An electromechanical system edge end fault diagnosis method based on improved fuzzy width learning
Patent Information
- Application Number
- CN202410669290.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-28
- Publication Date
- 2026-10-27
- Estimated Expiration
- 2044-05-28
AI Technical Summary
[0080]本发明的有益效果是：该改进模糊宽度学习的边缘端机电设备故障诊断方法，在边缘端实现故障诊断模型推理诊断以及模型在线训练更新。针对模型故障诊断精度不高的问题，使用模糊宽度学习故障诊断方法，提升诊断正确率；针对模型超参数选取复杂的问题，使用减法聚类改进的模糊C均值聚类实现对模糊宽度学习的改进，在保留了传统BLS结构简单、训练速度快优点的同时，有效减少了超参数讨论所需时间成本；提出改进模糊宽度学习增强节点的增量学习方法，进一步提升模型准确率，且实现模型增量更新；提出STM32模型部署方法，在边缘端实现改进模糊宽度学习的推理运算，能够进一步提高监测系统的实时性和故障响应速度，具有较高的实际意义和工程应用价值。
Smart Images
Figure CN118760918B_ABST
Abstract
Description
[0001] Field of study
[0002] This invention relates to the field of fault diagnosis, and more specifically to the application field of fault diagnosis of edge electromechanical equipment. Background Technology
[0003] Fault diagnosis is a crucial link in ensuring the safe and reliable operation of industrial processes. Timely, accurate, and efficient identification of abnormal faults is essential for the efficient and safe operation of industrial processes. Implementing fault diagnosis for electromechanical equipment at the edge can reduce the bandwidth pressure on the Internet of Things (IoT) and the computing load on the cloud, enabling timely fault detection, improving the real-time performance of fault diagnosis, and preventing significant losses. However, traditional data-driven fault diagnosis methods suffer from problems such as requiring large amounts of historical data and high computing power. The limited computing resources at the edge pose a challenge to their implementation. Under such circumstances, how to achieve model porting, deployment, and online training and updates at the edge has significant engineering significance and application value.
[0004] A literature review reveals that research on deploying machine learning methods at the edge is still in its early stages. Limited by the computational resources available at the edge, current edge machine learning deployment methods primarily focus on embedded microprocessors (MPUs) such as Cortex-A9 and Raspberry Pi. Furthermore, these methods are limited to deploying the relatively low-computation inference stage of intelligent fault diagnosis on embedded MPUs, failing to implement the training process at the edge. For example, Shang Shuyang et al., in their paper "Design of an Embedded Bearing Health Monitoring System Based on ARM," implemented bearing health monitoring using the ARM9 core S3C2440A main control chip, Linux operating system, and Qt graphical development software. However, this research still involved machine learning training on a PC, lacking real-time fault diagnosis and incremental updates to the fault diagnosis model, making it difficult to apply in practical engineering.
[0005] Embedded microcontroller units (MCUs) offer smaller size and lower power consumption compared to MPUs, making them commonly used in sensor data acquisition scenarios. Using MCUs as edge devices and deploying fault diagnosis algorithms closer to the device can further improve the real-time performance of health monitoring and the speed of fault diagnosis. However, the limited computing resources of MCUs compared to MPUs restrict the porting and deployment of fault diagnosis algorithms. Model porting and deployment are more challenging, and current research on real-time fault diagnosis based on MCUs is lacking. Proposed porting methods primarily rely on tools provided by MCU manufacturers. For example, Crocioni et al., in "Li-ion batteries parameter estimation with tiny neural networks embedded on intelligent IoT microcontrollers," built a lithium-ion battery health monitoring method based on Artificial Neural Networks (ANNs) and deployed a pre-trained ANN model onto an STM32 microcontroller using the STM32Cube.AI tool. However, this porting and deployment is limited by the tools used, making it difficult to adapt to newly proposed machine learning algorithms. Furthermore, since it involves porting pre-trained models, incremental updates at the edge are not possible.
[0006] In 2018, Chen et al. proposed a highly efficient incremental learning system without the need for deep architecture—Broad Learning System (BLS)—in their paper "Broad learning system: An effective and efficient incremental learning system without the need for deep architecture." This system aims to address problems in deep learning and edge computing, offering advantages such as simple structure, fast training speed, and high accuracy, while also providing the benefits of incremental learning and offering a new method for edge fault diagnosis. In 2021, Ren Chang'e et al. summarized the existing problems and research hotspots of Broad Learning System in their paper "Research Progress." Hyperparameter setting methods and cloud and embedded device deployment applications have become research hotspots.
