Intelligent fault diagnosis method for harmonic reducer based on multi-source information fusion

Abstract

The invention discloses a harmonic reducer intelligent fault diagnosis method based on multi-source information fusion, and the method comprises the steps: synchronously collecting four heterogeneous physical signals, namely vibration, acoustic emission, temperature and sound, and carrying out the modal specificity preprocessing of the four heterogeneous physical signals based on the physical mechanism of a typical fault of a harmonic reducer; according to the method, a self-adaptive multi-branch fusion network model is constructed, the model comprises a CNN local feature extraction branch and a Transform global dependence modeling branch which are parallel, a cross-modal, channel and sensor level triple attention mechanism is innovatively introduced, and fusion weight distribution dynamically matched with a fault state is realized. An adaptive network architecture is designed, and a fusion strategy and network complexity can be dynamically adjusted according to the number of online sensors and working conditions. The model is trained through a multi-objective loss function, and high-precision and high-robustness diagnosis of the health state and various early faults of the harmonic reducer is realized. The problems that a single signal source is poor in reliability and a traditional fusion method is poor in physical interpretability are solved.

Description

Technical Field

[0001] This invention relates to an intelligent fault diagnosis method for harmonic reducers, and more particularly to an intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion. Background Technology

[0002] Wave reducers are core transmission components of industrial robot joints, and their reliability directly affects the overall performance and safety of the robot. Operating under complex conditions such as heavy loads and high speeds, they are prone to early failures such as flexural cracks and bearing pitting. These failures are subtle and exhibit multi-physics coupling effects, making diagnosis difficult.

[0003] Existing fault diagnosis methods for harmonic reducers mainly follow three directions: first, feature enhancement methods based on signal processing, such as optimized variational mode decomposition and wavelet transform; second, model optimization methods based on deep learning, such as convolutional neural networks and residual networks; and third, multi-sensor information fusion methods. However, current research still has significant shortcomings: First, most methods rely on a single vibration signal, which limits their diagnostic performance under strong noise and variable operating conditions; second, even multi-sensor fusion research mostly focuses on data fusion from similar sensors (such as multiple vibration sensors), failing to fully utilize the natural complementary advantages of heterogeneous signals with different physical principles, such as vibration, acoustic emission, temperature, and sound, in characterizing different fault mechanisms; finally, existing fusion methods are mostly fixed-weight feature splicing or simple decision fusion, lacking an interpretable fusion mechanism that dynamically matches the physical state of the fault.

[0004] Therefore, developing an intelligent diagnostic method that can fully leverage the advantages of multi-source heterogeneous information, integrate strategies and fault mechanisms, and possess strong robustness is crucial for achieving precise operation and maintenance of harmonic reducers. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion. This method starts from the physical mechanism of the fault, constructs a heterogeneous signal optimization and fusion paradigm, and achieves intelligent diagnosis by dynamically matching the fault state through an innovative deep learning model, significantly improving diagnostic accuracy, robustness, and interpretability.

[0006] The technical solution of this invention: A method for intelligent fault diagnosis of harmonic reducers based on multi-source information fusion, comprising the following steps:

[0007] S1, based on the physical mechanism of typical faults of harmonic reducers, simultaneously collects four heterogeneous physical signals: vibration, acoustic emission, temperature and sound, and performs modal-specific preprocessing on each;

[0008] S2, construct an adaptive multi-branch fusion network model, which includes a parallel CNN local feature extraction branch and a Transformer global dependency modeling branch, and introduces a cross-modal, channel and sensor-level triple attention mechanism to achieve dynamic matching of fusion weight allocation with fault state;

[0009] S3, design an adaptive network architecture that dynamically adjusts the fusion strategy and network complexity based on the number of online sensors and operating conditions;

[0010] S4, the model is trained using a multi-objective loss function;

[0011] S5 performs online fault diagnosis.

[0012] In the aforementioned intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion, S1 specifically involves analyzing the excitation mechanism of harmonic reducer faults, clarifying the advantages of four physical signals—vibration, acoustic emission, temperature, and sound—in characterizing different faults, and building a multi-sensor synchronous acquisition platform accordingly.

