An equipment fault diagnosis method for adaptively updating model weights

Abstract

The application discloses an equipment fault diagnosis method for self-adaptive model weight updating. The application comprises the following steps: firstly, collecting historical data of equipment under multiple working conditions, and obtaining a plurality of auxiliary data sets and test data sets after data preprocessing, and constructing an auxiliary classification model by using the classification accuracy and classification energy of the auxiliary data sets; determining the prediction classification accuracy of the fault diagnosis model based on the classification energy of the test data set, and deciding whether to update the model weight according to the predicted classification accuracy, so as to realize self-adaptive model weight updating, thereby guaranteeing high-accuracy diagnosis of equipment faults. The application can reduce unnecessary model updating, run the domain adaptation method only when necessary, accelerate the fault diagnosis process, and thus can timely and effectively check and maintain the equipment in a fault state.

Description

Technical Field

[0001] This invention belongs to the field of rotating equipment fault diagnosis, and involves a fault diagnosis method for equipment in the fields of machine learning, deep learning and time series classification. Specifically, it involves a fault diagnosis method for equipment that adaptively updates model weights. Background Technology

[0002] Rotating equipment is widely used in modern industry, and its complexity and precision are constantly increasing, such as in aircraft engines and gas turbines. However, when these rotating devices experience failures such as bearing wear or blade breakage, it can lead to serious accidents and huge economic losses. Therefore, timely and accurate identification of equipment operating status and taking measures when signs of failure appear are crucial for improving production efficiency and reducing catastrophic losses.

[0003] Currently, a large number of research and patents on fault diagnosis have emerged, with data-driven methods being the most popular. These methods reveal the relationship between vibration signals and equipment status through the analysis of large-scale data. For example, patent CN113720605B uses a one-dimensional convolutional neural network and a fully connected neural network to extract vibration features and utilizes ensemble learning techniques to obtain diagnostic results from multiple classifiers. Patent CN112199888B uses a deep residual network to extract fault features from vibration signals.

[0004] While existing methods have achieved considerably high diagnostic accuracy, they require the features of the data used to train the model to follow the same distribution as the features of the data the model actually encounters. This differs from reality, as data features change when equipment conditions alter. When data features are inconsistent, model performance suffers a significant decline. Therefore, it is necessary to frequently check the model's accuracy in classifying new data to ensure that the model's diagnostic results remain acceptable. This involves human annotation of new data and comparison with the model's classification results.

[0005] Existing fault diagnosis methods mainly employ periodic update methods, such as running domain adaptation methods (e.g., Domain Adversarial Neural Networks, DANNs) every few days to update the weights of the diagnostic model to adapt to the latest data distribution. However, due to the irregularity of changes in data feature distribution, periodic update methods result in unnecessary or untimely model updates, wasting computational resources and potentially leading to erroneous model diagnostic results. Summary of the Invention

[0006] To address the problems and needs in the background technology, this invention proposes an equipment fault diagnosis method that adaptively updates model weights.

[0007] The technical solution of the present invention is as follows:

[0008] I. An Adaptive Method for Updating Model Weights in Equipment Fault Diagnosis

[0009] S1: Collect historical data of the equipment under multiple operating conditions and perform data preprocessing to obtain several auxiliary datasets and test datasets; among them, the auxiliary datasets are a collection of multi-label and multi-operating-condition data.

[0010] S2: Input all auxiliary datasets into the fault diagnosis model to determine the classification accuracy of each auxiliary dataset and calculate the classification energy of all auxiliary datasets.

[0011] S3: Construct an auxiliary classification model using the classification accuracy and classification energy of all auxiliary datasets;

[0012] S4: Calculate the classification energy corresponding to all test datasets, and then use the auxiliary classification model to calculate the predicted classification accuracy corresponding to all test datasets.

[0013] S5: Determine whether the accuracy of the current fault diagnosis model meets the equipment accuracy requirements based on the prediction classification accuracy of the test dataset. If not, update the weights of the fault diagnosis model and repeat S2-S4 until the accuracy of the updated fault diagnosis model meets the equipment accuracy requirements, and obtain the updated fault diagnosis model; otherwise, keep the model weights unchanged.

[0014] S6: Repeat S1-S5 at preset intervals to perform adaptive weight updates on the fault diagnosis model, thereby ensuring high accuracy in diagnosing equipment faults.

[0015] In S2, the formula for calculating the classification energy e of each auxiliary dataset is as follows:

[0016]

[0017] Where N represents the amount of data in each auxiliary dataset, x i x j Let f represent the i-th and j-th data points in the current auxiliary dataset, respectively. k (x i ) indicates that the model applies to data x i The output belonging to the k-th label, where T is the energy parameter and K represents the number of label types in the current auxiliary dataset.

