claim 2, characterized in that, The steps for obtaining the step response sample are as follows: The original time series data is format-converted, cleaned, and time-aligned to generate preprocessed data; The preprocessed data is segmented into independent step response samples based on the step control signal.
6. The fault diagnosis method based on an incremental learning model according to claim 1, characterized in that, The initial SCLIFD incremental learning model includes: The feature extraction module uses a ResNet18 network to extract deep features from the input data; The supervised comparative knowledge distillation module is used to guide the student model to learn discriminative features based on the deep features through the teacher model and output feature samples. The priority example selection module is used to calculate and scale the sample difference based on the feature samples and cached samples, and update the cached samples in the memory cache. The classification module uses a balanced random forest classifier to output fault classifications.
7. A fault diagnosis system based on an incremental learning model, characterized in that, include: The dataset construction module is used to construct a simulated dataset based on step signal test data of multiple fault modes; The model building module is used to build an initial SCLIFD incremental learning model, and train the initial SCLIFD incremental learning model using the simulated dataset to generate a new SCLIFD incremental learning model. The initial SCLIFD incremental learning model includes a feature extraction module, a supervised contrastive knowledge distillation module, a priority example selection module, and a classification module. The fault diagnosis module is used to input the data to be diagnosed into the SCLIFD incremental learning model to obtain the fault diagnosis results.
8. A storage medium, characterized in that, It stores executable instructions that, when executed by a processor, implement a fault diagnosis method based on an incremental learning model as described in any one of claims 1-6.