Mechanical fault diagnosis method based on multi-strategy improved SSA Authors: Xiangtao Kong Published: 2026 ABSTRACT AbstractTo address the unstable target localization and insufficient path planning efficiency caused by complex lighting and dynamic environments during automated robotic arm loading, this study proposes a method for automated robotic arm loading that integrates visual guidance and an improved sparrow search algorithm. This method constructs a visual guidance system based on a depth camera, achieving reliable target object recognition and 3D pose mapping through template matching and depth information fusion, providing stable spatial perception information for robotic arm grasping. Furthermore, non-dominated sorting and multinomial mutation mechanisms are introduced to improve the traditional sparrow search algorithm, establishing a trajectory planning model with motion time and impact smoothness as optimization objectives, thereby enhancing the algorithm’s optimization ability under multi-objective constraints and dynamic environments. Experimental results show that in typical industrial scenarios, this method achieves a high grasping success rate and outperforms comparative methods in positioning accuracy and path planning efficiency, validating the effectiveness and practicality of the proposed method in automated robotic arm loading tasks.Keywords: Sparrow search algorithm, Multi-strategy improvement method, Machine learning, Feature extraction, Mechanical fault diagnosisSubject terms: Engineering, Mathematics and computing Introduction The fault diagnosis (FD) and operation state monitoring technologies of mechanical equipment have steadily emerged as the essential to ensuring production efficiency and safety as modern industry has evolved toward artificial intelligence1. Conventional mechanical FD techniques mostly use artificial experience and signal processing methods such as Fourier transform and wavelet analysis. The efficiency and diagnostic accuracy of FD are relatively low when dealing with high-dimensional, nonlinear, and very noisy industrial data2. The quick advancement of machine learning (ML) technology in recent years has led to novel mechanical FD solutions. These include methods such as support vector machines (SVM) and convolutional neural networks (CNN)3. However, These methods are costly to train the model and need large amount of labeled data. Fault data are hard to come by and dispersed unevenly in actual industrial settings. Mechanical FD widely uses optimization algorithms to solve the above problems4. Due to its excellent global search capabilities and quick convergence speed, the sparrow search algorithm (SSA) is frequently employed in FD5. An FD method for nuclear power plants grounded on SSA optimized CNN has been offered by Zhang C et al. The method extracted features using CNN, fed feature data into long and short memory networks to mine temporal features, and optimized the network parameters by SSA. It was found that the fault recognition accuracy could reach 98.24%6. An integrated approach for rolling bearing FD based on SSA improvement was presented by Wang M et al. Following the optimization of the variational mode decomposition (VMD) parameters using SSA, the recovered features were subsequently input into SVM for fault classification. According to the research results, this technology successfully improved the accuracy of rolling bearing FD7. Wang B et al. developed a low-speed bearing FD method grounded on SSA to address the problems of substantial environmental noise and poor fault information at low speeds. It was found that the method could effectively reduce noise interference and accurately extract fault features8. Hu L et al. proposed a transformer fault classification model based on the probabilistic neural network optimized by SSA. The model established a mapping model between gas ratio and fault type by installing a smart gas sensor in the transformer. Experimental results showed that the prediction accuracy reached 90%9. Ding C et al. proposed an FD method for oil-immersed transformers based on improved SSA optimized SVM. The method extracted feature data through dissolved gas analysis and optimized the parameters of SVM using improved SSA. The findings demonstrated that the diagnosis accuracy was 98.1013%, a significant improvement over the conventional approach10. Recently, researchers have explored advanced intelligent techniques to further enhance the performance of mechanical fault diagnosis under complex industrial conditions. Wen Z et al. proposed a fine-grained visual-text feature alignment framework with prompt tuning for domain-adaptive object detection, demonstrating the potential of cross-modal feature learning in improving diagnostic generalization across different working conditions11. In the context of industrial defect detection, Wen Z et al. developed a triple semantic-aware knowledge distillation