[0013] In the aforementioned intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion, S2 specifically involves designing targeted preprocessing algorithms for the heterogeneity of four types of signals and the main noise sources: performing order tracking and resonant frequency band energy ratio extraction on vibration signals; performing wavelet soft threshold denoising on acoustic emission signals; performing environmental compensation and moving average filtering on temperature signals; performing echo cancellation on sound signals and generating time-spectrum diagrams; and finally, extracting mean, variance, kurtosis, and skewness statistical features from the preprocessed signals to form a multi-source heterogeneous feature set.

[0014] In the aforementioned intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion, the adaptive multi-branch fusion network model includes two parallel branches: one is a CNN branch composed of multiple one-dimensional convolutions, used to extract local fault features; the other is a global dependency modeling branch composed of a Transformer encoder. The output features of the two branches are dynamically fused through a triple attention fusion module, which sequentially includes:

[0015] (1) Cross-modal attention promotes the interaction between local details of CNN and global context information of Transformer;

[0016] (2) Channel attention, adaptively emphasizing important feature channels;

[0017] (3) Sensor-level attention, dynamically assigning decision weights to different physical sensors based on the current input;

[0018] Furthermore, the adaptive multi-branch fusion network model employs an adaptive architecture design, automatically selecting different configurations from simple CNNs to complex multi-attention fusion networks based on the number of input sensors, ensuring optimal resource utilization and the best balance between performance.

[0019] In the aforementioned intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion, S4 specifically involves using a multi-objective loss function to supervise the model. This multi-objective loss function integrates the cross-entropy loss of the main classifier and the auxiliary classification loss of the two parallel branches to promote collaborative learning and feature consistency between branches. The training uses the AdamW optimizer and a cosine annealing learning rate scheduling strategy.

[0020] In the aforementioned intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion, step S5 specifically involves preprocessing and feature constructing the multi-source signals of the harmonic reducer to be diagnosed, and then inputting them into a trained adaptive fusion diagnosis model to obtain the diagnosis results of its health status or fault category.

[0021] In the aforementioned intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion, the vibration signal is determined according to the formula:

[0022] ,

[0023] Order tracking based on rotational speed synchronization is performed to eliminate the influence of rotational speed fluctuations, and the frequency band energy ratio is calculated to focus on the structural resonance frequency band, thus separating the structural resonance frequency band.

[0024] In the formula The frequency band energy ratio, The distribution of signal power spectral density in the target frequency band. The total power spectral density across the entire frequency band. , These represent the upper and lower limits of the structural resonance frequency band.

[0025] In the aforementioned intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion, the acoustic emission signal is determined according to the formula:

[0026] ,

[0027] An improved wavelet soft thresholding denoising algorithm is used to effectively suppress high-frequency electromagnetic interference and highlight damage-related sudden events;

[0028] In the formula These are the denoised wavelet coefficients. These are the original wavelet transform coefficients. This is the threshold parameter.

[0029] In the aforementioned intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion, the temperature signal is determined according to the formula:

[0030] ,

[0031] Ambient temperature compensation is performed, and then a smooth temperature rise curve is obtained by moving average filtering.

[0032] The moving average filter window width is w=10s. To compensate for the temperature value, The original temperature sensor measurement value. This is the measured value of ambient temperature.

[0033] In the aforementioned intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion, the sound signal is determined according to the formula:

[0034] ,

[0035] Adaptive echo cancellation is used to suppress background noise, and a time-frequency diagram is generated through short-time Fourier transform;

[0036] In the formula is the window function for the short-time Fourier transform, and N is the window width.

[0037] The beneficial effects of the present invention: Compared with the prior art, the method of the present invention has the following advantages:

[0038] Mechanism-oriented and highly complementary: Starting from the physical essence of the fault, four heterogeneous signals, namely vibration, acoustic emission, temperature and sound, are selected to ensure the complementarity of multi-source information in characterizing different fault dimensions such as cracks, impacts and temperature rises from the source, thus laying a data foundation for high-precision diagnosis.

[0039] Dynamic adaptive fusion with high interpretability: The proposed triple attention mechanism (cross-modal, channel, and sensor level) enables the model to adaptively focus on the most sensitive physical signals and key features of the current fault based on real-time input signals, realizing dynamic matching between fusion weights and the physical state of the fault, and enhancing the physical interpretability of the diagnostic process.

[0040] Adaptive architecture with good engineering applicability: The designed adaptive network architecture can dynamically adjust the model complexity and fusion strategy according to the actual number of sensors installed. It can still maintain excellent performance on equipment with some sensor failures or different configurations, thus improving the engineering applicability and reliability of the method.