[0007] In summary, deploying intelligent fault diagnosis algorithms at the edge has significant industrial application value. Optimizing these algorithms and implementing online updates are crucial research areas. Therefore, the invention of a fault diagnosis method for electromechanical systems at the edge is urgently needed to address these existing problems. Summary of the Invention
[0008] The purpose of this invention is to address the shortcomings of existing technologies by proposing an edge-end fault diagnosis method for electromechanical systems based on improved fuzzy width learning. This method leverages the limited computing resources at the edge to enable fault diagnosis models to perform inference and diagnosis, as well as online model training and updates. To address the issue of low model fault diagnosis accuracy, a Takagi-Sugeno fuzzy model is adopted at the mapping node layer. Expert experience is incorporated into the width learning process, and incremental learning of enhanced nodes is implemented at the edge to improve diagnostic accuracy and enable online model training and updates. To address the complexity of model hyperparameter selection, fuzzy C-means clustering, improved using subtractive clustering, is employed to enhance fuzzy width learning, enabling automatic hyperparameter selection. For the incremental model issue, an enhanced node incremental method is proposed to achieve online model updates, further improving diagnostic accuracy. Finally, to address the difficulty of deploying machine learning models at the edge, an edge-end porting and deployment scheme is proposed based on the Matlab Coder toolbox, implementing width learning inference operations on an STM32 microcontroller.
[0009] The technical solution of this invention is as follows: (See attached document) Figure 1 A fault diagnosis method for the edge of an electromechanical system based on improved fuzzy width learning includes the following six steps: data processing, model building, model improvement, model increment, model porting, and fault diagnosis; the specific process is as follows:
[0010] Step 1: Data Processing
[0011] The collected fault diagnosis data is processed using the following sub-steps:
[0012] Sub-step 1: Data cleaning. Due to environmental interference, sensor malfunctions, and other factors, the data in the dataset may contain default or outlier values. This invention preferably uses the nearest neighbor method to process these values, that is, replacing the problematic value with the average of the preceding and following data.
[0013] Sub-step 2: Data standardization. The dataset is processed using the Z-score standardization method to eliminate differences between units.
[0014] Sub-step 3: Data partitioning. Based on the dataset, determine the fault type, select feature parameters, and partition the dataset into training and test sets.
[0015] Step 2: Model Building
[0016] Sub-step 1: Establish the mapping layer model
[0017] In the Takagi-Sugeno fuzzy system, let X = (x1, x2, ..., x...). n ), xs =(x s1 ,x s2 ,......,x sm ), s = 1, 2, ..., n, assuming the i-th fuzzy subsystem has K i If there are 10 rules, then there is the IF-Then rule, i.e., IF x s1 is andx s 2 is ...and x sm is Then k = 1, 2, ..., K i ,So
[0018]
[0019] in, Let f be a coefficient, and f be a function; then, the training sample x s The i-th fuzzy subsystem is represented as
[0020]
[0021] in, Here, Take 1, The output vector of the i-th fuzzy subsystem for all samples is obtained directly by the k-means algorithm.
[0022] Z i =[Z 1i Z 2i ,…Z ni (3)
[0023] Then the outputs of the n fuzzy subsystems, i.e. the mapping layer model, are:
[0024] Z n =[Z1,Z2,…,Z n (4)
[0025] Sub-step 2: Establish the enhancement layer model
[0026] The outputs of the fuzzy rules generated by each fuzzy subsystem are not immediately aggregated into a single value. Instead, they are all sent to the enhancement layer for further nonlinear transformation, resulting in the model of the j-th enhancement layer node.
[0027] H j =ξ j (Z n W hj +βhj (5)
[0028] Where ξ j W represents a random mapping. hj β hj Let represent random weights. Then the output of the m-th enhancement layer node, i.e., the enhancement layer model, is:
[0029] H m =[H1,H2,…,H m (6)
[0030] Sub-step 3: Defuzzification of the mapping layer model
[0031] To preserve the input characteristics, the defuzzified output of the fuzzy subsystem is...
