An IEMLLE-WKELM-based acceleration sensor fault diagnosis method
By using the IEMLLE-WKELM algorithm to reduce dimensionality and optimize accelerometer sensor data, the problems of detection speed and accuracy caused by the large number of features in traditional methods are solved, and efficient fault diagnosis is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WENZHOU VOCATIONAL COLLEGE OF SCI & TECH
- Filing Date
- 2026-03-10
- Publication Date
- 2026-06-05
AI Technical Summary
Existing accelerometer data features numerous characteristics, which affects detection speed and accuracy. Traditional dimensionality reduction algorithms perform poorly when processing sparse data, making it difficult to effectively extract information. Furthermore, they are prone to overfitting during machine learning, increasing training costs.
Dimensionality reduction is achieved using the Local Linear Embedding (IEMLLE) algorithm based on information entropy measurement, and the penalty factor and kernel parameters of the Weighted Kernel Extreme Learning Machine (WKELM) are optimized by combining the Chaotic Particle Swarm Optimization (CPSO) algorithm to construct a fault diagnosis model. Feature information is then extracted through time-frequency analysis.
The accuracy of accelerometer fault diagnosis was improved to 99.375%, which is better than traditional methods, demonstrating efficient feature extraction and model optimization capabilities.
Smart Images
Figure CN122153534A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of acceleration sensor technology, and more specifically to an acceleration sensor fault diagnosis method based on IEMLE-WKELM. Background Technology
[0002] Accelerometers are used to acquire vibration signals from mechanical equipment, which is crucial for monitoring the condition of such equipment and improving its safety and reliability. However, the complex characteristics of accelerometer output signals pose a significant challenge to developing effective fault diagnosis methods.
[0003] Common accelerometers include capacitive, inductive, strain gauge, piezoresistive, and piezoelectric types. In recent years, the emergence of microelectromechanical systems (MEMS) technology has marked the beginning of a new era, enabling accelerometers to be manufactured small enough to be easily integrated into mobile phones and wearable devices.
[0004] Taking a piezoelectric accelerometer as an example, its working principle is based on the piezoelectric effect. When certain materials, such as quartz crystals or piezoelectric ceramics, are subjected to external forces, they deform, causing polarization within the crystal and generating opposite charges on its two surfaces. When the external force is removed, the crystal returns to its uncharged state. This phenomenon is called the piezoelectric effect. In other words, when a piezoelectric material (such as quartz or ceramic) is subjected to a force, it generates a voltage, which is proportional to the applied force.
[0005] In other words, the design of a piezoelectric accelerometer involves a piezoelectric element (quartz crystal, piezoelectric ceramic, organic piezoelectric material) being subjected to an external force in a certain direction, resulting in deformation, internal polarization, and surface charge generation. The charge / voltage is amplified by a measuring circuit, and the acceleration is finally calculated.
[0006] Therefore, the sensor outputs a lot of characteristic data, such as: acceleration waveform data, tilt, and derived vibration parameters (velocity, amplitude, crest factor), etc.
[0007] Therefore, data can be obtained by extracting signal features. There are three main methods for signal feature extraction: time domain, frequency domain, and time-frequency domain. These three methods can be used to extract feature information under different fault states from different perspectives.
[0008] Time-domain analysis refers to the time-domain analysis of vibration signals. The parameters that can be extracted include dimensional parameters such as root mean square, maximum value, variance, skewness, kurtosis, and peak-to-peak value, as well as dimensionless parameters such as waveform factor, peak factor, kurtosis factor, and skewness factor.
[0009] Frequency domain analysis methods analyze and process vibration signals at the frequency domain level to obtain highly sensitive frequency domain characteristic parameters, thereby enabling further determination of the type and location of the fault.
[0010] In practical engineering, most fault signals exhibit nonlinear and non-stationary characteristics. Simple time-domain or frequency-domain analysis is only applicable to stationary signals and cannot simultaneously consider both local and global features of the signal. Time-frequency analysis overcomes the limitations of the former two methods and can effectively extract feature information from nonlinear and non-stationary signals.
[0011] After obtaining feature data, the next step is data processing. In machine learning and data science problems, the main goal is to find the most relevant features that play a dominant role in determining and influencing the output results.
[0012] In most data science problems, machine learning datasets are filled with a large number of features, which can easily lead to overfitting and increase training costs, making the process quite slow.
[0013] Because the numerous features of accelerometer data can affect both the accuracy and speed of diagnosis, it is essential to simplify these features. These features impact not only detection speed but also accuracy. Since high-dimensional information can be represented in a low-dimensional space with minimal information loss, dimensionality reduction can lead to low-dimensional data, thus diminishing the features of the accelerometer data. Current dimensionality reduction algorithms include Locally Linear Embedding (LLE), Principal Component Analysis (PCA), and t-distribution-based Random Nearest Neighbor Embedding (t-SNE).
