Acoustic vibration signal fusion RV reducer fault diagnosis method based on ALO-VMD-PCA

Abstract

The invention discloses an acoustic vibration signal fusion RV reducer fault diagnosis method based on ALO-VMD-PCA. The problems of low efficiency, local optimum and mode aliasing caused by the fact that existing VMD parameter optimization depends on experience / a traditional algorithm are solved. Sound and vibration signals of the RV speed reducer are synchronously collected; a variational mode decomposition (VMD) parameter is optimized through an ant lion optimization (ALO) algorithm in combination with a fitness function of'average envelope entropy + energy loss coefficient + mode orthogonality coefficient '; carrying out VMD decomposition and optimal modal screening, and carrying out dimension reduction fusion on sound and vibration characteristics by adopting principal component analysis (PCA); and inputting the fusion signal into a neural network classifier to output a diagnosis result. According to the method, the VMD parameter optimization time is only 36.3 s, the modal aliasing rate is 7.3%, the fault diagnosis accuracy rate reaches 98.5%, and the method is suitable for efficient and accurate fault diagnosis of the RV speed reducer and can be integrated into an industrial robot control system to achieve online monitoring.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault diagnosis technology, specifically to a fault diagnosis method for RV reducers based on ALO-VMD-PCA acoustic-vibration signal fusion. Background Technology

[0002] As a core power component of industrial robots and precision transmission equipment, the operational reliability of RV reducers directly determines the service life of the equipment and production safety. Under complex working conditions, RV reducers are subjected to alternating loads, impact vibrations, and other effects over long periods, making them prone to typical faults such as sun gear wear, planetary gear wear, needle roller bearing peeling, and cycloidal gear peeling. Failure to diagnose these faults in a timely manner may lead to equipment downtime, production interruption, or even safety accidents. Therefore, efficient fault diagnosis of RV reducers is of significant engineering importance.

[0003] Traditional RV reducer fault diagnosis methods mostly rely on single signal source analysis, and are mainly divided into two categories: one is based on vibration signals, which can effectively reflect the transient impact characteristics of internal mechanical faults. However, the internal structure of RV reducers is compact and the transmission relationship is complex, and the installation space for vibration sensors is limited, making it difficult to directly collect vibration data from key parts. The other is based on acoustic signals, which are convenient to collect and do not require direct contact with the equipment. However, the strong interference from environmental noise (such as motor noise and airflow noise) in the field makes the acoustic signal fault characteristics blurry and has low identification. The limitations of single signal source diagnosis can no longer meet the diagnostic needs under complex working conditions.

[0004] To address the shortcomings of single-source signals, existing technologies have shifted towards multi-sensor data fusion strategies, which improve diagnostic robustness and accuracy by leveraging complementary information from vibration and acoustic signals. Currently, various multi-source signal fusion diagnostic methods have emerged, such as the acoustic-vibration signal fusion algorithm based on VMD-KPCA and the multi-parameter fusion method based on DS evidence theory. However, these methods still have many problems in practical applications: First, the parameter selection of the core signal processing module (such as VMD) relies on manual experience adjustment. The combination of key parameters such as the number of modes K, penalty factor α, and time scale τ lacks scientific optimization basis, which easily leads to insufficient signal decomposition, severe mode aliasing, and inability to effectively extract weak fault features. Second, some methods use particle swarm optimization (PSO) and genetic algorithm (GA) to optimize VMD parameters, but PSO is prone to getting trapped in local optima, and GA has high computational complexity and long optimization time, making it difficult to meet the needs of real-time fault diagnosis. Third, the correlation between signals is not fully considered in the multi-modal feature fusion process, and the application of feature dimensionality reduction methods (such as traditional PCA) is not standardized, resulting in insufficient distinguishability of fused features. For example, although the EMD-PCA method can achieve feature fusion, the retention rate of detailed fault features is low, and the diagnostic accuracy is only 92%. Fourth, existing multi-source fusion methods generally suffer from the contradiction of being unable to balance "parameter optimization efficiency" and "feature extraction accuracy". Either the optimization speed is fast but the modal mixing rate is high (such as PSO-VMD modal mixing rate of 10.5%), or the modal separation effect is good but the computational cost is too high (such as GA-VMD), which cannot meet the dual requirements of diagnostic efficiency and accuracy in industrial scenarios.