network that effectively transfers multi-level semantic information, achieving high-precision defect recognition with limited labeled samples12. Additionally, Zhang G et al. introduced a motor current signal-based fault diagnosis method for harmonic drives in industrial robots under time-varying speed conditions, highlighting the importance of adaptive feature extraction in non-stationary operational environments13. These studies underscore the growing trend toward integrating semantic understanding, knowledge transfer, and adaptive signal processing into mechanical fault diagnosis frameworks, providing valuable insights for developing more robust and generalizable diagnostic models. In addition, the rapid development of ML technology has brought new opportunities and breakthroughs in the field of mechanical FD. Cen J et al. explored a data-driven mechanical FD approach based on ML algorithms. It was pointed out that ML algorithms could effectively improve the reliability of equipment in different industrial production scenarios, but they were deficient in dealing with complex faults or data imbalance problems14. The ML algorithms for mechanical FD and prediction in actual industrial manufacturing scenarios were investigated by Fernandes M et al. The study pointed out that the complexity of manufacturing systems and production processes were important challenges for fault detection and prediction15. Brusa E et al. investigated the interpretable artificial intelligence contributed by features in ML models in industrial equipment condition monitoring. The methodology analyzed how ML models identify and utilize features to diagnose mechanical faults. The research results indicated that a better understanding of feature contributions improved the reliability and transparency of diagnostic models16. An ML-based FD technique for multiphase drive systems was presented by Liu Z et al. To increase the precision of fault feature extraction, the adaptive secondary sampling methodology was used to optimize the data sampling procedure. The findings demonstrated the strong robustness, high accuracy, and remarkable generalization capabilities17. Ng EYK et al. examined how well different machine learning techniques performed in both theoretical underpinnings and real-world applications. Although ML methods demonstrated great potential in wind turbine FD, they still faced many challenges, such as data imbalance, feature selection, and model adaptation18. Despite the progress of improved SSA variants for FD, many existing approaches (e.g., Chaos-SSA, Lévy-SSA, and adaptive SSA) typically enhance SSA by introducing a single strategy (or loosely coupled strategies) without explicitly binding each strategy to a specific stage of the SSA workflow or tailoring it to the downstream classifier optimization task. In this work, the study proposes a multi-strategy improved SSA (MSISSA), whose key distinction lies in a stage-wise, task-driven integration: (1) Segmented linear chaotic mapping is used only at the initialization stage to increase population diversity and improve the coverage of the SVM hyperparameter search space; (2) Lévy-flight perturbation is embedded during iterative position updates to introduce occasional long-range exploration and reduce the risk of stagnation in local optima; and (3) an adaptive step-size control is adopted to balance exploration (early iterations) and exploitation (late iterations) within a unified update routine for SVM parameter optimization. Moreover, the proposed dual-channel CNN is not merely a parallel processing of conventional time-frequency features. Instead, it learns complementary representations from time-domain and frequency-domain inputs and performs feature-level fusion to generate a unified deep feature embedding, while the final decision is made by the MSISSA-optimized SVM. This hybrid design forms a “deep feature learning + optimized margin-based classification” pipeline, providing an explicit and reproducible workflow for improving diagnostic robustness under complex and nonlinear industrial data. Methods and materials This section first constructs a mechanical fault classification model that incorporates SVM with multi-strategy improved SSA. The study additionally suggests an mechanical FD approach based on SVM-MSISSA with two-channel CNN to further increase the precision and effectiveness of FD. Fusion of SVM and multi-strategy improved SSA for mechanical fault classification model construction There are various types and manifestations of mechanical failures. As an effective classification algorithm, SVM can effectively classify sample data with different failure modes by finding the optimal hyperplane. However, the parameters, including the kernel function (KF) and penalty factor, affects how well SVM performs. Overfitting of the model might result from unreasonable parameter values, which compromises fault classification’s accuracy and