[0032]
[0033] in, Let be the coefficient, then the defuzzified output of the i-th fuzzy subsystem is (8), and the mapping layer can be represented as (9).
[0034]
[0035] F n =(F1,F2,......,F n (9)
[0036] Sub-step 4: Establish the output layer model
[0037] The mapping layer model is combined with the enhancement layer model to obtain the final model.
[0038] Y = F n +H m W e (10)
[0039] make get
[0040]
[0041] Where Ω i Depend on composition,
[0042] Sub-step 5: Solving the weights of the output layer model
[0043] Using the pseudo-inverse solution technique of ridge regression, the weights are obtained as
[0044] W=(DΩ,H m ) + Y (12)
[0045] Among them, (DΩ,H m ) + =((DΩ,H m ) T (DΩ,H m )) -1 (DΩ,H m ) T ;
[0046] Step 3: Model Improvement
[0047] An improved FCM is used to replace K-means clustering, and FBLS is improved to achieve automatic selection of the hyperparameter of the number of cluster centers. Specifically, the optimal number of cluster centers is obtained through subtractive clustering to determine the number of cluster centers for FCM, and the cluster centers obtained by subtractive clustering are used as the initial cluster centers for FCM. The process includes the following sub-steps:
[0048] Sub-step 1: Subtractive clustering to determine the number of clusters
[0049] In subtractive clustering, each data point can potentially serve as a cluster center. Let data point x... j The probability of becoming a cluster center is P. j
[0050]
[0051] Where r a Let x represent the neighborhood radius of the data point, which is a positive constant. i Let i = 1, 2, ..., n be the data points. The probability of a data point becoming a cluster center is inversely proportional to its distance from other points. The point with the highest probability is chosen as the first cluster center, denoted as ∑_{i=1}^{n} ... The probability is Data point x j The probability of becoming a cluster center has been updated to
[0052]
[0053] Where r b It is also a positive constant, which actually defines the neighborhood radius where the probability of a data point becoming a cluster center is significantly reduced. The point with the highest probability within this neighborhood is selected as the second cluster center. Continue refining the probability of each data point becoming a cluster center, repeating k times to obtain k cluster centers. The criterion for stopping subtractive clustering is defined as: ε r and ε a , representing the upper and lower bounds for being accepted as cluster centers, when hour
[0054]
[0055] Where d min The parameter is λ, where λ≥1 represents the cluster center; the cluster centers obtained by subtractive clustering are... As the initial w i Here, a tiny amount must be added to the cluster centers obtained by subtractive clustering because directly substituting the cluster centers obtained by subtractive clustering into the original dataset would result in division by zero in the following text.
[0056] Sub-step 2: Improve the clustering method to determine cluster centers
[0057] The objective function of fuzzy C-means clustering is
[0058]
[0059] Where, x j (j = 1, 2, ..., N) represents the required sample space data, μ ji x represents j For w i The membership degree, and satisfying m represents the fuzziness index, which is taken as m = 2 in this paper, w i (i = 1, 2, ..., c) are the cluster centers. In fuzzy C-means clustering, w i The initial value is generated through random initialization.
[0060] The clustering objective of improved fuzzy C-means clustering is to minimize J. m First, calculate μ ji ;
[0061]
[0062] According to μ ji Correct the cluster center w i ;
[0063]
[0064] Calculation error
[0065]
[0066] Set the tolerance error E max , when e<E max The algorithm ends if the condition is met; otherwise, the above process is repeated from formula (17).
[0067] Step 4: Model Increment
[0068] If improving the fuzzy width learning training fails to achieve the desired performance, new augmentation nodes are added to reduce the loss function. Let p be the number of augmentation nodes added, and A... m =(DΩ,H m ) and W m For the enhancement layer and weights before adding enhancement nodes, A m+1 =[A m |ξ(Z n W hm+1 +β hm+1 )] and W m+1 To add enhancement layers and weights after enhancing nodes, W hm+1 To determine the weights of the p newly added augmenting nodes, we have:
[0069]
[0070] in When c = 0, b T =(1+d) T d) -1 d T A + When c≠0, b T =(c) + Therefore, the updated weights are:
[0071]
[0072] Step 5: Model Porting
[0073] This invention proposes an improved deployment scheme for porting a fuzzy width learning model to the edge, and implements intelligent fault diagnosis based on fuzzy width learning inference operation on an STM32 microcontroller. The steps required for porting are as follows.