[0005] Furthermore, existing single-signal diagnostic methods suffer from high false positive rates. For example, the false positive rate for single-source diagnosis based on vibration signals is approximately 15%-20%, while that based on acoustic signals exceeds 25%. Even with multi-source fusion, improper parameter optimization can still result in a false positive rate below 10%. Simultaneously, under complex operating conditions (such as high speed and strong noise), existing methods lack robustness. When noise interference reaches 30%, the accuracy of traditional single-signal CNN diagnosis is only around 75%, making it difficult to effectively identify early, subtle faults. These problems severely restrict the engineering application of multi-source signal fusion methods in RV reducer fault diagnosis. Therefore, a fault diagnosis method capable of intelligent parameter optimization, efficient signal decomposition, and accurate feature fusion is urgently needed. Summary of the Invention

[0006] This application proposes a fault diagnosis method for RV reducers based on ALO-VMD-PCA acoustic-vibration signal fusion. It aims to solve the core technical problems of "low parameter optimization efficiency, easy getting trapped in local optima, and severe mode mixing" caused by the use of manual experience or traditional optimization algorithms (PSO, GA) to adjust VMD parameters in the existing multi-source acoustic-vibration signal fusion diagnosis of RV reducers. Furthermore, it solves the derivative problems of insufficient fault feature extraction and difficulty in balancing diagnostic accuracy and real-time performance.

[0007] To achieve the above objectives, this application adopts the following technical solution: a fault diagnosis method for RV reducers based on ALO-VMD-PCA acoustic-vibration signal fusion, comprising the following steps: (1) Signal acquisition: Vibration and sound signals of the RV reducer under normal and fault conditions are acquired simultaneously using an accelerometer and a sound pressure sensor; (2) ALO optimization of VMD parameters: Set the population size and maximum number of iterations of the antlion optimization algorithm, and determine the range of values ​​for the number of modes K, penalty factor α, and time scale τ of variational mode decomposition (VMD); construct a fitness function based on the average envelope entropy Ee, energy loss coefficient Ec, and mode orthogonality coefficient Corth, and obtain the optimal parameter combination (K*, α*, τ*) of VMD through random walk of the antlion population, fitness value calculation, and elite antlion update. (3) VMD decomposition: The optimal parameter combination (K*, α*, τ*) is used to perform VMD decomposition on the vibration signal and the acoustic signal respectively to obtain multiple modal components (IMF). (4) Optimal mode selection: Calculate the correlation coefficient between each modal component and the original signal, select the optimal modes with a correlation coefficient greater than 0.7, and obtain the optimal mode Xa of the acoustic signal and the optimal mode Xv of the vibration signal; (5) PCA dimensionality reduction fusion: Construct the joint mode matrix of the optimal mode Xa of the acoustic signal and the optimal mode Xv of the vibration signal, calculate the covariance matrix and obtain the acoustic-vibration fusion signal through principal component projection; (6) Fault diagnosis: Input the acoustic vibration fusion signal into the neural network classifier and output the fault diagnosis result of the RV reducer.

[0008] Furthermore, in step (1), the fault conditions include cycloidal wheel peeling, needle roller bearing peeling, sun gear wear, and planetary gear wear. The sensor sampling frequency is 16384Hz, and 200 sets of valid samples are collected for each condition.

[0009] In further step (2), the parameters of the antlion optimization algorithm are set as follows: antlion population size 30, ant population size 50, maximum number of iterations 100; the value range of the VMD parameter is: K∈[3,12], α∈[1000,5000], τ∈[0.01,0.1].

[0010] Furthermore, in step (2), the expression for the fitness function is: The weight parameters are w1=0.4, w2=0.3, and w3=0.3; Ee is the average envelope entropy, Ec is the energy loss coefficient, and Corth is the modal orthogonality coefficient.

[0011] Furthermore, in step (2), the position update formula for the random walk of the antlion population is as follows: ,in Let i be the position of the i-th ant in generation t. This is the optimal solution for the t-th generation. .

[0012] Furthermore, the constrained optimization problem expression of the VMD decomposition is: Where α is the penalty factor and λ is the Lagrange multiplier.

