A multi-source information fusion mechanical equipment state monitoring method

Abstract

This invention relates to a multi-source information fusion method for monitoring the condition of mechanical equipment, belonging to the field of mechanical equipment condition monitoring technology. The invention involves installing acoustic sensors, temperature sensors, and vibration sensors on key parts of the equipment to collect acoustic, temperature, and vibration signals, respectively. Preprocessing is performed on the acoustic, temperature, and vibration modal data. The variational mode decomposition (VMD) parameters are optimized using the Crown Porcupine optimization algorithm. Time-domain features, frequency-domain features, and time-frequency features are extracted and processed from the intrinsic mode functions obtained from multiple Crown Porcupine-optimized variational mode decompositions. Then, the equipment condition is classified and identified through acoustic-vibration collaborative feature learning, feature fusion, and intelligent decision-making modules, outputting the equipment condition monitoring results. This invention solves the technical problems of incomplete information, temperature changes affecting monitoring accuracy, and low efficiency of multi-source information fusion in traditional single-mode monitoring methods, achieving high-precision and robust intelligent monitoring of mechanical equipment.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment condition monitoring technology, and in particular to a mechanical equipment condition monitoring method that integrates multi-source information. Background Technology

[0002] With the continuous improvement of industrial automation, the requirements for the reliability and safety of mechanical equipment are becoming increasingly stringent. Traditional equipment maintenance strategies mainly rely on regular inspections and post-incident repairs, which are not only costly but also difficult to avoid losses caused by sudden failures. Currently, mechanical equipment condition monitoring mainly adopts single sensor technology. Vibration analysis is the most mature monitoring method, which can effectively identify the fault characteristics of rotating parts, but it has limitations in low-speed equipment and early fault detection. Acoustic monitoring technology can capture abnormal sound signals during equipment operation and has high sensitivity to certain fault types, but it is easily affected by environmental noise. Temperature monitoring can reflect changes in the thermodynamic state of equipment and has a good indication effect on faults such as overheating and friction, but its response speed is relatively slow.

[0003] Existing multi-source information fusion methods mainly employ simple feature splicing or weighted averaging, neglecting the complex correlations and complementarities between different modal information, resulting in limited fusion effects. Furthermore, existing methods lack effective compensation mechanisms for the impact of temperature changes on sensor signals, leading to a significant decrease in monitoring accuracy in industrial environments with large temperature fluctuations. The performance of variational mode decomposition largely depends on the selection of the number of modes and the penalty factor. Traditional parameter selection methods mainly rely on manual experience, which is inefficient and prone to getting trapped in local optima. Therefore, predictive maintenance technology based on condition monitoring has become an important development direction for modern industry. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for monitoring the condition of mechanical equipment by fusing multi-source information.

[0005] The technical solution adopted in this invention is: A method for monitoring the condition of mechanical equipment by fusing multi-source information specifically includes the following steps: S1. Data Acquisition: Acoustic sensors, temperature sensors, and vibration sensors are installed in key parts of the equipment to collect acoustic signals, temperature signals, and vibration signals of the mechanical equipment, respectively. The acoustic, temperature, and vibration modal data are preprocessed. If the processed signal quality is qualified, it proceeds to the next step; otherwise, the signal quality is unsuitable and processing continues. S2. Optimize the variational mode decomposition (VMD) parameters using the porcupine optimization algorithm: The Crowned Porcupine optimization algorithm is used to adaptively optimize the number of modes and the penalty factor of variational mode decomposition, and the preprocessed signal is decomposed to obtain multiple intrinsic mode functions. If the VMD parameter optimization is completed, the next step is performed; if the VMD parameter optimization is not completed, the optimization continues. S3. Extract and process features: From the intrinsic mode functions obtained by multiple variational mode decompositions optimized by porcupines, time-domain features, frequency-domain features, and time-frequency features are extracted respectively, and the composite feature set of the extracted time-domain features, frequency-domain features, and time-frequency features is encoded by a temperature-adaptive Transformer encoder; S4, Acoustic-Vibration Co-functional Feature Learning: Computational acoustic signals With vibration signal The coherence function is used to evaluate frequency domain correlation, and the phase information of cross power spectral density is extracted to analyze phase synchronization characteristics. Then, a nonlinear coupling model of acoustic and vibration signals is established through bispectral analysis. S5, Feature Fusion: Feature fusion is performed using a multi-scale time-frequency fusion module, and the feature importance evaluation module within the multi-scale time-frequency fusion module evaluates mutual information. and normalized variance The importance score of each feature is calculated. The adaptive weight allocation module in the multi-scale time-frequency fusion module uses the softmax function to calculate the adaptive weights for the importance scores. The multi-level feature fusion module in the multi-scale time-frequency fusion module performs feature-level fusion and decision-level fusion respectively. If the feature fusion is completed, the process proceeds to the next step. If it is not completed, the process returns to step S2. S6. Intelligent decision-making module classifies and identifies device status: An ensemble classifier consisting of Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting Tree (GBDT) is constructed. The ensemble weights are determined based on the accuracy of each classifier on the validation set. The Monte Carlo Dropout method is used to estimate the prediction uncertainty. Finally, the device status monitoring results are output.