[0041] Superior diagnostic performance and outstanding robustness: Experiments show that the average diagnostic accuracy of the method proposed in this invention for typical faults of harmonic reducers is significantly higher than that of the optimal single-sensor method and the traditional fusion method, and it also exhibits excellent generalization ability and robustness under variable load and strong noise interference. Attached Figure Description

[0042] Figure 1 This is a flowchart of the harmonic reducer fault diagnosis method provided in the embodiments of the present invention;

[0043] Figure 2 This is a physical image of a harmonic reducer fault simulation test bench, showing the drive motor, sensor installation positions, the reducer under test, and the loading device.

[0044] Figure 3 It is a time-frequency characteristic spectrum of the vibration signal of the harmonic reducer under different loads (such as short-time Fourier transform results), used to show the load dependence and fault impact characteristics of the vibration signal;

[0045] Figure 4 It is a damage evolution characteristic diagram (such as waveform diagram or parameter diagram) of the acoustic emission signal of the harmonic reducer under different loads, used to show the sensitivity of the acoustic emission signal to early damage;

[0046] Figure 5 It is a temperature field distribution cloud map of key components (flexible bearing, flex wheel, rigid wheel) of the harmonic reducer under multiple operating conditions (0Nm, 2Nm, 4Nm), used to illustrate the characteristics and effectiveness of the temperature signal;

[0047] Figure 6 It is a nonlinear response spectrum (such as a spectrum diagram) of the acoustic signal of the harmonic reducer under different loads, used to demonstrate the complementarity of the acoustic signals;

[0048] Figure 7 This is a schematic diagram of a four-sensor multi-attention fusion network architecture in an embodiment of the present invention;

[0049] Figure 8 This is a visual comparison of features after t-SNE dimensionality reduction under no-load and full-load conditions for different sensor combinations (single sensor vs. four sensors), used to demonstrate the advantages of multi-source fusion features;

[0050] Figure 9 This is a comparison chart of the confusion matrix between the single-sensor diagnostic model and the four-sensor fusion diagnostic model, which intuitively demonstrates the improvement effect of the fusion strategy on diagnostic accuracy. Detailed Implementation

[0051] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0052] Embodiments of the present invention: This embodiment takes the diagnosis of four states of a harmonic reducer: normal state, flexural wheel crack, bearing pitting, and rigid wheel wear as examples, but the scope of protection of the present invention is not limited thereto.

[0053] S1: Fault Experiment Design and Multi-Source Heterogeneous Data Acquisition

[0054] To verify this invention, a real fault dataset must first be constructed. For example... Figure 2 As shown, a fault simulation test bench for harmonic reducers was constructed. The test bench mainly consists of a drive servo motor, an input torque sensor, the harmonic reducer under test, an output torque sensor, and a magnetic powder brake. To simulate real faults and common fault types of harmonic reducers, cracks, pitting, and wear faults were pre-fabricated on the flex wheel, rigid wheel, and flexible bearings through methods such as electrical discharge machining and fatigue loading.

[0055] Based on the structural analysis and fault mechanism analysis of the harmonic reducer, four types of sensors were simultaneously configured to achieve multi-physics sensing (specific parameters are shown in Table 1):

[0056] Table 1 Configuration of Multi-Sensor Synchronous Acquisition System

[0057] signal type model Installation location / method Sampling parameters calibration vibration PCB 356A16 Radial housing, magnetic attachment, 50mm from flange 51.2kHz Standard vibration table calibration acoustic emission PAC R15 Flexible wheel surface, coupling agent, magnetic clamp 1MHz Electrical discharge calibration temperature K-type thermocouple Flexible bearing outer ring, thermally conductive adhesive bonding 1hz Comparison calibration method sound BSW MPA416 100mm from the housing, bracket, directly opposite 20kHz Sound level calibrator calibration

[0058] All sensor signals were acquired via a synchronous data acquisition card, ensuring strict time alignment of different physical signals. Data was collected under different load conditions (0Nm, 2Nm, 4Nm, 6Nm, 8Nm, 10Nm) and health conditions, constructing a multi-source heterogeneous raw dataset covering multiple operating conditions.

[0059] S2: Modality-Specific Preprocessing and Deep Feature Visualization Analysis

[0060] For the four types of signals, specific preprocessing is performed based on their physical characteristics and noise sources, and deep feature maps are generated to verify the signal validity.