[0013] Furthermore, in step (5), the formula for calculating the fusion weight of the principal component projection is: The fused signal expression is ; where ρ is the intermodal correlation coefficient, and σa and σv are the standard deviations of the optimal modes of the acoustic and vibration signals, respectively.

[0014] Furthermore, in step (6), the neural network classifier is a convolutional neural network (CNN) or a long short-term memory network (LSTM); the network structure of the CNN includes 3 Conv1D layers, 2 MaxPooling1D layers and 1 GlobalAveragePool1D layer.

[0015] The present invention has the following beneficial effects: 1. This application provides a fault diagnosis method for RV reducers based on ALO-VMD-PCA acoustic-vibration signal fusion. It employs the Antlion Optimization (ALO) algorithm to globally optimize the key parameter combination (mode number K, penalty factor α, and time scale τ) of Variational Mode Decomposition (VMD), overcoming the limitations of traditional parameter selection relying on manual experience. This effectively avoids the problems of Particle Swarm Optimization (PSO) easily getting trapped in local optima and the low computational efficiency of Genetic Algorithm (GA). Experimental results show that the optimization time of this parameter optimization method is only 36.3s, which is 25.3% shorter than PSO-VMD and 48.0% shorter than GA-VMD. The mode aliasing rate is as low as 7.3%, significantly improving the accuracy and efficiency of signal decomposition. Simultaneously, Principal Component Analysis (PCA) is introduced to reduce the dimensionality of the optimal acoustic-vibration modes through fusion. By dynamically calculating the fusion weights, the complementary features of the dual-source signals are fully explored, significantly improving the distinguishability of different fault modes. Compared with the traditional EMD-PCA fusion method, the fault feature point retention rate is improved by 15%. The above results show that the fault identification accuracy rate reaches 98.5%, which is 9.3 percentage points higher than the single signal diagnosis method, demonstrating significant advantages in both feature representation capability and fault identification accuracy.

[0016] 2. The ALO-VMD-PCA-based acoustic-vibration signal fusion method for RV reducer fault diagnosis provided in this application guides the ALO algorithm to optimize VMD parameters by constructing a multi-objective fitness function (fusion of average envelope entropy, energy loss coefficient, and modal orthogonality coefficient). This ensures the independence of modal components while maximizing the preservation of energy and detailed features of the fault signal, overcoming the signal decomposition bias problem caused by traditional single evaluation indicators and effectively solving the pain point of difficulty in extracting weak fault features. Simultaneously, by combining Z-score normalization and principal component projection fusion strategies, the dimensional differences and redundant information of the acoustic-vibration signals are eliminated, making the fault features of the fused signal more concentrated. Even under 30% noise interference, the fault diagnosis accuracy remains above 93.7%, significantly improving robustness compared to the traditional single-signal CNN diagnostic method (accuracy of only 75% under noise interference). This provides reliable technical support for early fault diagnosis of RV reducers under complex operating conditions.

[0017] 3. The RV reducer fault diagnosis method based on ALO-VMD-PCA acoustic-vibration signal fusion provided in this application adopts an integrated framework of "ALO optimized VMD + PCA fusion + neural network classification". It inherits the advantages of ALO algorithm's strong global search capability and fast convergence speed, and also leverages the characteristics of PCA's dimensionality reduction and redundancy removal and neural network's accurate classification. The entire process requires no manual intervention, realizing automated processing from signal acquisition to fault diagnosis. Moreover, the method's memory usage is only 83MB, which is 61.7% less than GA-VMD-PCA. It can be directly integrated into the industrial robot control system, solving the problems of high computational complexity and poor engineering applicability of traditional multi-source fusion methods. While ensuring diagnostic accuracy, it significantly improves the method's real-time performance and engineering feasibility, and reduces equipment maintenance costs and production safety risks.