dependability19. Therefore, the study improves the SSA to construct an method that fuses SVM and multi-strategy improved sparrow search algorithm (SVM-MSISSA) for mechanical fault classification to accurately classify mechanical faults. To improve the performance of SVM in mechanical fault classification, MSISSA adjusts the parameters of the model. Before mechanical FD, the study first uses wavelet packet (WP) transform method to WP decomposition of the acquired fault signal. Then, the signal is reconstructed by thresholding20. The calculation is shown in Eq. (1). In Eq. (1), is the WP coefficients of the k th node of the k th layer. h(n) denotes the coefficients of the wavelet filter. x(n) denotes the original signal value. n is the integer variable. denotes the wavelet basis function. denotes the reconstructed signal. After signal reconstruction, the signal energy and mechanical failure feature vectors are calculated. The Eq. (2) displays the expression. In Eq. (2), means the WP reconstructed signal of the j th node of the p th layer. E(p, j) denotes the signal energy. T denotes the fault feature vector. J denotes the number of nodes in layer p. The study breaks down the complex signal into numerous modal components (MC) using the VMD method to further increase the stability of signal decomposition. Moreover, the complexity of the signal is measured using the arrangement entropy21. The Eq. (3) displays the expression. In Eq. (3), x(t) denotes the input signal. uk means the k th MC. denotes the center frequency. means the Dirac function. means the balance parameter. K means the number of MCs. Pe is the alignment entropy value. m! is the total number of alignment modes. m denotes the embedding dimension. The work used the principle component analysis (PCA) method for feature dimensionality reduction of the collected signals in order to eliminate the fault redundancy information22. Figure 1 displays the schematic diagram of WP decomposition with VMD. Schematic diagram of WP decomposition and VMD. Figure 1(a) displays the process of generating different subbands of signal X by WP decomposition. Figure 1(b) graph depicts the iterative process of VMD which outputs the intrinsic mode function when the iteration is satisfied. It is investigated to find the optimal hyperplane by SVM to classify different classes of data in order to achieve linear separability23. Equation (4) displays the expression. In Eq. (4), denotes the KF. denotes the parameters of the KF. denotes the Lagrange multiplier. C denotes the penalty factor. yi is the label of the sample . In an effort to maximize the parameters of SVM, SSA is used to quickly identify the global optimal solution in challenging optimization issues. Equation (5) provides an expression for it. In Eq. (5), denotes the position update (PU) of the discoverer. and denote the PU of the joiner and the vigilant, respectively. means the position of the eth individual at the v th iteration. Q denotes the step factor. L denotes the Lévy flight vector. denotes the optimal solution. F denotes a random number. G denotes the vigilance factor. To address the susceptibility of SSA to local optima and insufficient convergence accuracy, the study attempts to increase population diversity and GSA by initializing the population using segmented linear chaotic mapping, as shown in Eq. (6)24. In Eq. (6), denotes the dynamically adjusted step factor. The greatest and lowest values of the step factor are and . T means the highest number of iterations. means the next value after chaotic mapping. is the current value. denotes the chaotic mapping parameter. The optimal classification hyperplane of SVM with chaotic mapping distribution is shown in Fig. 2. Schematic diagram of optimal classification hyperplane and chaotic mapping distribution for SVM. The ideal classification hyperplane for SVM is schematically depicted in Fig. 2(a). The red dots and blue triangles denote the two types of data, the black lines are the classification boundaries, and the arrows display the boundary spacing. The distribution of chaotic values in various dimensions is displayed in Fig. 2(b). The Lévy flight perturbation is introduced during the iteration process to further enhance the GSA, as shown in Eq. (7). In Eq. (7), denotes the Lévy flight step vector. and denote standard normally distributed random numbers. denotes the parameters of Lévy distribution. Unlike Chaos-SSA that mainly replaces random initialization with a chaotic sequence, Lévy-SSA that only injects Lévy flight as a perturbation, or adaptive SSA that focuses on step-size scheduling, the proposed MSISSA explicitly binds (1) chaotic mapping to population initialization for diversity control, (2) Lévy flight to iterative updates for exploration enhancement, and (3) adaptive step-size to global-local search balancing within a single optimization pipeline dedicated to SVM hyperparameter tuning. Therefore, MSISSA should be viewed as a stage-wise