[0074] Sub-step 1: Function encapsulation. Since the model's inference process is executed at the edge, the computational process required for model inference is encapsulated into a function. The inputs are the training dataset, model weights, and parameters, and the output is the fault diagnosis result.
[0075] Sub-step 2: Model Generation. Use the MATLAB Coder toolbox to generate a library from the encapsulated functions. It is recommended to generate a mex file, execute the mex file to verify the translation results, and then generate the library.
[0076] Sub-step 3: Program porting. Use Keil uVision5 to create a project file, port the lib library into the project file and declare it, and copy the Matlab header functions into the project file and declare them.
[0077] Sub-step 4: Program writing. Complete the header file declarations, initialization, and main program configuration in the project file main.c. Add other functions or project modules as needed. Compile and program the program, and then verify it on the STM32.
[0078] Step 6: Fault Diagnosis
[0079] Test data is sent to the STM32 via serial port, and the improved fuzzy width learning fault diagnosis method is run on the STM32 to perform fault diagnosis.
[0080] The beneficial effects of this invention are as follows: This improved fuzzy width learning-based fault diagnosis method for edge-end electromechanical equipment enables fault diagnosis model inference and online model training and updating at the edge. Addressing the issue of low model fault diagnosis accuracy, a fuzzy width learning-based fault diagnosis method is used to improve diagnostic accuracy. To address the complexity of model hyperparameter selection, a subtractive clustering-improved fuzzy C-means clustering method is used to improve fuzzy width learning, retaining the advantages of traditional BLS structure simplicity and fast training speed while effectively reducing the time cost required for hyperparameter discussion. An improved incremental learning method for enhancing nodes in fuzzy width learning is proposed to further improve model accuracy and achieve incremental model updates. An STM32 model deployment method is proposed to implement improved fuzzy width learning inference operations at the edge, further improving the real-time performance and fault response speed of the monitoring system, demonstrating high practical significance and engineering application value. Attached Figure Description
[0081] Figure 1 This is a schematic diagram of the implementation framework of the present invention.
[0082] Figure 2 This is a schematic diagram of improved fuzzy width learning.
[0083] Figure 3 This is a schematic diagram of the electromechanical actuator dataset in the embodiment.
[0084] Figure 4 This is a graph showing the impact of the improved fuzzy width learning hyperparameter selection on the fault diagnosis accuracy in the embodiment.
[0085] Figure 5 This is a graph showing the relationship between the accuracy of the improved fuzzy width learning fault diagnosis and the distance parameter in the embodiment.
[0086] Figure 6 This is a physical diagram of the improved fuzzy width learning fault diagnosis in the embodiment. Detailed implementation method:
[0087] The effectiveness of this invention is demonstrated and verified using NASA's Flyable Electromechanical Actuator (FLEA) dataset. NASA developed an experimental platform for Flyable Electromechanical Actuators (FLEAs) to reproduce FLEA failures and monitor data. This platform consists of two test FLEAs and one load FLEA. One test FLEA operates normally, while the other simulates the failure process. The load FLEA simulates the dynamic load conditions during actual aircraft flight. The FLEA experimental platform measures acceleration, temperature, current, and voltage using deployed sensors. The dataset includes 17 dimensions, such as desired location, A test actuator location, B test actuator location, load actuator location, A test actuator current, B test actuator current, load actuator current, A test actuator voltage, and B test actuator voltage. Figure 3 As shown.
[0088] All experiments in this embodiment were tested using the MATLAB R2022a software platform, with the computer equipped with an Intel(R) Xeon(R) W-2133 CPU @ 3.60GHz and 64GB of memory. The specific process of the edge fault diagnosis method for electromechanical systems based on improved fuzzy width learning in this embodiment is as follows:
[0089] Step 1: Data Processing. The process is the same as described in the invention. This case selects six features for subsequent experiments: Y-axis acceleration of test actuator A, X-axis acceleration of test actuator A, Z-axis acceleration of test actuator A, temperature of test actuator A, X-axis acceleration of test actuator B, and Y-axis acceleration of test actuator B. Four scenarios are considered: normal operation (Include Normal), ball screw jam fault (Jam), ball screw spalling fault (Spall), and motor winding short circuit fault (Motor). The selected experimental dataset contains 600 sets of original data, divided into four scenarios: normal operation, ball screw jam fault, ball screw spalling fault, and motor winding short circuit fault. Each set of data uses the six features mentioned above. During the experiment, the dataset is divided into training and test sets at a ratio of 80% and 20%, respectively.