[0018] 4. The RV reducer fault diagnosis method based on ALO-VMD-PCA acoustic and vibration signal fusion provided in this application further balances the global exploration and local development capabilities of the algorithm by introducing an adaptive weighting factor and a Latin hypercube sampling strategy into the ALO algorithm. This reduces the VMD parameter optimization time to 32.1s, increases the modal orthogonality coefficient to 0.902, and achieves a fault diagnosis accuracy of up to 99.1%. Simultaneously, the PCA fusion process accurately quantifies the correlation of acoustic and vibration signals through the covariance matrix, enabling the fused signal to more accurately reflect the nature of the fault. Even in scenarios with similar characteristics, such as needle roller bearing spalling and normal operating conditions, the false judgment rate is as low as 0.9%, further improving the reliability of diagnosis compared to the basic scheme. It is particularly suitable for the accurate identification of early, weak faults, providing stronger technical support for the preventive maintenance of RV reducers. Attached Figure Description

[0019] The accompanying drawings, which form part of this specification, illustrate embodiments disclosed in this application and, together with the specification, serve to explain the principles disclosed in this application.

[0020] This disclosure will become clearer with reference to the accompanying drawings and the following detailed description, wherein: Figure 1 This is a flowchart of the steps of the present invention; Figure 2 Here is a flowchart of the ALO-VMD-PCA algorithm; Figure 3 Diagram showing the experimental platform and sensor layout; Figure 4 A schematic diagram of the fault location in the RV reducer; Figure 5 The time-domain and frequency-domain plots of the acoustic vibration signal; Figure 6Trend charts for VMD parameter optimization for different algorithms; Figure 7 The modes and optimal mode diagrams of the vibration signal after VMD processing; Figure 8 The modalities and optimal modal diagrams of the acoustic signal after VMD processing; Figure 9 A comparison chart of the fused signals from VMD-PCA and EMD-PCA; Figure 10 The graph shows the trends of accuracy, loss value, F1 score, and recall for different models. Figure 11 This is a comparison chart of the accuracy rates of different diagnostic models.

[0021] In the diagram: 1. Input motor; 2. RV reducer; 3. Accelerometer; 4. Sound pressure sensor. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0023] Please see Figure 1-11 The RV reducer fault diagnosis method based on ALO-VMD-PCA acoustic and vibration signal fusion provided in this application achieves efficient and accurate diagnosis by globally optimizing VMD parameters through the Antlion Optimization (ALO) algorithm and combining it with PCA dimensionality reduction fusion of acoustic and vibration signal features.

[0024] Variational Mode Decomposition (VMD) processes signals by decomposing the input signal into multiple modal components with different frequency characteristics. Its decomposition performance is directly determined by two key parameters: the number of modes and the penalty factor; therefore, these parameters need to be optimized. Optimization algorithms can efficiently determine the optimal parameter combination.

[0025] See appendix Figure 2 Antlion Optimization Algorithm (ALO) is a nature-inspired swarm intelligence algorithm that simulates the predation behavior of ants and antlions. Each antlion is considered a potential solution, and its current position is updated by simulating the random walk behavior of ants foraging. It moves towards regions with better solutions to generate new candidate solutions, ultimately finding the global optimum. ALO enhances the algorithm's global exploration capability and can efficiently optimize the parameter selection process of Virtual Machine Determination (VMD). The ALO-VMD-PCA algorithm flow is as follows: Figure 2 As shown VMD solves variational problems to transform the original signal It is decomposed into multiple modal components containing a specific center frequency \({\omega_k}\). Its mathematical expression is: in, The number of modal components; For the first One mode; For the corresponding center frequency; It is a unit impulse function.

[0026] Introducing the Lagrange multiplier With penalty factor Construct a constrained optimization problem: in, Controlling the sparsity of the decomposition; It is a Lagrange multiplier.

[0027] ALO optimizes the parameters K, α, and τ of VMD by simulating antlion predation behavior. The specific steps are as follows: (1) Parameter initialization and initial population generation: Set the relevant parameters of the Antlion optimization algorithm (including population size, maximum number of iterations, etc.), and generate the initial population in a random manner. Each individual represents a set of VMD parameter combinations (K, α, τ).

[0028] (2) Fitness function calculation: Design a fitness function, apply the corresponding parameters to VMD for signal decomposition for each individual, and then calculate the fitness value of the decomposition result: in, The average envelope entropy; This is the energy loss coefficient; The modal component orthogonality coefficients (range (0,1), larger values ​​are better); w is the weighting parameter, set to a default value. .