[0018] Specifically, S3 is:

[0019] Treat the classification accuracy acc and classification energy e of each auxiliary dataset as a point, and then use the points of all auxiliary datasets to fit a linear function, which is denoted as the auxiliary classification model.

[0020] In step S4, the classification energy of each test dataset is calculated in the same way as the classification energy of the auxiliary dataset.

[0021] In step S5, if the prediction classification accuracy of multiple test datasets is less than the preset accuracy, then the accuracy of the current fault diagnosis model does not meet the equipment precision requirements; if the prediction classification accuracy of all test datasets is greater than the preset accuracy, then the accuracy of the current fault diagnosis model meets the equipment precision requirements.

[0022] II. A computer device

[0023] The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method.

[0024] III. A computer-readable storage medium

[0025] The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method.

[0026] IV. A computer program product

[0027] The product includes a computer program / instructions that, when executed by a processor, implement the steps of the method.

[0028] The beneficial effects of this invention are as follows:

[0029] This invention uses an auxiliary classification model to determine the classification accuracy of a fault diagnosis model. If the accuracy is less than a preset threshold, a domain-adaptive fault diagnosis method is used to adjust the model parameters; otherwise, the model weights remain unchanged. This invention reduces unnecessary model updates, running the domain-adaptive method only when necessary, thus accelerating the fault diagnosis process and enabling timely and effective troubleshooting and repair of faulty equipment. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the specific steps of the present invention. Detailed Implementation

[0031] The following uses data from the DIRG rolling bearing test bench at the Polytechnic University of Turin as a specific example to further illustrate the present invention. This dataset can be obtained via the FTP protocol.

[0032] DIRG is a high-speed aerospace bearing dataset. It provides information on the degree of bearing damage and its acceleration signals under different speeds and radial loads. Bearings are classified into three states: normal, inner ring fault, and outer ring fault. Inner and outer ring faults can be further subdivided into three levels: 150, 250, and 450 μm, resulting in a total of seven bearing states. Sensors were placed at two locations on the bearing to collect acceleration signals along the x, y, and z axes, generating a total of six channels of data. The dataset covers 17 operating conditions, corresponding to 17 domains. Within each domain, data collection was performed seven times, corresponding to the seven bearing states. This resulted in a total of 119 subsets. Data labels were assigned based on the bearing state. Each subset was obtained by sampling 10 seconds of data from the sensor at a frequency of 51200 Hz. In this embodiment, data from seven operating conditions in the dataset were used. Each of these seven conditions has a unique speed or load and was assigned an index as an identifier, as shown in Table 1.

[0033] Table 1 shows the seven working conditions in the dataset.

[0034]

[0035]

[0036] In this embodiment, ResNet is used as the fault diagnosis model, denoted as f, and dataset 1 (D1) is selected as the training set to train the model. The following data processing procedure was performed on D1: each data point was divided into multiple non-overlapping parts using a sliding window, each with a length of 1024 and the same label. These data were then standardized using min-max scaling. In practical applications, the model is not required to be ResNet.

[0037] like Figure 1 As shown, the equipment fault diagnosis method for adaptively updating model weights proposed in this invention includes the following steps:

[0038] S1: Data sets 2, 3, 4, 5, and 6 are considered auxiliary datasets, denoted as D2, D3, D4, D5, and D6 respectively. The same data preprocessing method as the model training set (D1) is used: each data point is split into multiple non-overlapping parts of length 1024 using a sliding window, with the same label. These data are then standardized using min-max scaling.

[0039] S2: Input all auxiliary datasets into the fault diagnosis model to obtain the classification results for the data, denoted as c2, c3, c4, c5, and c6 respectively. Compare these classification results with the actual labels, count the correct classification results on each auxiliary dataset, and divide them by the total data volume of that dataset to obtain the classification accuracy, denoted as acc2, acc3, acc4, acc5, and acc6 respectively. Treat datasets 7, 8, and 9 as test datasets, denoted as D7, D8, and D9. Input D7, D8, and D9 into the fault diagnosis model f to obtain the classification results for this dataset, denoted as c7, c8, and c9.

[0040] In this embodiment, the classification accuracy (acc) of the model on the auxiliary dataset are as follows:

[0041] acc2=97.32, acc3=95.41, acc4=92.66, acc5=90.35, acc6=91.20

[0042] The formula for calculating the classification energy e for each dataset is as follows:

[0043]

[0044] Where N represents the amount of data in each auxiliary dataset, and N is 50. i x j Let f represent the i-th and j-th data points in the current auxiliary dataset, respectively. k (x i ) represents the model for data x i The output belonging to the k-th label, T is the energy parameter, a hyperparameter, and T is set to 1. K represents the number of label types in the current auxiliary dataset. In this embodiment, the data has 7 states, so K is set to 7. The classification energies of the 5 auxiliary datasets are denoted as e2, e3, e4, e5, e6, and the classification energies of the 3 test datasets are denoted as e7, e8, e9. The specific values ​​are as follows:

[0045] e2=4.87, e3=5.56, e4=6.57, e5=7.41, e6=7.10,

[0046] e7=7.55, e8=7.11, e9=7.78

[0047] S3: Construct an auxiliary classification model using the classification accuracy and classification energy of all auxiliary datasets;

[0048] S3 specifically refers to:

[0049] Treating the classification accuracy (acc) and classification energy (e) of each auxiliary dataset as a point, and then fitting a linear function using the points from all auxiliary datasets, we obtain the auxiliary classification model, as shown in the following formula:

[0050] acc = -2.739229422066556e + 110.64828371278466

[0051] S4: Substitute the classification energy of the three test datasets into the auxiliary classification model to obtain the predicted classification accuracy. Substituting e7 = 7.45 into the above function, we obtain the classification accuracy acc7 = 89.97 predicted by the model on dataset D7. Substituting e8 = 7.11 into the above function, we obtain the classification accuracy acc8 = 91.17 predicted by the model on dataset D8. Substituting e9 = 7.7 into the above function, we obtain the classification accuracy acc9 = 89.56 predicted by the model on dataset D9.

[0052] S5: In the above results, acc7 and acc9 are less than the set threshold of 90. The weights of the fault diagnosis model are updated using the running domain adversarial neural network method. After the model is updated, steps S2 to S4 are executed again. The new classification accuracies are acc7 = 94.7, acc8 = 90.41, and acc9 = 93.56, all of which are greater than the set threshold. The update is then complete.

[0053] S6: Repeat S1-S5 at preset intervals to perform adaptive weight updates on the fault diagnosis model. This allows for flexible determination of the weight update time of the fault diagnosis model based on different working conditions or equipment status, making it easier to manage the fault diagnosis model and ensuring high accuracy in diagnosing equipment faults. It also enables timely and effective troubleshooting and repair of equipment with faulty conditions.

[0054] This invention uses an auxiliary classification model to determine the classification accuracy of a fault diagnosis model. If the accuracy is less than a preset threshold, a domain adaptation fault diagnosis method is used to adjust the model parameters; otherwise, the model weights remain unchanged. Therefore, this invention can reduce unnecessary model updates and only run the domain adaptation method when necessary, thus accelerating the fault diagnosis process.

[0055] The above embodiments are implementations of the present invention on the rolling bearing test bench dataset of the University of London. However, the specific implementation of the method of the present invention is not limited to bearings. Any similar scheme that collects equipment operation data through sensors and performs model evaluation in accordance with the principles and ideas of the present invention should be considered within the scope of protection of the present invention patent.

Claims

1. A method for equipment fault diagnosis that adaptively updates model weights, characterized in that, Includes the following steps: S1: Collect historical data of the equipment under multiple operating conditions and perform data preprocessing to obtain several auxiliary datasets and test datasets; S2: Input all auxiliary datasets into the fault diagnosis model to determine the classification accuracy of each auxiliary dataset and calculate the classification energy of all auxiliary datasets. Classification energy of each auxiliary dataset The calculation formula is as follows: in, This indicates the amount of data in each auxiliary dataset. , These represent the first and second elements in the current auxiliary dataset. Article, No. data, The model represents the data. Belongs to the The output of each label, For energy parameters, This indicates the number of label types in the current auxiliary dataset; S3: Construct an auxiliary classification model using the classification accuracy and classification energy of all auxiliary datasets; S4: Calculate the classification energy corresponding to all test datasets, and then use the auxiliary classification model to calculate the predicted classification accuracy corresponding to all test datasets. S5: Determine whether the accuracy of the current fault diagnosis model meets the equipment accuracy requirements based on the prediction classification accuracy of the test dataset. If not, update the weights of the fault diagnosis model to obtain an updated fault diagnosis model. S6: Repeat S1-S5 at preset intervals to perform adaptive weight updates on the fault diagnosis model, thereby ensuring high accuracy in diagnosing equipment faults.

2. The equipment fault diagnosis method for adaptively updating model weights according to claim 1, characterized in that, Specifically, S3 is: The classification accuracy of each auxiliary dataset and classification energy Treating a point as an example, we then fit a linear function using points from all auxiliary datasets and denote it as the auxiliary classification model.

3. The equipment fault diagnosis method for adaptively updating model weights according to claim 1, characterized in that, In step S4, the classification energy of each test dataset is calculated in the same way as the classification energy of the auxiliary dataset.

4. The equipment fault diagnosis method for adaptively updating model weights according to claim 1, characterized in that, In step S5, if the prediction classification accuracy of multiple test datasets is less than the preset accuracy, then the accuracy of the current fault diagnosis model does not meet the equipment precision requirements; if the prediction classification accuracy of all test datasets is greater than the preset accuracy, then the accuracy of the current fault diagnosis model meets the equipment precision requirements.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1 to 4.

Images