integrated and task-oriented framework rather than a simple engineering concatenation of common heuristics. Finally, the study identifies the mechanical failure types by SVM-MSISSA. The process is schematically shown in Fig. 3. Schematic diagram of mechanical fault classification model based on SVM-MSSSA. Fig. 3 shows the flowchart of SVM-MSSSA-based mechanical fault classification model. The flow includes signal acquisition, data preprocessing, PCA dimensionality reduction, and SVM-MSSSA fault classification model. After parameter initialization, random numbers are generated by chaotic mapping and Lévy flight strategy to determine whether the iteration is sufficient. If not, adaptive distribution is performed and the optimal solution is retained until the condition is satisfied. Finally, the fault classification results are output. To avoid ambiguity in the implementation of the MSISSA, this study explicitly integrates different enhancement strategies into distinct stages of the original SSA framework. Specifically, segmented linear chaotic mapping is introduced at the population initialization stage to increase population diversity and improve the coverage of the search space. Compared with random initialization, chaotic mapping makes the distribution of candidate solutions more uniform and reduces premature convergence. During the iterative optimization stage, Lévy flight perturbation is embedded into the position update of sparrows to enhance global exploration capability. By introducing occasional long-distance jumps, Lévy flight helps the algorithm escape local optima and improves its global search ability in high-dimensional parameter spaces. In addition, an adaptive step-size mechanism is adopted to dynamically balance exploration and exploitation during iterations. Larger step sizes are applied in early iterations to encourage global search, while smaller step sizes are used in later iterations to refine local optimization near promising solutions. Through the coordinated integration of chaotic initialization, Lévy flight perturbation, and adaptive step-size control, the proposed MSISSA forms a unified optimization framework that enhances convergence stability and global search performance without increasing algorithmic complexity. The pseudocode is shown in Table 1. Table 1Pseudocode of the MSISSA. Pseudocode of the MSISSA. Mechanical FD method based on SVM-MSISSA with two-channel CNN Although the SVM-MSISSA mechanical fault classification method is able to effectively identify and classify mechanical fault types, with the increase in the complexity of mechanical systems, the method is unable to effectively capture the intrinsic characteristics of faults, which leads to misdiagnosis. Therefore, the research suggests an mechanical FD approach based on SVM-MSISSA with dual-channel CNN to further increase the precision and resilience of mechanical FD. This technology uses dual-channel CNN for deep learning of mechanical defect features, improving the accuracy and efficiency of mechanical FD while reducing reliance on large amounts of labeled data. The dual-channel CNN consists of two independent convolutional channels for processing TD signals and FD signals, respectively. This dual-channel structure can fully utilize the TD and FD information to improve the extraction of fault features. The dual-channel CNN is designed to learn complementary fault representations from time-domain and frequency-domain inputs and to form a fused deep embedding via feature-level fusion. It acts as a deep representation learner, whose fused feature vectors are subsequently classified by the MSISSA-optimized SVM. Such a two-stage design (deep embedding followed by an optimized margin-based classifier) explicitly separates feature representation learning from decision boundary optimization, which is beneficial for improving robustness in industrial FD scenarios with limited or unevenly distributed labeled data. Each channel contains three convolutional layers with 32, 64, and 128 filters, respectively, using a kernel size of 3 × 1 and ReLU activation. Each convolutional layer is followed by a max-pooling layer with a pooling size of 2 × 1 to reduce feature dimensionality and enhance robustness. After the final pooling operation, feature maps from both channels are flattened and fused through feature-level concatenation, forming a unified representation without introducing additional trainable parameters. The fused features are passed through a fully connected layer with 256 neurons and a Dropout layer with a dropout rate of 0.5. During training, a Softmax layer together with the cross-entropy loss function is employed only to guide feature discrimination within the CNN, while the final fault classification is performed by the SVM-MSISSA model using the extracted deep feature vectors. The Adam optimizer is used with an initial learning rate of 0.001, a batch size of 64, and a maximum of 150 training