[0061] Vibration signals are based on the formula:

[0062] ,

[0063] Order tracking based on rotational speed synchronization is performed to eliminate the influence of rotational speed fluctuations, and the frequency band energy ratio is calculated to focus on the structural resonance frequency band, thus separating the structural resonance frequency band.

[0064] In the formula The frequency band energy ratio, The distribution of signal power spectral density in the target frequency band. The total power spectral density across the entire frequency band. , These are the upper and lower limits of the structural resonance frequency band; Figure 3The time-frequency characteristic spectrum reveals the load dependence of the vibration response. Under load, high-frequency energy concentration and periodic impact components are obvious, proving that the vibration signal can effectively capture the fault characteristics aggravated by dynamic excitation.

[0065] Acoustic emission signals are based on the formula:

[0066] ,

[0067] An improved wavelet soft thresholding denoising algorithm is adopted to effectively suppress high-frequency electromagnetic interference and highlight damage-related sudden events.

[0068] In the formula These are the denoised wavelet coefficients, used to eliminate high-frequency electromagnetic interference. These are the original wavelet transform coefficients. This is the threshold parameter. Figure 4 The damage evolution characteristic diagram shows that the amplitude of the AE signal increases and sudden waveforms appear under load. These characteristics correspond to the propagation of the flexible wheel crack and the friction slip of the bearing, proving that acoustic emission is extremely sensitive to early damage.

[0069] The temperature signal is based on the formula:

[0070] ,

[0071] Ambient temperature compensation is performed, followed by moving average filtering to obtain a smooth temperature rise curve. Through the ambient temperature compensation model and moving average filtering, environmental thermal disturbances are eliminated, resulting in a clean component temperature rise curve.

[0072] Moving average filtering (window width w=10s), To compensate for the temperature value, The original temperature sensor measurement value. This is the measured ambient temperature value; Figure 5 The temperature field distribution of key components of the harmonic reducer under multiple operating conditions is presented. Analysis shows that the temperature rise of the outer ring of the flexible bearing is significant, making it the most effective indicator of the system's thermal load and friction state. This verifies the effectiveness of temperature signals as indicators of slow-moving fault variables from a thermodynamic perspective.

[0073] Sound signals are based on the formula:

[0074] ,

[0075] Adaptive echo cancellation (NLMS algorithm) is used to suppress background noise, and time-frequency diagrams are generated through short-time Fourier transform.

[0076] In the formula is the window function for the short-time Fourier transform, and N is the window width. Figure 6The nonlinear response spectrum shows that the nonlinear response of the acoustic signal becomes more complex with increasing load, providing a complementary diagnostic perspective independent of shell vibration.

[0077] For the various preprocessed signals, 14 statistical features, including mean, standard deviation, kurtosis, and skewness, are extracted to form an initial 56-dimensional multi-source heterogeneous feature vector.

[0078] S3: Construction and Training of Adaptive Multi-branch Fusion Diagnostic Model

[0079] The core of this invention is as follows Figure 7 The optimal architecture (four sensors) of the adaptive multi-branch fusion network shown is explained in detail below:

[0080] Input and Feature Projection: The 56-dimensional feature vector is mapped to a 256-dimensional unified feature space through a linear projection layer.

[0081] Parallel feature extraction branch:

[0082] The CNN branch consists of four one-dimensional convolutional modules connected in series (with the number of filters being 32, 64, 128, and 256 respectively). Each layer includes batch normalization, ReLU activation, and max pooling, and finally outputs a 128-dimensional local feature vector through global average pooling.

[0083] Transformer branch: After adding positional encoding to the features, it is input into a 4-layer Transformer encoder (8-head self-attention, feedforward dimension 512), and finally outputs a 128-dimensional global dependency feature vector through global average pooling.

[0084] Triple attention fusion mechanism (key invention point):

[0085] Cross-modal attention: Using CNN branch features as queries and Transformer branch features as keys and values, attention weights are calculated to promote the complementarity of local details and global contextual information.

[0086] Channel Attention: Introduces a compression-excitation (SE) module to adaptively recalibrate the importance of feature channels.

[0087] Sensor-level attention: Dynamically evaluate and weight the contribution of four types of sensors—vibration, acoustic emission, temperature, and sound—to the current diagnostic task through a learnable parameter matrix.

[0088] Classification decision layer: The fused 256-dimensional high-level features are transformed by a 128-dimensional fully connected layer (including Dropout), and the Softmax classifier outputs the diagnostic probabilities of four health states.