[0006] The preprocessing steps of the multi-source information fusion mechanical equipment condition monitoring method are as follows: S11. Calculate acoustic signals using the data synchronization alignment module. With vibration signal The cross-correlation function, by finding The maximum value determines the optimal delay. : ; ; Time alignment of acoustic signals: ; S12, Signal Quality Assessment: Within each time window, with a length of T_win and an overlap rate of 50%, the following three indices are calculated for both acoustic and vibration signals: (1) Signal-to-noise ratio (SNR): First, calculate the main signal energy as signal power Ps based on the operating frequency band of the equipment, and then estimate the noise power Pn based on the average energy of the non-operating frequency band. The SNR is taken as 10×log10(Ps / Pn). When the acoustic signal SNR ≥ 10 dB and the vibration signal SNR ≥ 12 dB, it is considered qualified. (2) Integrity Index (CI): This represents the proportion of valid sampling points to the total number of sampling points. A CI below 0.95 indicates incomplete sampling or packet loss. (3) Consistency Index (COI): Represents the stability of signal fluctuations. It is calculated by subtracting the ratio of variance to the square of the mean from 1. The result is limited to between 0 and 1. When COI ≥ 0.6, the signal is considered stable. When all three indicators meet the above conditions, the signal quality is considered qualified; if not, denoising or resampling operations are performed, and signal windows that fail twice consecutively are automatically downweighted during subsequent feature fusion. S13, Temperature Adaptive Correction: A sensor temperature response model is established to linearly compensate for the temperature sensitivity of acoustic and vibration signals; the specific method is as follows: (1) Under constant operating conditions, signals were collected under multiple temperature conditions, and the signal amplitude of each group was statistically analyzed; (2) By least squares fitting, the linear relationship between the signal amplitude and temperature is obtained; (3) Define temperature compensation coefficients βa (acoustic) and βv (vibration). During real-time compensation, the correction coefficient is calculated based on the difference between the current temperature T and the reference temperature Tref. ; When the temperature change exceeds ±10℃, the compensation coefficient can be calibrated separately in different temperature ranges; After the above processing, a temperature-corrected multimodal signal is obtained, which is used for subsequent VMD parameter optimization and feature extraction.

[0007] The multi-source information fusion mechanical equipment condition monitoring method, in step S2, specifically includes the following sub-steps of the Crowned Porcupine optimization algorithm: S21. Initialize the Crowned Hog pig population, initializing the population size to... The maximum number of iterations in the Crown Porcupine pig population Each individual represents a set of VMD parameter combinations; S22. Define the fitness function as a weighted combination of envelope entropy and correlation coefficient; S23. Dynamically adjust the search parameters and modal number based on the current temperature information. The search scope is Punishment factor The search scope is ; S24, Crowned Porcupine Defensive Behavior and Location Update; S25. Update the population position and evaluate fitness, iterate to find the best option until convergence.

[0008] In the aforementioned multi-source information fusion method for monitoring the condition of mechanical equipment, the temperature adjustment factor of the temperature adaptive Transformer encoder in step S3 is calculated based on the difference between the current average temperature and the reference temperature. ; Temperature regulation coefficient , , The average temperature within the current time window. This is a reference temperature.