[0029] (3) Antlion optimization execution: Simulate antlion random walks to search for the optimal solution, and update the parameter positions of individuals in the population using the following formula: in, For the tth generation The location of a single ant (i.e., VMD parameters); This is the current optimal solution; Random numbers used to control the search range.

[0030] (4) Iterative update: Record the optimal fitness value of each generation, and gradually optimize the VMD parameter combination through multiple iterations to achieve the optimal decomposition effect.

[0031] (5) Output optimal parameters: When the required number of iterations is reached or the convergence condition is met, the optimal parameters K*, α* and τ* of VMD can be obtained.

[0032] Using optimal parameters K∗, α∗, and τ∗ to set the VMD, the vibration and acoustic signals are processed respectively to obtain the optimal modes Xa=[x1a,x2a,...,xma]T and Xv=[x1v,x2v,...,xkv]T that retain the most feature points of the original signal, and the joint mode matrix is ​​constructed: Then the covariance matrix is ​​calculated: in, Let be the intermodal correlation coefficient, and .

[0033] Project the principal components and calculate the fusion weights: The final fused signal is: See appendix Figure 1-11 As shown, the specific steps of the RV reducer fault diagnosis method based on ALO-VMD-PCA acoustic-vibration signal fusion provided in this application are as follows: 1. Steps for synchronous acquisition of acoustic and vibration signals: A signal acquisition platform was constructed, comprising an input motor 1, an RV reducer 2, an accelerometer 3, and a sound pressure sensor 4. An NV9822 accelerometer 3 (measuring range ±50g, sensitivity 100mV / g) and an INV9206 sound pressure sensor 4 (frequency range 20Hz-20kHz, sensitivity 50mV / Pa) were used, respectively installed at key locations on the housing of the RV reducer 4 (vibration sensor close to the bearing mounting point, sound pressure sensor 10cm from the housing surface). The input motor 1 provides driving force to the RV reducer 4, and the output shaft of the RV reducer 4 is connected to the load. Signals from the RV reducer under five operating conditions were collected simultaneously: normal, sun gear wear, planetary gear wear, needle roller bearing peeling, and cycloidal gear peeling. The sampling frequency was set to 16384Hz. 200 valid samples were collected for each operating condition. Each sample contained 16384 data points, forming a dataset of "5 operating conditions × 2 signal modes × 200 samples".

[0034] In summary, ensuring the time synchronization of vibration and acoustic signals and acquiring raw data covering different fault characteristics provides a high-quality data source for subsequent signal processing.

[0035] As shown in Figure 4, the top of the input shaft has a sun gear peeling fault, with a peeling length of 5mm and a depth of 3mm; the planet gear has a peeling and broken tooth fault, with a depth of 2mm; the needle roller bearing has a peeling fault with a depth of 1.5mm; the cycloidal gear has a tooth tip peeling fault, with a diameter of 0.8mm and a depth of 0.4mm; and the needle roller bearing has a single roller peeling fault with a depth of 2mm.

[0036] The sampling frequency and number of sampling points were both set to 16384, with 200 sets of data collected for each operating condition. The complete fault types and corresponding labels in the database are shown in Table 1. The time-domain plot of the acoustic and vibration signals of the RV reducer collected in the experiment is shown in Figure 5. Analysis of the experimental data shows that the vibration and acoustic signals of the same fault have high similarity, but the vibration signals of different faults are difficult to distinguish. Specifically, the needle roller bearing peeling fault and the vibration and acoustic signals of a normal RV reducer have certain similarities in both the time and frequency domains, making direct differentiation difficult. Other faults, although different, cannot be directly used as fault judgment criteria.

[0037] Table 1 Database Fault Types and Corresponding Tags 2. Steps for ALO optimization of VMD parameters: 2.1 Parameter Initialization: First, parameter initialization is required. ALO parameter settings are shown in Table 2: Table 2 ALO Algorithm Initialization Parameter Settings As can be seen from Table 2 above, the core parameters of the antlion optimization algorithm are: antlion population size 30, ant population size 50, maximum number of iterations 100, and parameter search precision 0.01. The range of values ​​for the VMD optimization parameters is as follows: number of modes K∈[3,12], penalty factor α∈[1000,5000], time scale τ∈[0.01,0.1], lower limit of parameters is [3,1000,0.01], upper limit is [12,5000,0.1]; This step clarifies the search space and algorithm operation rules for parameter optimization, providing the basic conditions for ALO to globally search for the optimal VMD parameters, and avoiding the optimization efficiency being reduced due to an excessively large search range or the optimal solution being missed due to an excessively small search range.