epochs. Batch normalization is applied after convolutional layers to stabilize training and improve convergence. For TD signals, the CNN structure of its channel includes a convolutional layer (CL), a pooling layer (PL), and a fully connected layer (FCL). The CL is used to extract local features, the PL is used to reduce the feature dimension and enhance the robustness of the features, and the FCL is used to classify the extracted features. The classifier used is Softmax25. The output of CL and PL is shown in Eq. (8). In Eq. (8), f means the activation function. is the output of the CL. denotes the convolutional kernel weights. S means the output signal. denotes the bias term (BT). means the output of the PL. The expressions for the FCL and Softmax classifier are shown in Eq. (9). In Eq. (9), is the FCL’s output. f means the activation function. and mean the weights and BTs of the FCL. is the category probability. means the value of the ith element in the output vector. is the exponential function used to convert to a positive value. A is the total number of categories. Figure 4 roughly depicts the two-channel CNN’s structure. Diagrammatic representation of the dual channel structure CNN. Figure 4 contains two channels, each with a CL, a PL, and a flat layer. The channel output passes through the feature fusion layer, then through the FCL and Dropout layer. Finally, to enhance classification performance, it outputs the model classification results at the classification layer26. Equation (10) illustrates how the study measures the model’s prediction performance using the cross-entropy loss function. In Eq. (10), denotes the cross-entropy loss value. and denote the true and predicted values, respectively. N denotes the number of samples. To improve the training efficiency of the model, the study stabilizes the distribution of each feature by batch normalization, as shown in Eq. (11). In Eq. (11), denotes the result after batch normalization. denotes the dth dimension of the input. denotes the result after scaling and panning. and mean the mean and variance of the sample set. denotes the scaling factor. denotes the bias. Equation (12) illustrates how the study uses the stochastic deactivation strategy to produce the regularization effect to avoid overfitting. In Eq. (12), the pre-activation output of the u-th neuron in the l+1-th layer, and represents the output vector of the previous layer. The function g denotes the linear transformation, while f represents the activation function. and denote the corresponding weight vector and bias term, respectively. Equation (12) describes the stochastic deactivation mechanism applied during training, which introduces randomness into neuron outputs and produces a regularization effect to alleviate overfitting. Equation (13) illustrates how the study extracts the local features of the frequency domain signal using the continuous wavelet transform. In Eq. (13), denotes the continuous wavelet transform coefficient of the signal x(t) at scale u and translation v. t represents the time index. denotes the mother wavelet function, and represents the corresponding scaled and translated sub-wavelet basis. The parameters u and v control the scale and displacement, respectively, and serves as a normalization factor. After comprehensively capturing the mechanical fault signal features by two-channel CNN, the study utilizes SVM-MSISSA for further FD. The schematic diagram of mechanical FD based on SVM-MSISSA with dual-channel CNN is shown in Fig. 5. Schematic diagram of mechanical FD based on SVM-MSSSA and dual channel CNN. In Fig. 5, the FD method starts with signal acquisition and dataset partitioning. Then, the model parameters are initialized and backpropagation weights are adjusted. Chaos mapping and L é vy flight strategy generate random numbers for adaptive distribution and optimal solution preservation until sufficient iterations are made, and finally output FD results. Results The performance of the built mechanical fault classification model is initially verified in this part. Then, the effectiveness of the proposed mechanical FD method is analyzed. Fusion of SVM and MSISSA for mechanical fault classification performance testing Experiments are carried out utilizing the Case Western Reserve University (CWRU) bearing fault dataset and the Prognostics and Health Management (PHM) dataset to illustrate the effects of the suggested mechanical fault classification model. The CWRU bearing failure dataset contains bearing failure data under different operating conditions. The PHM dataset contains operation data and fault data of various mechanical equipment (The data of the test and training set is 3:7). The CWRU dataset can be accessed at: https://engineering.case.edu/bearingdatacenter. The PHM dataset can be accessed at: https://www.nasa.gov/content/prognostics-center-of-excellence-data-set-repository. For both the CWRU and PHM datasets, raw vibration