[0090] Step 2: Model Building. The process is the same as described in the invention. The experimental results are analyzed, as shown in Table 1. The FBLS and BLS network structures are unified as 3-6-20, i.e., Nf is 3, Nm is 6, and Ne is 20. The SVM is determined using the libsvm Matlab toolbox. On the NASA electromechanical actuator dataset, Fuzzy Broad Learning System (FBLS) achieves a higher fault diagnosis accuracy compared to Broad Learning System (BLS) and Support Vector Machines (SVM).
[0091] Table 1 Comparison of fault diagnosis accuracy between fuzzy width learning and other methods
[0092]
[0093] Step 3: Model Improvement. The process is the same as described in the invention. Analyze the experimental results.
[0094] Sub-step 1: Verify the feasibility of the improved fuzzy width learning. The different clustering methods used in FBLS are discussed, as shown in Table 2. FBLS uses the k-means clustering method, FCM-BLS is FBLS using FCM clustering, and IFBLS is the improved method proposed in this invention. For ease of comparison, all three methods are trained with the same hyperparameters: Nf = 3, Nm = 6, and Ne = 20.
[0095] Table 2. Impact of different clustering methods on the accuracy of fuzzy set width learning fault diagnosis.
[0096]
[0097] The table shows that IFBLS has the highest accuracy in fault diagnosis. Although the training time is longer compared to other methods, it directly determines a reasonable network structure without requiring a grid search for hyperparameters, resulting in lower time costs in practice. Furthermore, the testing time is almost identical for all three methods during inference.
[0098] Sub-step 2: Verify the effectiveness of the improved fuzzy width learning. The improved model IFBLS is compared with FBLS, BLS, and SVM, as shown in Table 3. For ease of comparison, the network structures of IFBLS, FBLS, and BLS are unified as 3-6-20, that is, Nf is 3, Nm is 6, and Ne is 20. SVM is determined using the libsvm Matlab toolbox.
[0099] Table 3 Comparison of fault diagnosis accuracy between improved fuzzy width learning and other methods
[0100]
[0101] The following conclusions can be drawn from the table:
[0102] ① Both IFBLS and FBLS can achieve high fault diagnosis accuracy in a relatively short runtime. Among them, IFBLS has the highest diagnosis accuracy, especially on the test set, reaching 97.98%, which further demonstrates the advantages of the IFBLS method proposed in this invention.
[0103] ② In terms of time cost, IFBLS is similar to BLS. Especially compared with SVM, although IFBLS and FBLS have a slightly increased training time, it is still at a relatively low level. This proves that the IFBLS proposed in this invention retains the fast computation speed of BLS.
[0104] ③ Compared to width learning, IFBLS and FBLS have simpler network structures but higher fault diagnosis accuracy. This indicates that IFBLS requires only a small number of fuzzy rules to achieve higher fault diagnosis accuracy, thus alleviating the rule explosion problem in fuzzy systems.
[0105] Sub-step 3: Discuss the impact of hyperparameters on the performance of the model of this invention. First, we discuss the impact of the selection of hyperparameters Nf, Nm, and Ne on the accuracy of FBLS fault diagnosis, such as... Figure 4 As shown in the figure. The following conclusions can be drawn from the figure:
[0106] ① As the hyperparameters increase, the fault diagnosis accuracy gradually increases. It goes through three stages: rapid growth, fluctuating growth, and near-stabilization, achieving a relatively satisfactory diagnosis accuracy, while the training time also increases.
[0107] ② Changes in the mapping node layer have a more significant impact on the model's fault diagnosis accuracy. Specifically, as Nf and Nm increase, the diagnosis accuracy fluctuates significantly, while as Ne increases, the diagnosis accuracy tends to stabilize.