[0089] The model was trained using a multi-objective loss function with the AdamW optimizer. The initial learning rate was set to 0.001, and cosine annealing was used to adjust the learning rate. An early stopping strategy was implemented (training was stopped when the validation set loss did not decrease for 20 consecutive epochs) to prevent overfitting.

[0090] S4: System Adaptive Architecture Configuration

[0091] This invention designs an adaptive architecture, the mapping relationship of which is shown in Table 2. The system can automatically select the optimal processing strategy based on the actual number of connected sensors.

[0092] Table 2 Adaptive Network Architecture Mapping Table

[0093] Number of sensors Architecture Configuration 1 Deep CNN 6-layer convolution for dense feature extraction 2 CNN + Simple Fusion 5-layer convolution, feature cascade fusion 3 Hierarchical converged network 4 convolutional layers, attention-guided fusion 4 Multiple attention fusion CNN+Transformer+Triple Attention

[0094] One sensor: Enables a deep CNN model for intensive feature extraction.

[0095] Two sensors: CNN + simple feature cascade fusion model enabled.

[0096] Three sensors: Enable hierarchical attention-guided fusion model.

[0097] Four sensors: Enable the complete CNN-Transformer triple attention fusion model (optimal configuration).

[0098] S5: Experimental Results Analysis and Comparison

[0099] To fully verify the effectiveness, superiority, and robustness of the framework proposed in this invention, systematic experiments were conducted, and the key results are as follows:

[0100] Table 3 shows the diagnostic performance of the sensor combination. Experimental results show that the four-sensor (vibration, acoustic emission, temperature, and sound) fusion scheme proposed in this invention achieves optimal diagnostic performance, with an average accuracy of 94.51%.

[0101] Table 3 Diagnostic performance under modal combinations

[0102] Modal Combination accuracy Accuracy Recall rate F1 score Acoustic emission + vibration + temperature + microphone 94.51% 94.72% 94.35% 94.52%

[0103] Model decision visualization and interpretability analysis:

[0104] Feature space visualization: Figure 8 The distribution of features after t-SNE dimensionality reduction under different sensor configurations is shown. It can be seen that when four sensors are fused, the feature clusters of different fault categories are more compact and the separation is higher, especially showing significant robustness under variable load conditions.

[0105] Confusion matrix analysis: Figure 9The confusion matrices of single-sensor and four-sensor fusion were compared. Four-sensor fusion significantly reduced the false positive rate among various fault types, indicating that the fusion strategy effectively eliminated diagnostic blind spots through information complementarity.

[0106] Model generalization and robustness testing:

[0107] Variable load test: Table 4 shows that the model maintains high accuracy at most load points (98.75% at 4Nm) within the 0-10Nm variable load range, demonstrating its good adaptability to various working conditions.

[0108] Table 4 Comparison of diagnostic accuracy under different loads

[0109] load accuracy F1 score 0Nm 94.51% 94.52% 2Nm 96.25% 96.23% 4Nm 98.75% 98.75% 6Nm 97.35% 97.20% 8Nm 96.25% 96.23% 10Nm 93.75% 93.65%

[0110] Noise robustness test: Table 5 shows that under 30% Gaussian noise interference, the diagnostic accuracy of the four-sensor fusion scheme can still be maintained at 82.21%, and the performance degradation is much smaller than that of the single-sensor scheme, highlighting the great advantage of multi-source fusion in anti-interference.

[0111] Table 5. Comparison of diagnostic accuracy under different noise levels

[0112] Noise level 4 sensors (AE+Vib+Temp+Sound) Optimal 3-sensor setup (AE + Temp + Sound) Optimal 2-sensor setup (AE+Vib) Best Sensor 1 (AE) 0% (Original) 94.51% 90.43% 88.66% 80.12% 10% noise 89.34% 86.56% 84.79% 78.29% 20% noise 86.37% 83.62% 79.57% 74.39% 30% noise 82.21% 78.27% 75.24% 69.20%

[0113] Contributions to the core module for ablation experiment verification:

[0114] Attention mechanism ablation: Table 6 shows that after removing all attention mechanisms, the model performance decreased by 4.74%, with cross-modal attention contributing the most, confirming the key role of dynamic fusion mechanism.