signals were first subjected to normalization to eliminate scale differences among samples. The continuous signals were then segmented into fixed-length samples using a sliding window strategy without overlap to ensure sample independence. For each segment, wavelet packet decomposition was applied for noise reduction and feature enhancement, followed by variational mode decomposition to extract modal components. Feature vectors were subsequently constructed based on signal energy and entropy-related measures and further reduced in dimensionality using principal component analysis. The same preprocessing and partitioning strategy was consistently applied to all comparative methods to ensure fair evaluation and reproducibility. Table 2 displays the parameters and experimental conditions. Table 2Parameter configurations and the experimental environment.Experimental environmentParameter settingsHardware environmentProcessorIntel Core i7-10700 K @ 3.80 GHzPenalty factor[0.1,100]Graphics cardNVIDIA GeForce RTX 3080Kernel function parameters[0.001,1]Memory32GB DDR4Population size50Storage1 TB SSDMaximum number of iterations250Software environmentOperating systemWindows 10Chaos mapping parameters0.5Programming languagePython 3.8Lévy flight parameters1.5Deep learning frameworkTensorFlow 2.5.0Step size factor (maximum value)1.0Signal processing toolsPyWavelets 1.1.1Step size factor (minimum value)0.1 Parameter configurations and the experimental environment. The experiment initially compares the loss of the designed model with the existing popular mechanical fault classification models in an attempt to identify the effect of the raised method. The comparison methods include traditional SVM3, SVM-sparrow search algorithm (SVM-SSA)7, SVM-particle swarm optimization (SVM-PSO)27, and SVM-genetic algorithm (SVM-GA)28. In Fig. 6(a), in the CWRU dataset, when the iteration is 50, the losses of the SVM and SVM-PSO models are 1.05 and 0.87. The losses of the SVM-SSA, SVM-GA, and SVM-MSISSA models are 0.58, 0.49, and 0.36. The losses of the proposed models tend to be stable when the number of iterations continues to increase. In Fig. 6(b), in the PHM dataset, the losses of the SVM, SVM-PSO, SVM-SSA, and SVM-GA are 1.35, 1.17, 0.89, and 0.52, respectively, when the iteration is 50. In comparison, the loss of the SVM-MSISSA model is 0.43, which is a reduction of 68.15%, 63.25%, 51.69%, and 17.31%. The outcomes demonstrate the higher convergence performance and stability of the suggested model. Comparison of loss between different fault classification models in two datasets. To further validate the superiority of SVM-MSISSA, the study analyzes the fault classification accuracy of this model and compares it with existing models. In Fig. 7(a), in the CWRU dataset, when the iteration is 100, the accuracy of SVM, SVM-PSO, and SVM-SSA models are 81.03%, 82.36%, and 87.64%. The accuracy of SVM-GA and SVM-MSISSA models are 89.25% and 92.30%. When the iteration reaches 70, the accuracy of the suggested model stabilizes in comparison to the other four models, improving by 11.27%, 9.94%, 4.66%, and 3.05%, respectively. In Fig. 7(b), the accuracy of SVM, SVM-PSO, SVM-SSA, and SVM-GA models are 77.62%, 81.35%, 84.59%, and 88.24% in the PHM dataset after 100 iterations. The accuracy of the SVM-MOSSA model is 91.52%, with improvements of 13.90%, 10.17%, 6.93%, and 3.28%, respectively. The findings demonstrate the effectiveness of the study’s suggested model in identifying and categorizing various fault kinds. Accuracy comparison of two datasets using several fault classification models. The experiments are conducted to compare and analyze the fault recognition accuracy and running time (RT) of the proposed model before and after PCA dimensionality reduction. In Fig. 8(a), the fault recognition accuracy of SVM, SVM-PSO, and SVM-SSA models before PCA dimensionality reduction is 78.63%, 83.72%, and 86.24%, respectively. The fault recognition accuracy after dimensionality reduction is 80.53%, 85.67%, and 87.79%, respectively. The fault recognition accuracy of SVM-GA and SVM-MSISSA models before PCA dimensionality reduction is 82.34% and 88.25%, respectively. The recognition accuracy after dimensionality reduction is 84.36% and 93.72%, respectively. In Fig. 8(b), the RT of SVM, SVM-PSO, SVM-SSA, and SVM-GA models before PCA dimensionality reduction is 93.6s, 82.7s, 91.2s, and 87.3s, respectively. The RT after dimensionality reduction is 69.3s, 50.7s, 45.6s, and 46.3s, respectively. Compared with the SVM-MSISSA model the RT before and after PCA dimensionality reduction is 85.2s and 20.3s, respectively. The research results indicate that the proposed model is significantly superior to other comparative versions in mechanical fault recognition accuracy and operation duration. The fault recognition accuracy and RT of different fault classification models before and after PCA dimensionality reduction. Effectiveness analysis of