[0009] The multi-source information fusion mechanical equipment condition monitoring method, the encoding processing step in step S3 is as follows: S31. Map the input multimodal feature matrix to a unified high-dimensional feature space through a linear transformation to obtain the initial feature embedding sequence; S32. Calculate the temperature adjustment factor based on the difference between the average temperature Tcur and the reference temperature Tref within the current time window. γ is the temperature adjustment coefficient, which is used to dynamically adjust the distribution of attention weights in the Transformer. In the self-attention mechanism encoding process, the temperature adjustment factor is introduced into the multi-head attention module to perform weighted correction on the query, key and value vectors. The temperature-adjusted attention mechanism strengthens the suppression of temperature-sensitive features and enhances robust features. S33. The attention output is layer normalized and deep nonlinear features are extracted through a feedforward fully connected network. In the output stage, a temperature-compensated attention layer is introduced to perform weighted fusion of acoustic, vibration and temperature modal features. The softmax function is used to calculate the adaptive weights of each modality to achieve multimodal feature fusion based on temperature perception. Finally, the fused feature vector after temperature adaptive Transformer encoding is output to provide input for the subsequent acoustic-vibration collaborative feature learning and multi-scale time-frequency fusion module.

[0010] In the aforementioned multi-source information fusion method for monitoring the condition of mechanical equipment, the coherence function in step S4 is calculated using the Welch method, the window function is selected as the Hanning window, the window length is set to 1024 points, and the overlap rate is set to 50%.

[0011] In the multi-source information fusion mechanical equipment condition monitoring method, the weights of the integrated classifiers in step S6 are determined based on the accuracy of each classifier on the validation set. The weights of the Support Vector Machine (SVM) are set to 0.4, the Random Forest (RF) weights are set to 0.3, and the Gradient Boosting Tree (GBDT) weights are set to 0.3.

[0012] Due to the adoption of the technical solution described above, the present invention has the following advantages: This invention presents a multi-source information fusion method for monitoring the condition of mechanical equipment. By constructing a temperature-adaptive Transformer encoding module and employing a temperature-regulated attention mechanism, it effectively addresses the impact of temperature changes on sensor signals, maintaining high monitoring accuracy over a wide temperature range. It utilizes acoustic-vibration collaborative feature learning to establish a physical correlation model between acoustic and vibration signals, fully leveraging the complementarity of multi-modal information, resulting in a 12-18% improvement in fault detection accuracy compared to traditional single-modal methods. An improved hog optimization algorithm is employed to automatically optimize VMD parameters, avoiding the subjectivity and inefficiency of manual parameter tuning, improving parameter optimization efficiency by over 30% compared to traditional genetic algorithms. Through multi-scale time-frequency fusion and intelligent decision-making layer integration, it achieves accurate identification and reliable early warning for various fault types, with an overall diagnostic accuracy exceeding 90%. This invention solves the technical problems of incomplete information, the impact of temperature changes on monitoring accuracy, and low efficiency of multi-source information fusion in traditional single-modal monitoring methods, achieving high-precision and robust intelligent monitoring of mechanical equipment. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the status monitoring process of the present invention. Detailed Implementation

[0014] The present invention will be further explained and described below with reference to the accompanying drawings and embodiments. However, this should not be construed as limiting the scope of protection of the present invention. The purpose of disclosing the present invention is to protect all technical improvements within the scope of the present invention.