[0038] 2.2. Fitness function construction steps: Construct a multi-objective fitness function to evaluate the VMD decomposition effect, the expression of which is: Where Ee is the average envelope entropy (reflecting the complexity of the modal components; the smaller the value, the more concentrated the fault characteristics), Ec is the energy loss coefficient (reflecting the degree of signal energy retention during decomposition; the smaller the value, the better), and Corth is the modal orthogonality coefficient (reflecting the independence of modal components; the value ranges from 0 to 1; the larger the value, the better the modal separation effect). The weight parameters are set to w1=0.4, w2=0.3, and w3=0.3, and the three objectives of "feature concentration, energy retention rate, and modal independence" are balanced by weighting. This step establishes a quantitative correspondence between VMD parameter combinations and decomposition results, providing an evaluation standard for the ALO algorithm to search for optimal parameters and avoiding decomposition bias caused by a single evaluation index.

[0039] 2.3 Parameter Optimization Iteration: 30 sets of VMD parameters were randomly generated as initial "antlion individuals," and then optimized using the formula... Update ant location ( Let i be the position of the i-th ant in generation t. This is the optimal solution for the t-th generation. Each set of parameters is input into the VMD decomposition signal, the fitness value is calculated, the top 5 best individuals of each generation are retained as guides, and the iteration is continued until the maximum number of iterations (100 times) or the fitness change is < 0.001 for 10 consecutive generations, and the optimal parameter combination (K*, α*, τ*) is output.

[0040] By using ALO global search capability to replace manual / traditional algorithm parameter tuning, local optima are avoided, and the optimal VMD parameters adapted to the acoustic and vibration signals are obtained.

[0041] In summary, to compare the optimization performance of ALO, Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) were introduced as a comparison. As shown in Figure 6 and Table 3, PSO-VMD performed moderately in all indicators, with higher optimization time and memory consumption than ALO-VMD, but lower than Genetic Algorithm-VMD. Although GA-VMD outperformed the other two algorithms in terms of mode aliasing rate and orthogonality coefficient, it had the highest computational cost and the longest optimization time, which may not be able to achieve real-time fault diagnosis in the field of fault monitoring. ALO-VMD, on the other hand, performed best in terms of computational efficiency, memory consumption, and optimization accuracy. It not only had the shortest optimization time, the lowest memory consumption, and the lowest computational cost, but also showed a faster convergence speed and lower computational overhead.

[0042] Table 3. Results and performance comparison of the three optimization methods 3. VMD Signal Decomposition The optimal parameters (K*, α*, τ*) are input into the VMD module to decompose the vibration signal and the acoustic signal respectively, obtaining K* modal components (IMFs); the core constrained optimization problem of VMD decomposition is: The optimized parameters enable adaptive frequency band segmentation of the signal, separating noise and fault characteristics.

[0043] 4. Optimal mode selection steps: Calculate the correlation coefficient between each modal component and the original signal, select the top 3 modal components with a correlation coefficient > 0.7, and sum them according to their correlation coefficient ratios to obtain the optimal modes Xa (acoustic signal) and Xv (vibration signal).

[0044] The optimized VMD parameters enable accurate signal decomposition, separating noise and fault features. Optimal mode selection eliminates redundant information, retains the feature components that contribute most to fault diagnosis, and reduces the complexity of subsequent fusion calculations.

[0045] In summary, VMD was initialized using ALO-optimized VMD parameters to decompose all vibration and acoustic signals of the RV reducer. Figures 7 and 8 show all modes of different vibration and acoustic signals after VMD processing, with the red portion representing the optimal mode that retains the most feature points of the original signal.

[0046] 5. PCA dimensionality reduction and fusion steps: Z-score normalization is performed on Xa and Xv to construct the joint mode matrix. (N is the number of sample points); Calculate the covariance matrix. (ρ is the correlation coefficient between modes; principal components with a variance contribution rate > 95% are selected through eigenvalue decomposition, and the fusion weights are calculated.) Generate fusion signal .