[0108] Further analysis of the hyperparameter selection yielded the following results: Figure 5 As shown in the figure. Traditional methods typically employ a grid search approach, which is time-consuming. For example, considering all three hyperparameters within the range [0, 10], a grid search would require 1000 iterations to determine the optimal number of cluster centers. Therefore, this invention proposes a more intuitive method for selecting Nf. By discussing the distance between cluster centers, the value of Nf, i.e., the optimal number of cluster centers, can be determined more quickly. This is one of the advantages of the improved model in this invention. The following conclusions can be drawn from the figure:
[0109] ① As the distance parameter increases, the number of cluster centers decreases, the time cost decreases, and the accuracy of fault diagnosis decreases. This is in line with expectations and also proves the rationality of the improved method proposed in this invention.
[0110] ② Considering all factors, a distance parameter value between [0.2 0.7] is more suitable. In the initial stage, as the parameter increases, the decrease in training time is more significant than the decrease in diagnostic accuracy. Therefore, a suitable parameter value can be selected by considering all factors.
[0111] Step 4: Model Increment. The process is the same as described in the invention. Analyze the experimental results to verify the incremental effect of the model enhancement nodes of the present invention. Set the initial network structure to 2-6-20, and dynamically increase the number of enhancement nodes by 20 each time, with the final structure being 2-6-30. The results are shown in Table 4.
[0112] Table 4. Improved Fuzzy Width Learning Enhances the Accuracy of Incremental Fault Diagnosis for Nodes
[0113]
[0114] The table shows that the model accuracy improves with the increase of augmentation nodes, proving the effectiveness of the incremental augmentation node method proposed in this invention. When the improved fuzzy width learning method proposed in this invention fails to achieve ideal performance after training, the model's diagnostic accuracy can be improved by adding new augmentation nodes.
[0115] Step 5: Model Porting. The process is the same as described in the invention. Using the porting method, IFBLS is ported to an STM32 microcontroller, and IFBLS inference operations are implemented on the STM32 microcontroller using the Anfulai STM32-V7 development board (MCU is STM32H750XBH6, Flash is 2MB).
[0116] Step 6: Fault Diagnosis. The process is the same as described in the invention. While performing fault diagnosis, the performance of the IFBLS edge is primarily verified, with the main indicators being diagnostic accuracy and diagnostic time.
[0117] Sub-step 1: Verify effectiveness. Set Nf to 3, Nm to 6, and Ne to 20. Nf is not directly set but obtained through the distance parameter. The results are shown in Table 5. The test time in the table is the time required for the device to complete the entire inference process, not the time required to execute one inference operation, and includes the time required for both data reception and inference diagnosis. If calculated based on the time required to execute one inference operation, a millisecond-level fault diagnosis response speed can be achieved, meeting the requirements for real-time fault diagnosis response rate in actual industrial applications. (The last sentence appears to be incomplete and possibly refers to a different product or service.) Figure 6 As shown.
[0118] Table 5 Comparison of STM32 fault diagnosis results based on IFBLS
[0119]
[0120] The table shows that the accuracy of IFBLS fault diagnosis at the edge is almost identical to that on the PC. However, due to limited computing power, the computation time is increased, which proves the feasibility of the deployment method described in this invention.
[0121] Sub-step 2: Verify stability. Repeat the experiment 5 times on both the PC and edge devices to verify the stability of the IFBLS fault diagnosis model and the stability of the proposed porting and deployment method. Set Nf to 3, Nm to 6, and Ne to 20; the results are shown in Table 6.
[0122] Table 6. Fault diagnosis accuracy of IFBLS under repeated experiments.
[0123]
[0124]
[0125] The table shows that the accuracy of IFBLS fault diagnosis remained relatively stable at 97.50% in repeated experiments, with 97.80% accuracy on the PC side and 97.60% accuracy on the edge side. This demonstrates that the improved fuzzy width learning method and its edge deployment method both possess strong stability.
[0126] Sub-step 3: Measure the power consumption required to perform fault diagnosis. The results are as follows: Figure 6 As shown in Table 7, Figure 6 This table shows the actual power consumption test results for IFBLS fault diagnosis. The lower left corner shows the power supply, which provides 12V power. The upper left corner shows the ammeter. The test circuit is measured in amperes (A). Table 7 shows the measurement results of the current and voltage required by the STM32 development board in the working and standby states.