mechanical FD methods To further evaluate the practical applicability of the proposed method, an industrial FD experiment was conducted using vibration data collected from motor-driven equipment in a large-scale production plant. The dataset includes four typical fault types, namely normal condition, bearing wear, rotor imbalance, and misalignment, which were commonly encountered in industrial motor systems. Fault labels were obtained based on maintenance records and expert inspection reports provided by on-site engineers. A total of 1,200 signal samples were collected under stable operating conditions, with each sample generated using the same segmentation and preprocessing procedures described in Sect. 3.1. The dataset was processed and evaluated using the identical feature extraction, training, and testing pipeline to ensure consistency with the previous experiments. To demonstrate the impact of the proposed mechanical FD method, the study looks at the FD loss values and accuracy of the model test set and compares them with those of other methods. The comparison methodologies are the CNN-based FD method4 and the long short-term memory and recurrent neural network (LSTM-RNN) FD method29. In Fig. 9(a), when the iteration is 50, the accuracy of CNN, LSTM-RNN, and the proposed methods are 38.27%, 80.05%, and 92.36%, respectively. When the iteration reaches 150, the accuracy of the three mechanical FD methods are 79.62%, 84.78%, and 94.32%, respectively. In Fig. 9(b), when the iteration is 50, the losses of CNN, LSTM-RNN, and the proposed methods are 1.021, 0.387, and 0.189, respectively. When the iteration reaches 150, the losses of the three methods are 0.465, 0.267, and 0.184. The findings demonstrate the higher performance of the proposed FD method. Comparison of precision and loss of different mechanical FD methods. The trials are further examined for F1 score (F1), recall, and precision of the suggested mechanical FD approach using the CWRU and PHM datasets. The suggested strategy is also contrasted with the current approaches to confirm its efficacy. In Fig. 10(a), the CNN method in the CWRU dataset are 0.842, 0.836, and 0.879. The three indices of the LSTM-RNN are 0.896, 0.918, and 0.905. In comparison, the precision of the proposed method is 0.943, which is improved by 10.71% and 4.98%, respectively. The recall is 0.949 and F1 is 0.958. In Fig. 10(b), the precision of CNN, LSTM-RNN, and the proposed method in the PHM dataset is 0.847, 0.908, and 0.950, respectively. The recall is 0.825, 0.902, and 0.966, respectively. The F1 is 0.972, 0.889, and 0.938. Compared with the other two methods, the precision of the proposed method has increased by 10.84% and 4.42%, respectively. The findings reveal that the recommended mechanical FD technique performs better. Comparing the F1, precision, and R of several approaches using the CWRU and PHM datasets. The study is conducted at a large production facility to identify the issues with numerous motor equipment pieces to validate the effectiveness of the recommended approach for real-world application. The motors have a rated power of 1000 kW and a rated speed of 1500 rpm. The FD accuracy and RT of mechanical equipment are experimentally analyzed. In Fig. 11(a), the average FD accuracy of CNN and LSTM-RNN is 85.24% and 89.75%, respectively. The average FD accuracy of the proposed method is 95.05%, which is an improvement of 9.81% and 5.30%, respectively. In Fig. 11(b), the FD RT of CNN, LSTM-RNN, and the proposed method are 81.36s, 53.62s, and 18.75s, respectively. Compared with them, the FD RT of the proposed method is shortened by 76.95% and 65.03%, respectively. The outcomes demonstrate how efficiently and successfully the suggested approach may identify mechanical issues. Comparison of actual accuracy of diagnosis and RT of different FD methods. Table 3 shows the performance comparison results with recent state-of-the-art methods after multiple random runs. In Table 3, the proposed method consistently outperformed the two recent advanced methods under multiple random runs. Compared with the Transformer-based FD approach, the proposed method improved the average accuracy from 92.38% to 95.01%, representing an absolute gain of 2.63% points. In addition, the standard deviation of accuracy was reduced from 1.08 to 0.68, indicating improved stability and robustness across different random initializations. When compared with the hybrid CNN-Transformer method, the proposed approach also achieved superior performance, with an accuracy improvement of 1.29% points and a smaller variance. This suggested that combining deep feature extraction with MSISSA-optimized SVM classification provided more stable decision boundaries than end-to-end deep architectures alone. From the perspective of classification quality, the proposed method attained the highest precision (0.95), recall (0.95), and F1-score (0.96) among all compared approaches. The balanced