[0014] Manifold learning has various variants that address the problem of reducing the dimensionality and feature set of data from non-uniform singular surfaces through suboptimal data representations. This data representation selectively chooses data points from a low-dimensional manifold embedded in a high-dimensional space, attempting to generalize to linear frameworks like PCA. Manifolds have a planar and featureless feel, much like Euclidean space. The manifold learning problem is unsupervised; it learns the high-dimensional structure of the data from the data itself, without using pre-defined classifications or losing important information about the original variables. The goal of manifold learning algorithms is to recover the original domain structure, achieving a certain degree of scaling and rotation. The nonlinearity of these algorithms allows them to reveal the domain structure even when the manifold is not linearly embedded. Due to the sparsity of much accelerometer data, these dimensionality reduction algorithms perform poorly when processing accelerometer data, struggling to effectively extract information from sparse data.
[0015] Extreme Learning Machine (ELM) is an algorithm for solving single-hidden-layer neural networks. While maintaining learning accuracy, it is more efficient than traditional single-layer neural networks. To improve the performance of ELM in accelerometer fault diagnosis, a Weighted Kernel Extreme Learning Machine (WKELM) algorithm is proposed. This algorithm uses kernel functions instead of the random feature maps in the hidden layers containing activation functions, which improves the nonlinear processing capability and robustness of the weighted extreme learning machine. Furthermore, the weighted solution method is more suitable for imbalanced datasets, improving the machine's ability to identify minority class samples. Summary of the Invention
[0016] This invention proposes a fault diagnosis method for accelerometers. By using the Local Linear Embedding (LLE) algorithm based on Information Entropy Measurement (IEM) to reduce the features of accelerometer data, it can solve the influence of position differences of unaligned samples. The Chaotic Particle Swarm Optimization (CPSO) algorithm is used to optimize the penalty factor and kernel parameters of the weighted kernel extremum learning machine to avoid getting trapped in local optima.
[0017] To achieve the above objectives, the technical solution adopted by this invention is: an accelerometer fault diagnosis method based on IEMLE-WKELM, the method comprising the following steps:
[0018] Step 1: Feature Extraction: The time-frequency analysis method is used to extract feature information of the accelerometer under different fault states, including normal, offset, gain, and drift.
[0019] Step 2, Data Dimensionality Reduction: The extracted high-dimensional feature data is reduced in dimensionality using the Local Linear Embedding (LLE) algorithm based on Information Entropy Measurement (IEM) to obtain low-dimensional feature data. The implementation process of the IEMLLE algorithm is as follows:
[0020] For a high-dimensional dataset consisting of arbitrary sample points with N features, according to the formula... Calculate information entropy ,in The probability of a feature appearing;
[0021] According to the formula Calculate the entropy difference and select the k nearest neighbors for each sample based on the entropy difference;
[0022] According to the formula Calculate the reconstruction weight coefficient Under the condition that the constraints are met, the formula can be used to... st Minimizing the quadratic cost function completes the reconstruction of vector v;
[0023] According to the formula To embed high-dimensional data into a low-dimensional space;
[0024] Step 3: Model Construction and Parameter Optimization: A fault diagnosis model is constructed based on Weighted Kernel Extreme Learning Machine (WKELM). The Chaotic Particle Swarm Optimization (CPSO) algorithm is used to optimize the penalty parameter C and kernel parameters of WKELM. The implementation process of the WKELM algorithm is as follows:
[0025] Based on the Extreme Learning Machine (ELM) model, according to the formula
[0026] st Minimize the weighted cumulative error for each sample, where It is a diagonal matrix. For error;
[0027] According to the formula Calculate the weights of the ELM, where C is the penalty parameter and I is the identity matrix;
[0028] Introducing a kernel function to replace the hidden layer random feature mapping of ELM, according to the formula Construct the WKELM model;
[0029] The implementation process of the chaotic particle swarm optimization algorithm is as follows:
[0030] Using the logistic chaotic mapping according to the formula , Generate initial particle position and speed ,in It is a variable parameter, and its value range is... , To control the parameters, and then combine the formula Update particle position and velocity to complete parameter optimization, among which Indicates inertia weight, Indicates the number of iterations. and All are within the range of 0 to 1. ,in It is the current optimal position of the i-th particle. It is the optimal solution among all particles;
[0031] Step 4: Fault Diagnosis: Input the dimensionality-reduced low-dimensional feature data into the CPSO-optimized WKELM fault diagnosis model, and output the fault diagnosis results of the accelerometer.