[0015] Combined with appendix Figure 1 The aforementioned method for monitoring the condition of mechanical equipment based on multi-source information fusion specifically includes the following steps: S1. Multimodal data preprocessing: Acoustic sensors, temperature sensors, and vibration sensors are installed in key parts of the equipment. The acoustic sensors utilize omnidirectional condenser microphones with a frequency response range of 20Hz to 20kHz and a sensitivity of -38dB. These are installed near rotating parts or bearing housings to capture subtle acoustic changes during operation. The temperature sensors employ PT100 platinum resistance thermometers with a measurement range of -50°C to 150°C and an accuracy of ±0.1°C. These are installed in critical heat-generating areas, such as bearing housings, gearbox housings, or motor stator surfaces, to monitor changes in thermal characteristics. The vibration sensors use piezoelectric accelerometers with a frequency response range of 10Hz to 5kHz and a sensitivity of 100mV / g. These are installed in vibration-sensitive parts of the equipment. Positions such as bearing covers, machine bases, or support base plates are used to capture mechanical vibration signals. The above three types of sensors respectively collect acoustic, temperature, and vibration signals from the mechanical equipment. The collected multimodal data undergoes time synchronization alignment, signal quality assessment, and temperature adaptive correction in sequence. The data acquisition parameters are set as follows: acoustic signal sampling frequency is 48kHz, quantization bit depth is 16bit; temperature signal sampling frequency is 10Hz, quantization bit depth is 12bit; vibration signal sampling frequency is 10kHz, quantization bit depth is 16bit; the data acquisition time window length T_win is 2 seconds; the time window overlap rate is 50%, meaning a monitoring result is output every 1 second, thereby achieving continuous monitoring of the mechanical equipment's operating status. S11. Time synchronization and alignment of multimodal data: Through the data synchronization and alignment module, calculate the acoustic signal. With vibration signal cross-correlation function By searching The maximum value determines the optimal delay. : ; ; Time alignment of acoustic signals: ; S12, Signal Quality Assessment: Within each time window, with a length of T_win and an overlap rate of 50%, the following three indices are calculated for both acoustic and vibration signals: (1) Signal-to-noise ratio (SNR): First, calculate the main signal energy as signal power Ps based on the operating frequency band of the equipment, and then estimate the noise power Pn based on the average energy of the non-operating frequency band. The SNR is taken as 10×log10(Ps / Pn). When the acoustic signal SNR ≥ 10 dB and the vibration signal SNR ≥ 12 dB, it is considered qualified. (2) Integrity Index (CI): This represents the proportion of valid sampling points to the total number of sampling points. A CI below 0.95 indicates incomplete sampling or packet loss. (3) Consistency Index (COI): Represents the stability of signal fluctuations. It is calculated by subtracting the ratio of variance to the square of the mean from 1. The result is limited to between 0 and 1. When COI ≥ 0.6, the signal is considered stable. When all three indicators meet the above conditions, the signal quality is considered qualified; if not, denoising or resampling operations are performed, and signal windows that fail twice consecutively are automatically downweighted during subsequent feature fusion. S13, Temperature Adaptive Correction: A sensor temperature response model is established to linearly compensate for the temperature sensitivity of acoustic and vibration signals; the specific method is as follows: (1) Under constant operating conditions, with constant speed and load, signals are collected under multiple temperature conditions, and the signal amplitude of each group is statistically analyzed, such as the RMS value. (2) By least squares fitting, the linear relationship between the signal amplitude and temperature is obtained; (3) Define temperature compensation coefficients βa (acoustic) and βv (vibration). During real-time compensation, the correction coefficient is calculated based on the difference between the current temperature T and the reference temperature Tref. ; When the temperature change exceeds ±10℃, the compensation coefficient can be calibrated separately in different temperature ranges; After the above processing, a temperature-corrected multimodal signal is obtained, which is used for subsequent VMD parameter optimization and feature extraction. S2. Optimize the variational mode decomposition (VMD) parameters using the porcupine optimization algorithm: The Crowned Porcupine optimization algorithm is used to adaptively optimize the number of modes and the penalty factor in variational mode decomposition, and the preprocessed signal is decomposed to obtain multiple intrinsic mode functions. The optimization algorithm for the crowned porcupine specifically includes the following sub-steps: S21. Initialize the Crowned Hog pig population, initializing the population size to... The maximum number of iterations in the Crown Porcupine pig population Each individual This represents a set of VMD parameter combinations; S22. Define the fitness function as a weighted combination of envelope entropy EE and correlation coefficient CC: ; Among them, weight ; Envelope entropy calculation process: For each IMF component The analytic signal is obtained by calculating the Hilbert transform: ; Calculate the instantaneous amplitude: ; Normalization process: ; Envelope entropy: ; Correlation coefficient calculation of the original signal With reconstructed signal Correlation coefficient: ; Dynamically adjust search parameters based on current temperature information: ; in , ; S23. Dynamically adjust the search parameters based on the current temperature information. The parameter settings are as follows: Population size. Maximum number of iterations Modal number The search scope is Punishment factor The search scope is ; S24, Crowned Porcupine Defensive Behavior and Location Update: Each candidate individual represents a set of VMD parameters: the number of modes K and the penalty factor α. In each iteration of the algorithm, the individual sequentially performs the following four defensive behaviors to adaptively search for the optimal parameters in the global and local spaces: (1) Visual defense (local