improvement in recall and F1-score indicated that the method not only enhanced overall accuracy but also effectively reduced misclassification across different fault categories. Overall, the results demonstrated that the proposed method achieved both higher diagnostic performance and better robustness under repeated random experiments. Table 3Performance comparison with recent advanced methods under multiple random runs.MethodsRunsAccuracy (%)PrecisionRecallF1Transformer-based FD30191.050.910.910.91293.800.950.930.91392.450.930.920.91491.500.920.910.91593.100.940.930.91Average ± SD92.38 ± 1.080.93 ± 0.020.92 ± 0.010.91 ± 0.00Hybrid CNN–Transformer31192.650.910.930.91294.850.970.950.95393.900.940.930.93493.000.920.920.92592.400.960.940.94Average ± SD93.72 ± 0.920.94 ± 0.030.94 ± 0.010.93 ± 0.02Method of this study194.300.930.940.95295.850.970.960.97395.100.950.950.96494.500.940.940.95595.300.960.960.96Average ± SD95.01 ± 0.680.95 ± 0.020.95 ± 0.010.96 ± 0.01 Performance comparison with recent advanced methods under multiple random runs. To comprehensively evaluate the classification performance and intuitively display the misclassification among different operating states, the confusion matrices for the three methods are presented in Fig. 12. In the confusion matrix of the Transformer-based FD method, although the diagonal cells exhibit a relatively high correct recognition rate, notable misclassifications are observed in the off-diagonal cells, which appear as light blue blocks. This indicates that the pure transformer architecture has limitations in distinguishing mechanical similar fault features under limited data. For the Hybrid CNN-Transformer method, the correct classification numbers on the diagonal are improved, accompanied by a visible deepening of the blue color. The off-diagonal misclassifications are reduced, demonstrating that the convolutional layer enhances local feature extraction and mitigates inter-class confusion. In contrast, the confusion matrix of the proposed method displays the darkest blue blocks strictly along the diagonal, achieving the highest correct classification sample counts across all categories. Correspondingly, the off-diagonal cells are the palest, with misclassification numbers being significantly minimized. This visual and numerical comparison explicitly demonstrates that the proposed method extracts more discriminative features and establishes the most robust decision boundaries, effectively overcoming the confusion prevalent in the comparative methods. Comparison of confusion matrices for different methods. Discussion and conclusion A multi-strategy improved SSA-based mechanical FD approach was suggested by the study. Additionally, it worked in tandem with SVM and dual-channel CNN to enhance FD’s precision and effectiveness. The results indicated that SVM-MSISSA had lower loss values, higher accuracy, and faster convergence in CWRU and PHM datasets. Especially in the PHM dataset, the loss value of SVM-MSISSA was significantly lower than that of the other methods, and the accuracy was improved from 3.05% to 13.90%. In addition, the dual-channel CNN structure could process both TD signals and FD signals, which fully utilized the time-frequency domain information and further improved the effect of fault feature extraction. The experimental outcomes demonstrated that the precision, recall, and F1 of the FD method based on SVM-MSISSA with dual-channel CNN were better than those of the FD method based on traditional CNN and LSTM-RNN on both CWRU and PHM datasets. Practical application tests demonstrated that the proposed method performed well in FD of motor equipment in a large manufacturing plant. The average FD accuracy was improved by 9.81% and 5.30% compared with CNN and LSTM-RNN, respectively. The RT was shortened by 76.95% and 65.03%, respectively. This indicated that the method not only had advantages in theory and experiment, but also had good potential for practical application. The results indicate that the proposed mechanical FD method performs very well in many publicly available datasets and real-world application environments, with high accuracy and short RT. Although this study has the potential to optimize algorithms and reduce reliance on labeled data, obtaining a large amount of high-quality defect labeled data in actual industrial environments remains challenging. Future research can develop data augmentation strategies that use a large amount of unlabeled data and limited labeled data during the training process to improve the robustness and generalization ability of the model. Author contributions Xiangtao Kong wrote the main manuscript text. Xiangtao Kong reviewed the manuscript. Data availability All data generated or analysed during this study are included in this published article. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement All data generated or analysed during this study are included in this published article.