[0032] Finally, tests showed that the accuracy rate of accelerometer diagnosis using IEMLE-WKELM was 99.375%, while the accuracy rate using LLE-WKELM was 95.625%, LLE-ELM was 94.375%, and PCA-ELM was 92.5%. This leads to the conclusion that IEMLE-WKELM has a higher accuracy rate for accelerometer diagnosis than other methods, demonstrating that the fault diagnosis method based on IEM, LLE, and weighted extreme learning machine in this invention is feasible for accelerometer diagnosis.
[0033] The method proposed in this invention reduces the features of accelerometer data through an IEM-based LLE algorithm, addressing the impact of misaligned sample position differences. It optimizes the penalty factor and kernel parameters of the weighted kernel extreme learning machine (ELM) using a chaotic particle swarm optimization (CPSO) algorithm to avoid getting trapped in local optima. The weighted kernel extreme learning machine algorithm with kernel functions improves the robustness and nonlinear processing capabilities of the traditional weighted extreme learning machine, enhancing the ELM's ability in accelerometer fault diagnosis. The feasibility of the IEM-based LLE and CPSO-based WKELM methods for identifying accelerometer faults is verified.
[0034] Finally, this invention also demonstrates that the fault diagnosis method based on IEM-based LLE and weighted extreme learning machine is feasible for diagnosing accelerometers. Experimental results show that the accuracy of accelerometer diagnosis using IEM-based LLE and weighted extreme learning machine is 99.375%, which is higher than other methods. Attached Figure Description
[0035] Figure 1 This is a flowchart of the fault diagnosis method of the present invention.
[0036] Figure 2 A flowchart illustrating the parameter optimization process of WKELM for CPSO. Detailed Implementation
[0037] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments and features described in these embodiments can be combined with each other.
[0038] refer to Figure 1 The present invention provides a fault diagnosis method for accelerometers based on IEMLE-WKELM, as follows:
[0039] I. Characteristic Acquisition of Signals Output by Accelerometers
[0040] This invention uses time-frequency domain analysis to extract feature information under fault conditions;
[0041] II. Local Linear Embedding (LLE) Algorithm Based on Information Entropy Measurement (IEM)
[0042] In the dimensionality reduction process of the LLE algorithm, Euclidean distance is used to select the nearest neighbor during feature extraction; however, there is a significant problem of sample position misalignment.
[0043] To address this issue, this invention proposes an LLE algorithm based on IEM to resolve the impact of sample location differences, where information entropy can solve the problem of information quantification.
[0044] For high-dimensional datasets (in (Represents any sample point with N features), calculate the information entropy. See formula (1).
[0045] (1)
[0046] in It represents the probability of occurrence. The entropy difference is calculated using formula (2):
[0047] (2)
[0048] The k nearest neighbors of a sample are selected by calculating the entropy difference.
[0049] Reconstructing weight coefficients The calculation method is shown in formula (3):
[0050] (3)
[0051] In minimizing the quadratic cost function, by To reconstruct vector v, its expression is shown in formula (4):
[0052] st (4)
[0053] Then, the formal expression of the low-dimensional embedding is given by formula (5):
[0054] (5)
[0055] III. Weighted Kernel Extreme Learning Machine (WKELM) Algorithm
[0056] The Extreme Learning Machine (ELM) model is shown in formula (6):
[0057] , (6)
[0058] in It is the hidden layer feature mapping matrix; It is the training target matrix.
[0059] The process of minimizing the weighted cumulative error for each sample is shown in formula (7):
[0060] st (7)
[0061] in It is a diagonal matrix. It is the error, while the weights of the Extreme Learning Machine (ELM) are determined by... Add to the main diagonal The result is shown in formula (8). Where C is the penalty parameter and I is the identity matrix.
[0062] (8)
[0063] Introducing kernel function pairs After substitution, the Weighted Extreme Learning Machine (WKELM) can be expressed as Equation (9).
[0064] (9)
[0065] in For kernel function ( (For kernel parameters).
[0066] Obviously, parameters C and It needs to be determined.
[0067] IV. Chaotic Particle Swarm Optimization (CPSO) Algorithm
[0068] In the Particle Swarm Optimization (PSO) algorithm, the position of the i-th particle is known. and speed Then, the position and velocity of each particle are updated according to the following formula (10).
[0069] (10)
[0070] in Indicates inertia weight; Indicates the number of iterations; and All are within the range of 0 to 1; ;in It is the current optimal position of the i-th particle. It is the optimal solution among all particles.
[0071] However, particle swarm optimization algorithms are prone to getting trapped in local optima.
[0072] This invention chooses to introduce chaotic dynamics to improve the particle swarm optimization algorithm, which can avoid getting trapped in local optima. It uses logistic mapping to solve the chaotic queue, that is, logistic chaotic mapping is used in the initialization stage.