[0115] Table 6 Results of the ablation experiment on the attention mechanism

[0116] Model Configuration accuracy F1 score Complete model (triple attention mechanism) 94.51% 94.52% Remove sensor-level attention mechanism 92.22% 93.05% Remove channel attention mechanism 92.13% 91.89% Remove cross-modal attention mechanism 91.05% 90.76% Remove all attention mechanisms 89.77% 89.42%

[0117] Sensor contribution ablation: Table 7 assesses the contribution of each sensor by removing them one by one. The acoustic emission sensor contributes the most (performance decreases by 5.30% after removal), which is highly consistent with the fault mechanism; the temperature sensor seems to contribute the least, but it is indispensable in trend prediction as a slow-variable indicator. This experiment reverse-engineers the signal selection principle based on the fault mechanism from a data perspective.

[0118] Table 7 Ablation Experiment Results of Sensor Contribution

[0119] Removed sensor Sensor combination accuracy none Acoustic emission + vibration + temperature + microphone 94.51% acoustic emission Vibration + Temperature + Microphone 89.21% vibration Acoustic emission + temperature + microphone 89.78% temperature Acoustic emission + vibration + microphone 90.43% microphone Acoustic emission + vibration + temperature 90.00%

[0120] in conclusion

[0121] This invention, starting from the physical mechanism of harmonic reducer faults, constructs an intelligent diagnostic framework of "mechanism-oriented, signal optimization, and adaptive fusion." By designing a multi-source heterogeneous signal synchronous acquisition scheme, a modality-specific preprocessing workflow, a parallel dual-branch CNN-Transformer network, and a cross-modal, channel, and sensor triple attention dynamic fusion mechanism, it achieves a synergistic improvement in diagnostic accuracy, robustness, and interpretability. Detailed experimental data and system ablation analysis fully demonstrate the effectiveness and advancement of this invention in early-stage fault diagnosis of harmonic reducers, providing a reliable technical solution for the intelligent operation and maintenance of high-end equipment.

[0122] The method of this invention optimizes signals based on the physical nature of the fault, simultaneously acquiring four heterogeneous physical signals: vibration, acoustic emission, temperature, and sound. These four signals are not simply selected and piled up, but rather precisely correspond to different dimensions of fault characterization: vibration captures impact and modal response, acoustic emission captures crack propagation stress waves, temperature reflects the gradual process of frictional temperature rise, and sound perceives the overall dynamic behavior of air radiation. They form a natural complementary advantage in the time domain, frequency domain, and physical mechanism, constructing an information foundation covering all dimensions of the fault from the source.

[0123] In terms of modeling methodology, this invention proposes a more advanced adaptive multi-branch fusion network model. Its core innovations include:

[0124] 1. Dual-branch parallel architecture: A structure was constructed that combines the local feature extraction branch of CNN with the global dependency modeling branch of Transformer, realizing the collaborative learning of local details and global context information.

[0125] 2. Triple Attention Dynamic Fusion Mechanism: This innovative mechanism introduces cross-modal attention, channel attention, and sensor-level attention. Based on real-time input signals, it adaptively focuses on the most sensitive physical signals and key features of the current fault, achieving dynamic matching between fusion weights and the physical state of the fault.

[0126] 3. Adaptive architecture design: The model can automatically adjust the network complexity and fusion strategy according to the actual number of sensors installed (1-4), ensuring the optimal performance and practicality of the method under different engineering configurations.

[0127] In terms of preprocessing, this invention designs a mode-specific preprocessing workflow, targeting the physical characteristics and main noise sources of each type of signal: order tracking and resonance band energy ratio extraction are performed on vibration signals; wavelet soft thresholding is used for acoustic emission signals; environmental compensation and moving average filtering are performed on temperature signals; and echo cancellation and time-spectrum generation are performed on sound signals. This targeted processing can more effectively extract essential fault information from various types of signals, laying a solid foundation for subsequent fusion diagnosis.

[0128] In summary, this invention proposes original technical solutions for diagnostic objects, signal systems, fusion models, and preprocessing procedures, solving problems such as poor reliability of single signal sources, weak physical interpretability of traditional fusion methods, and insufficient adaptability of models to operating conditions, and achieving a synergistic improvement in diagnostic accuracy, robustness, and interpretability.