[0047] It can eliminate dimensional differences and redundant information, integrate complementary features of acoustic and vibration signals, and enhance the distinguishability of fault features.

[0048] Figure 9 shows the fused signal obtained by PCA fusion and compares it with the result after Empirical Mode Decomposition (EMD) fusion. The results show that both VMD-PCA and EMD-PCA can effectively extract useful features from cluttered time-domain signals and distinguish different fault signals that were originally similar. However, VMD can retain more feature points and better preserve the detailed information in the signal, especially in the preservation of certain key points and anomalies. The decomposition result of EMD is smoother, but the preservation effect of some detailed features is poor, and the feature point extraction accuracy is lower than that of VMD. Moreover, the computation time of the two is similar.

[0049] 6. Fault diagnosis steps: Input the fused signal into a convolutional neural network (CNN) or a long short-term memory network (LSTM): CNN structure: Input layer (16384×2) → 3 Conv1D layers → 2 MaxPooling1D layers → GlobalAveragePool1D layer → Dense layer → Output layer (Softmax activation); LSTM structure: Input layer (16384×2) → 3 LSTM layers → GlobalAveragePool1D layer → Dense layer → Output layer (Softmax activation).

[0050] Set the number of iterations to 10, batch size to 32, and learning rate to 0.001, and output the fault diagnosis results.

[0051] By leveraging the feature extraction capabilities of neural networks, accurate fault classification of fused signals can be achieved.

[0052] In this step, to verify whether the fused signal retains the characteristics of the original signal, the original vibration signal, acoustic signal, and PSO-VMD-PCA fused signal were simultaneously input into a convolutional neural network (CNN) and a long short-term memory network (LSTM) for classification. The CNN and LSTM network structures constructed in the experiment are shown in Table 4 and Table 5, respectively.

[0053] Table 4 CNN Network Structure Table 5 LSTM Network Structure In the experiment, the number of iterations for both CNN and LSTM was set to 10. Figure 10 This demonstrates the trends in accuracy, loss, F1 score, and recall across different iteration rounds; Figure 11The accuracy of different models was compared. Chart analysis revealed that the model using vibration signals and VMD showed a more rapid decrease in loss value and a more significant improvement in accuracy. Both the constructed CNN and LSTM models demonstrated excellent performance in the RV reducer fault classification task, with CNN outperforming LSTM. Comparative analysis of model accuracy showed that both EMD and VMD effectively extracted fault features from vibration and acoustic signals, but the features extracted by VMD exhibited stronger discriminative power in identification and classification. Experimental results validated the effectiveness of the proposed ALO-VMD-PCA algorithm.

[0054] In summary, to address the challenges of feature extraction from multi-source signals, modal aliasing, and low efficiency in parameter optimization for RV reducers, this application proposes an ALO-VMD-PCA acoustic-vibration signal fusion fault diagnosis method. This method adaptively optimizes key parameters (mode number K, penalty factor α, and time scale τ) of Variational Mode Decomposition (VMD) using the Antlion Optimization Algorithm (ALO), effectively overcoming the limitations of traditional methods that rely on human experience and are prone to getting trapped in local optima, significantly improving the efficiency and accuracy of signal decomposition. Furthermore, combining the complementary characteristics of vibration and acoustic signals, Principal Component Analysis (PCA) is used to reduce the dimensionality of multimodal components through fusion, preserving signal details while enhancing the discriminative power of fault features. Experimental results show that compared to Particle Swarm Optimization (PSO) and Genetic Algorithm (GA), ALO-optimized VMD has higher parameter selection efficiency, with an optimization time of only 36.3 seconds, reduced memory usage to 83MB, a modal orthogonality coefficient of 0.887, and a 3.2% reduction in modal aliasing rate (7.3%) compared to PSO-VMD. Compared with the Empirical Mode Decomposition-Principal Component Analysis (EMD-PCA) method, VMD-PCA fused signals have advantages in feature point preservation and detail representation, and the computation time of both methods is similar, verifying the applicability of the proposed method in real-time scenarios. The multi-source signal fusion strategy overcomes the shortcomings of single-sensor signal representation by constructing a joint mode matrix and principal component projection, providing a new approach for diagnosing early, subtle faults under complex operating conditions.