[0073] Traditional particle swarm optimization algorithms typically use random numbers to generate initial particle positions and velocities during the initialization phase. However, randomly generated initial values can cause the algorithm to get stuck in local optima.
[0074] To address this issue, this invention proposes using logistic chaotic mapping to generate initial particle positions and velocities. Chaotic mapping has good randomness and can generate better initial values, thereby improving the global search capability of the algorithm, as shown in formula (11).
[0075] , (11)
[0076] in It is a variable parameter with a value range of 1000. ,and These are control parameters.
[0077] Introducing a chaotic queue into the Particle Swarm Optimization (PSO) algorithm yields the Chaotic Particle Swarm Optimization (CPSO) algorithm. The process of using CPSO to optimize the parameters of WKELM is described below. Figure 2 .
[0078] Combining the methods described above, the state types based on the accelerometer include four types: normal, offset, gain, and drifting.
[0079] In this invention, test samples are sampled in the experiment to obtain time-frequency image features of accelerometer state types based on empirical wavelet transform.
[0080] The training and test sample sets are dimensionality reduced by using IEM-based LLE to obtain training and test sample sets with low-dimensional features.
[0081] By using a low-dimensional feature training sample set and CPSO to optimize the penalty factor C and kernel function parameters (weights), fault diagnosis models for accelerometers based on IEM (Integrated Electron Mechanics) and CPSO (Programmable Calculator) were established. A comparison of the optimization processes of CPSO and PSO revealed that CPSO is superior to PSO.
[0082] The fault diagnosis results of IEMLE-WKELM for accelerometers show that the accuracy rate of IEMLE-WKELM for accelerometer diagnosis is 99.375%, compared to 95.625% for LLE-WKELM, 94.375% for LLE-ELM, and 92.5% for PCA-ELM. Therefore, it can be concluded that IEMLE-WKELM has a higher accuracy rate for accelerometer fault diagnosis than other methods. Experimental results demonstrate that the IEMLE-WKELM method has higher accuracy in accelerometer fault diagnosis than other methods.
[0083] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A fault diagnosis method for an accelerometer based on IEMLE-WKELM, characterized in that, The method includes: Step 1: Feature Extraction: The time-frequency analysis method is used to extract feature information of the accelerometer under different fault states, including normal, offset, gain, and drift. Step 2, Data Dimensionality Reduction: The extracted high-dimensional feature data is reduced in dimensionality using a local linear embedding algorithm based on information entropy measurement to obtain low-dimensional feature data; Step 3: Model Construction and Parameter Optimization: A fault diagnosis model is constructed based on a weighted kernel extreme learning machine, and the penalty parameter C and kernel parameters of the WKELM algorithm are optimized using a chaotic particle swarm optimization algorithm. Step 4: Fault Diagnosis: Input the dimensionality-reduced low-dimensional feature data into the CPSO-optimized WKELM fault diagnosis model and output the fault diagnosis results of the accelerometer.
2. The method for fault diagnosis of an accelerometer based on IEMLE-WKELM according to claim 1, characterized in that: The implementation process of the IEMLLE algorithm in step two is as follows: For a high-dimensional dataset consisting of arbitrary sample points with N features, according to the formula... Calculate information entropy ,in The probability of a feature appearing; According to the formula Calculate the entropy difference and select the k nearest neighbors for each sample based on the entropy difference; According to the formula Calculate the reconstruction weight coefficient Under the condition that the constraints are met, the formula can be used to... st Minimizing the quadratic cost function completes the reconstruction of vector v; According to the formula To enable the embedding of high-dimensional data into a low-dimensional space.
3. The method for fault diagnosis of an accelerometer based on IEMLE-WKELM according to claim 1, characterized in that: The implementation process of the WKELM algorithm in step three is as follows: Based on the Extreme Learning Machine model, according to the formula st Minimize the weighted cumulative error for each sample, where It is a diagonal matrix. For error; According to the formula Calculate the weights of the ELM, where C is the penalty parameter and I is the identity matrix; Introducing a kernel function to replace the hidden layer random feature mapping of ELM, according to the formula Construct the WKELM model.
4. The accelerometer fault diagnosis method based on IEMLE-WKELM according to claim 1, characterized in that: The implementation process of the chaotic particle swarm optimization algorithm in step three is as follows: Using the logistic chaotic mapping according to the formula , Generate initial particle position and speed ,in It is a variable parameter, and its value range is... , To control the parameters, and then combine the formula Update particle position and velocity to complete parameter optimization, among which Indicates inertia weight, Indicates the number of iterations. and All are within the range of 0 to 1. ,in It is the current optimal position of the i-th particle. It is the optimal solution among all particles.