[0127] Table 7 Power Consumption Test Results for Improved Fuzzy Width Learning Fault Diagnosis
[0128]
[0129] The following conclusions can be drawn from the chart:
[0130] ① Using the IFBLS fault diagnosis model, performing a fault diagnosis at the edge consumes approximately 0.048W of power (including peripherals such as the LCD screen). The operating current is 0.22917A, the standby current is 0.22515A, the difference is 0.00398A, and the voltage is constant at 12V. The calculated power is approximately 0.048W.
[0131] ② Comprehensive analysis shows that the program's fault diagnosis execution consumes relatively low levels of computing and storage resources, making it suitable for edge-end fault diagnosis. Analysis of the project file using Keil 5 reveals that the program occupies 24.3KB of Flash memory and 9.0KB of RAM, while the development board is equipped with only 2MB of Flash memory. The required storage resources are far less than the development board's rated memory, providing strong support for actual deployment and development.
[0132] The above experiments demonstrate that the improved fuzzy width learning method proposed in this invention performs excellently in the fault diagnosis problem of the NASA actuator dataset. It retains the learning speed of the BLS method, allowing for easy model training on a regular PC in a short time. The fault diagnosis accuracy is significantly improved compared to other methods. The improved fuzzy C-means clustering using subtractive clustering effectively reduces the time cost required for hyperparameter discussion, and the enhanced incremental learning method for nodes further improves the model's diagnostic accuracy. Furthermore, an IFBLS edge deployment scheme based on STM32 is proposed, maintaining high fault diagnosis performance at the edge while keeping power consumption low. This indicates that the proposed method has advantages such as fast computation speed, high diagnostic accuracy, and easy lightweight deployment.
Claims
1. A method for edge fault diagnosis of electromechanical systems based on improved fuzzy width learning, characterized in that, Includes the following steps: Step 1: Data Processing: The collected fault diagnosis data is processed using the following sub-steps: Sub-step 1: Data cleaning; Sub-step 2: Data standardization; Sub-step 3: Data partitioning: Based on the dataset, determine the fault type, select feature parameters, and partition the dataset into training and test sets; Step 2: Model building, including the following sub-steps: Sub-step 1: Establish the mapping layer model: In the Takagi-Sugeno fuzzy system, let X = (x1, x2, ..., x...). n ), x s =(x s1 ,x s2 ,......,x sm ), s = 1, 2, ..., n, assuming the i-th fuzzy subsystem has K i If there are 10 rules, then there is the IF-Then rule, i.e., IF x s1 is and x s2 is ...and x sm is Then k = 1, 2, ..., K i ,So in, Let f be a coefficient, and f be a function; then, the training sample x s The i-th fuzzy subsystem is represented as in, Here, Take 1, The output vector of the i-th fuzzy subsystem for all samples is obtained directly by the k-means algorithm. WITH i =[Z 1i ,WITH 2i ,…WITH ni ] (3) Then the outputs of the n fuzzy subsystems, i.e. the mapping layer model, are: WITH n =[Z1,Z2,…,Z n ] (4) Sub-step 2: Establish the enhancement layer model: The outputs of the fuzzy rules generated by each fuzzy subsystem are not immediately aggregated into a single value. Instead, they are all sent to the enhancement layer for further nonlinear transformation, resulting in the model of the j-th enhancement layer node. H j =ξ j (Z n W hj +b hj ) (5) Where ξ j W represents a random mapping. hj β hj Representing random weights; then the output of the m-th enhancement layer node, i.e., the enhancement layer model, is: H m =[H1,H2,…,H m ] (6) Sub-step 3: Defuzzification of the mapping layer model: To preserve the input characteristics, the defuzzified output of the fuzzy subsystem is... in, Let be the coefficient, then the defuzzified output of the i-th fuzzy subsystem is (8), and the mapping layer can be represented as (9). F n =(F1,F2,......,F n ) (9) Sub-step 4: Establish the output layer model: The mapping layer model is combined with the enhancement layer model to obtain the final model. Y=F n +H m W e (10) make get Where Ω i Depend on composition, Sub-step 5: Solving for the weights of the output layer model: Using the pseudo-inverse solution technique of ridge regression, the weights are obtained as W=(DΩ,H m ) + Y (12) Among them, (DΩ, H m ) + =((DΩ,H m ) T (DΩ,H m )) -1 (DΩ,H m ) T ; Step 3: Model Improvement An improved FCM is used to replace K-means clustering, and FBLS is improved to achieve automatic selection of the hyperparameter of the number of cluster