Claims

1. A fault diagnosis method for RV reducers based on ALO-VMD-PCA acoustic-vibration signal fusion, characterized in that, Includes the following steps: (1) Signal acquisition: Vibration and sound signals of the RV reducer under normal and fault conditions are acquired simultaneously using an accelerometer and a sound pressure sensor; (2) ALO optimization of VMD parameters: Set the population size and maximum number of iterations of the antlion optimization algorithm, and determine the range of values ​​for the number of modes K, penalty factor α, and time scale τ of variational mode decomposition (VMD); construct a fitness function based on the average envelope entropy Ee, energy loss coefficient Ec, and mode orthogonality coefficient Corth, and obtain the optimal parameter combination (K*, α*, τ*) of VMD through random walk of the antlion population, fitness value calculation, and elite antlion update. (3) VMD decomposition: The optimal parameter combination (K*, α*, τ*) is used to perform VMD decomposition on the vibration signal and the acoustic signal respectively to obtain multiple modal components (IMF). (4) Optimal mode selection: Calculate the correlation coefficient between each modal component and the original signal, select the optimal modes with a correlation coefficient greater than 0.7, and obtain the optimal mode Xa of the acoustic signal and the optimal mode Xv of the vibration signal; (5) PCA dimensionality reduction fusion: Construct the joint mode matrix of the optimal mode Xa of the acoustic signal and the optimal mode Xv of the vibration signal, calculate the covariance matrix and obtain the acoustic-vibration fusion signal through principal component projection; (6) Fault diagnosis: Input the acoustic vibration fusion signal into the neural network classifier and output the fault diagnosis result of the RV reducer.

2. The method for fault diagnosis of RV reducers based on ALO-VMD-PCA acoustic-vibration signal fusion according to claim 1, characterized in that, In step (1), the fault conditions include cycloidal wheel peeling, needle roller bearing peeling, sun gear wear, and planetary gear wear. The sensor sampling frequency is 16384Hz, and 200 valid samples are collected for each condition.

3. The method for fault diagnosis of RV reducers based on ALO-VMD-PCA acoustic-vibration signal fusion according to claim 1, characterized in that, In step (2), the parameters of the antlion optimization algorithm are set as follows: antlion population size 30, ant population size 50, maximum number of iterations 100; the value range of the VMD parameter is: K∈[3,12], α∈[1000,5000], τ∈[0.01,0.1].

4. The method for fault diagnosis of RV reducers based on ALO-VMD-PCA acoustic-vibration signal fusion according to claim 1, characterized in that, In step (2), the expression for the fitness function is: The weight parameters are w1=0.4, w2=0.3, and w3=0.3; Ee is the average envelope entropy, Ec is the energy loss coefficient, and Corth is the modal orthogonality coefficient.

5. The method for fault diagnosis of RV reducers based on ALO-VMD-PCA acoustic-vibration signal fusion according to claim 1, characterized in that, In step (2), the position update formula for the random walk of the antlion population is as follows: ,in Let i be the position of the i-th ant in generation t. This is the optimal solution for the t-th generation. .

6. The method for fault diagnosis of RV reducers based on ALO-VMD-PCA acoustic-vibration signal fusion according to claim 1, characterized in that, The constrained optimization problem expression of the VMD decomposition is: Where α is the penalty factor and λ is the Lagrange multiplier.

7. The method for fault diagnosis of RV reducers based on ALO-VMD-PCA acoustic-vibration signal fusion according to claim 1, characterized in that, In step (5), the formula for calculating the fusion weight of the principal component projection is: The fused signal expression is ; where ρ is the intermodal correlation coefficient, and σa and σv are the standard deviations of the optimal modes of the acoustic and vibration signals, respectively.

8. The method for fault diagnosis of RV reducers based on ALO-VMD-PCA acoustic-vibration signal fusion according to claim 1, characterized in that, In step (6), the neural network classifier is a convolutional neural network (CNN) or a long short-term memory network (LSTM); the network structure of the CNN includes 3 Conv1D layers, 2 MaxPooling1D layers and 1 GlobalAveragePool1D layer.

Images