Claims

1. A method for intelligent fault diagnosis of harmonic reducers based on multi-source information fusion, characterized in that: It includes the following steps: S1, based on the physical mechanism of typical faults of harmonic reducers, simultaneously collects four heterogeneous physical signals: vibration, acoustic emission, temperature and sound, and performs modal-specific preprocessing on each; S2, construct an adaptive multi-branch fusion network model, which includes a parallel CNN local feature extraction branch and a Transformer global dependency modeling branch, and introduces a cross-modal, channel and sensor-level triple attention mechanism to achieve dynamic matching of fusion weight allocation with fault state; S3, design an adaptive network architecture that dynamically adjusts the fusion strategy and network complexity based on the number of online sensors and operating conditions; S4, the model is trained using a multi-objective loss function; S5 performs online fault diagnosis.

2. The intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion according to claim 1, characterized in that: Specifically, S1 involves analyzing the excitation mechanism of harmonic reducer faults, clarifying the advantages of four physical signals—vibration, acoustic emission, temperature, and sound—in characterizing different faults, and building a multi-sensor synchronous acquisition platform accordingly.

3. The intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion according to claim 2, characterized in that: Specifically, S2 involves designing targeted preprocessing algorithms for the heterogeneity of the four types of signals and the main noise sources: order tracking and resonance band energy ratio extraction for vibration signals; wavelet soft threshold denoising for acoustic emission signals; environmental compensation and moving average filtering for temperature signals; echo cancellation and generation of time-spectrum diagrams for sound signals; and finally, extraction of mean, variance, kurtosis, and skewness statistical features from the preprocessed signals to form a multi-source heterogeneous feature set.

4. The intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion according to claim 1, characterized in that: The adaptive multi-branch fusion network model includes two parallel branches: one is a CNN branch composed of multiple one-dimensional convolutions, used to extract local fault features; One branch is a global dependency modeling branch consisting of a Transformer encoder. The output features of the two branches are dynamically fused through a triple attention fusion module, which includes, in sequence: (1) Cross-modal attention promotes the interaction between local details of CNN and global context information of Transformer; (2) Channel attention, adaptively emphasizing important feature channels; (3) Sensor-level attention, dynamically assigning decision weights to different physical sensors based on the current input; Furthermore, the adaptive multi-branch fusion network model employs an adaptive architecture design, automatically selecting different configurations from simple CNNs to complex multi-attention fusion networks based on the number of input sensors, ensuring optimal resource utilization and the best balance between performance.

5. The intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion according to claim 1, characterized in that: Specifically, S4 involves using a multi-objective loss function to supervise the model. This multi-objective loss function combines the cross-entropy loss of the main classifier and the auxiliary classification loss of the two parallel branches to promote collaborative learning and feature consistency between branches. Training uses the AdamW optimizer and a cosine annealing learning rate scheduling strategy.

6. The intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion according to claim 1, characterized in that: Specifically, S5 involves preprocessing and feature constructing the multi-source signals of the harmonic reducer to be diagnosed, and then inputting them into the trained adaptive fusion diagnostic model to obtain the diagnostic results of its health status or fault category.

7. The intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion according to claim 3, characterized in that: The vibration signal is based on the formula: ， Order tracking based on rotational speed synchronization is performed to eliminate the influence of rotational speed fluctuations, and the frequency band energy ratio is calculated to focus on the structural resonance frequency band, thus separating the structural resonance frequency band. In the formula The frequency band energy ratio, The distribution of signal power spectral density in the target frequency band. The total power spectral density across the entire frequency band. , These represent the upper and lower limits of the structural resonance frequency band.

8. The intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion according to claim 3, characterized in that: The acoustic emission signal is based on the formula: ， An improved wavelet soft thresholding denoising algorithm is used to effectively suppress high-frequency electromagnetic interference and highlight damage-related sudden events; In the formula These are the denoised wavelet coefficients. These are the original wavelet transform coefficients. This is the threshold parameter.

9. The intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion according to claim 3, characterized in that: The temperature signal is based on the formula: ， Ambient temperature compensation is performed, and then a smooth temperature rise curve is obtained by moving average filtering. The moving average filter window width is w=10s. To compensate for the temperature value, The original temperature sensor measurement value. This is the measured value of ambient temperature.

10. The intelligent fault diagnosis method for harmonic reducers based on multi-source information fusion according to claim 3, characterized in that: The sound signal is based on the formula: ， Adaptive echo cancellation is used to suppress background noise, and a time-frequency diagram is generated through short-time Fourier transform; In the formula is the window function for the short-time Fourier transform, and N is the window width.

Images