centers. Specifically, the optimal number of cluster centers is obtained through subtractive clustering to determine the number of cluster centers for FCM, and the cluster centers obtained by subtractive clustering are used as the initial cluster centers for FCM. The process includes the following sub-steps: Sub-step 1: Subtractive clustering to determine the number of clusters: In subtractive clustering, each data point can potentially serve as a cluster center. Let data point x... j The probability of becoming a cluster center is P. j Where r a Let x represent the neighborhood radius of the data point, which is a positive constant. i Let i = 1, 2, ..., n be the data points. The probability of a data point becoming a cluster center is inversely proportional to its distance from other points. The point with the highest probability is chosen as the first cluster center, denoted as ∑_{i=1}^{n} ... The probability is Data point x j The probability of becoming a cluster center has been updated to Where r b It is also a positive constant, which actually defines the neighborhood radius where the probability of a data point becoming a cluster center is significantly reduced. The point with the highest probability within this neighborhood is selected as the second cluster center. Continue refining the probability of each data point becoming a cluster center, repeating k times to obtain k cluster centers. The criterion for stopping subtractive clustering is defined as: ε r and ε a , representing the upper and lower bounds for being accepted as cluster centers, when hour Where d min The parameter is λ, where λ≥1 represents the cluster center; the cluster centers obtained by subtractive clustering are... As the initial w i Here, a tiny amount must be added to the cluster centers obtained by subtractive clustering because directly substituting the cluster centers obtained by subtractive clustering into the original dataset will result in division by zero in the following text. Sub-step 2: Improve the clustering method to determine cluster centers: The objective function of fuzzy C-means clustering is Where, x j (j = 1, 2, ..., N) represents the required sample space data, μ ji x represents j For w i The membership degree, and satisfying m represents the fuzziness index, which is taken as m = 2 in this paper, w i (i = 1, 2, ..., c) are the cluster centers. In fuzzy C-means clustering, w i The initial value is generated through random initialization. The clustering objective of improved fuzzy C-means clustering is to minimize J. m First, calculate μ ji ; According to μ ji Correct the cluster center w i ; Calculation error Set the tolerance error E max , when e<E max The algorithm ends if the condition is met; otherwise, the above process is repeated from formula (17). Step 4: Model Increment: If improving the fuzzy width learning training fails to achieve the desired performance, new augmentation nodes are added to reduce the loss function. Let p be the number of augmentation nodes added, and A... m =(DΩ,H m ) and W m For the enhancement layer and weights before adding enhancement nodes, A m+1 =[A m |ξ(Z n W hm+1 +β hm+1 )] and W m+1 To add enhancement layers and weights after enhancing nodes, W hm+1 To determine the weights of the p newly added augmenting nodes, we have: in When c = 0, b T =(1+d) T d) -1 d T A + When c≠0, b T =(c) + Therefore, the updated weights are: Step 5: Model Porting: Intelligent fault diagnosis based on fuzzy width learning inference was implemented on the STM32 microcontroller. The sub-steps required for porting are as follows: Sub-step 1: Function encapsulation; Since the model's inference process is executed at the edge, the computational process required for model inference is encapsulated into a function. The input is the training dataset, model weights, and parameters, and the output is the fault diagnosis result. Sub-step 2: Model generation; Use the MATLAB Coder toolbox to generate the encapsulated functions into a library; it is recommended to generate a mex file, execute the mex file to verify the translation result, and then generate the library. Sub-step 3: Program porting; Use Keil uVision5 to create a project file, port the lib library into the project file and declare it, copy the Matlab header functions into the project file and declare them; Sub-step 4: Program writing; Complete the header file declaration, initialization, and configuration of the main program in the project file main.c. Add other functions or project modules as needed; Compile and burn the program and verify it on the STM32. Step 6: Fault Diagnosis Test data is sent to the STM32 via serial port, and the improved fuzzy width learning fault diagnosis method is run on the STM32 to perform fault diagnosis.