fine-tuning): Randomly generate candidate solutions with small variations near the current optimal solution to refine the parameters. The magnitude of the variation gradually decreases with the number of iterations, which is equivalent to the annealing process; (2) Auditory defense (neighborhood guidance): Using the better performing individuals in the neighborhood as a reference, move along their direction by a certain proportion and add slight random perturbation for local exploration; (3) Odor defense (global jump): Periodically perform random jumps with large step sizes to escape local optimum traps. The jump amplitude is large in the early iterations and gradually converges in the later stages. (4) Physical defense (boundary constraints): If the updated parameters exceed the allowable range, they are reflected back to the feasible range and a slight random perturbation is superimposed to maintain population diversity. Each individual generates four candidate parameter sets in one generation, corresponding to the four defense methods mentioned above. The fitness function value (based on a weighted combination of envelope entropy and correlation coefficient) is calculated for each individual, and the best one is retained to enter the next generation. When the difference between the current average temperature and the reference temperature exceeds a set threshold (e.g., 10℃), the algorithm will automatically increase the execution probability of global jump (odor defense) to expand the search range. When the temperature difference is small or the algorithm is close to convergence, visual and auditory defenses are mainly activated to improve local search accuracy. The algorithm is considered converged when the fitness improvement is less than 0.001 over 10 consecutive generations or when the maximum number of iterations (e.g., 150 generations) is reached, and the optimal parameters are output. ; S25. Update the population position and evaluate fitness, iteratively optimizing until convergence: Iteratively update the population position and evaluate fitness. The process terminates early when the fitness improvement is less than 0.001 for 10 consecutive generations, or stops when the maximum number of iterations is reached. The acoustic and vibration signals are decomposed using optimized VMD parameters; the constrained variational problem of VMD is expressed as: ; ; By introducing the Lagrange multiplier and secondary penalty factor This transforms the constrained optimization problem into an unconstrained problem: ; The alternating direction multiplier method is used to solve the problem, and the results are obtained. One intrinsic mode function (IMF) is used to extract the first-order gradient from the temperature signal. Second gradient and rate of temperature change feature; S3. Extract and process features: Time-domain features, frequency-domain features, and time-frequency features are extracted from multiple intrinsic mode functions. The extracted features from the three modes are then normalized and standardized, and constructed into a feature matrix in chronological order. ,in For sequence length, The feature dimension is defined as follows: The sequence length represents the number of time windows, and the feature dimension represents the combined dimension of each modal feature. A temperature-adaptive Transformer encoder is used to encode the multimodal features, including the composite feature set of time-domain, frequency-domain, and time-frequency features. The specific processing steps are as follows: S31. The constructed multimodal feature matrix is ​​mapped to a unified high-dimensional feature space through a linear transformation to obtain the initial feature embedding sequence; S32. Calculate the temperature adjustment factor based on the difference between the average temperature Tcur and the reference temperature Tref within the current time window. γ is the temperature adjustment coefficient, which is used to dynamically adjust the distribution of attention weights in the Transformer. In the self-attention mechanism encoding process, the temperature adjustment factor is introduced into the multi-head attention module to perform weighted correction on the query, key and value vectors. The temperature-adjusted attention mechanism strengthens the suppression of temperature-sensitive features and enhances robust features. S33. The attention output is layer-normalized and deep nonlinear features are extracted through a feedforward fully connected network. In the output stage, a temperature-compensated attention layer is introduced to perform weighted fusion of acoustic, vibration, and temperature modal features. The softmax function is used to calculate the adaptive weights of each modality to achieve temperature-sensing-based multimodal feature fusion. Finally, the fused feature vector encoded by the temperature-adaptive Transformer is output to provide input for the subsequent acoustic-vibration collaborative feature learning and multi-scale time-frequency fusion module. Through the above processing, the temperature-adaptive Transformer encoder introduces a temperature sensing mechanism in the feature representation stage, which can effectively compensate for the influence of environmental temperature changes on the distribution of acoustic and vibration signals, and enhance the robustness and monitoring accuracy of the model under complex working conditions. S4, Acoustic-Vibration Co-functional Feature Learning: The acoustic signal is calculated using the coherence analysis module. With vibration signal The coherence function was used to evaluate frequency domain correlation. The coherence function was calculated using the Welch method, with the Hanning window selected and the window length set to [value missing]. Points, with an overlap rate set to 50%; The coherence function is defined as: ; in The cross-power spectral density of the acoustic signal and the vibration signal: ; and The self-power spectral density of the acoustic signal and the vibration signal are respectively: ; Phase information of the cross-power spectral density is extracted using the phase synchronization detection module, and phase synchronization characteristics are analyzed. Instantaneous phase difference: ; Phase lock value: ; Where ⟨·> represents the time average; Phase synchronization index: ; Phase difference variance: ; Then, a nonlinear coupling model of acoustic and vibration signals is established through bispectral analysis using the coupling relationship modeling module; Bispectral definition:

[0016] Bicoherence function: ; in and These are the power spectra of the acoustic signal and the vibration signal, respectively. Nonlinear coupling strength: ; Third-order cumulant:

[0017] S5. Feature Fusion: Feature fusion is performed through a multi-scale time-frequency fusion module. The feature importance evaluation module in the multi-scale time-frequency fusion module evaluates mutual information. and normalized variance The importance score of each feature is calculated. The adaptive weight allocation module in the multi-scale time-frequency fusion module uses the softmax function to calculate the adaptive weights for the importance score. The multi-level feature fusion module in the multi-scale time-frequency fusion module performs feature-level fusion and decision-level fusion respectively. Mutual Information and normalized variance Calculate the importance score for each feature: Mutual information calculation: ; in For joint probability density, and These are the marginal probability densities; Normalized variance: ; Importance score: ; in , ; The importance scores are calculated using the softmax function to determine adaptive weights. ; in For temperature parameters, ensure

[0018] Then perform feature-level fusion and decision-level fusion separately: Feature-level fusion employs weighted summation:

[0019] Decision-level fusion adopts the Dempster-Shafer evidence theory; let m_i be the basic probability assignment function of the i-th classifier, and the evidence combination rule is: ; S6. Intelligent decision-making module classifies and identifies device status: An ensemble classifier was constructed using Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting Tree (GBDT). SVM decision function: ; in For the radial basis function kernel, the penalty parameter is... ; Random forest prediction: ; in For the number of decision trees, For the first The predicted results for each tree; GBDT Prediction:

[0020] in For learning rate, For the first A weak learner, This represents the number of iterations. The ensemble weights are determined based on the accuracy of each classifier on the validation set: ; ; ; Final prediction result:

[0021] The Monte Carlo Dropout method is used to estimate the prediction uncertainty: conduct Each forward propagation involves randomly discarding a portion of neurons: ; Calculate the prediction variance: ; in This represents the average prediction result; Confidence assessment module: Calculates confidence based on prediction variance. ; in The maximum prediction variance among all samples; when The prediction was considered reliable at the time, among which This is the confidence threshold.

[0022] The parts of this invention not described in detail are prior art.

[0023] The embodiments selected herein for the purpose of disclosing the inventive objectives are currently considered suitable; however, it should be understood that the invention is intended to include all variations and modifications of the embodiments that fall within the scope of this concept and invention.

Claims

1. A method for monitoring the condition of mechanical equipment through multi-source information fusion, characterized in that: Specifically, the following steps are included: S1. Data Acquisition: Acoustic sensors, temperature sensors, and vibration sensors are installed in key parts of the equipment to collect acoustic signals, temperature signals, and vibration signals of the mechanical equipment, respectively. The acoustic, temperature, and vibration modal data are preprocessed. If the processed signal quality is qualified, it proceeds to the next step; otherwise, the signal quality is unsuitable and processing continues. S2. Optimize the variational mode decomposition (VMD) parameters using the porcupine optimization algorithm: The Crowned Porcupine optimization algorithm is used to adaptively optimize the number of modes and the penalty factor of variational mode decomposition, and the preprocessed signal is decomposed to obtain multiple intrinsic mode functions. If the VMD parameter optimization is completed, the next step is performed; if the VMD parameter optimization is not completed, the optimization continues. S3. Extract and process features: From the intrinsic mode functions obtained by multiple variational mode decompositions optimized by porcupine, time-domain features, frequency-domain features, and time-frequency features are extracted from the intrinsic mode functions. The extracted features under the three modes are then normalized and standardized, and constructed into a feature matrix in chronological order. ,in For sequence length, The feature dimension is defined as follows: sequence length represents the number of time windows, and feature dimension represents the comprehensive dimension of each modality feature. The extracted multimodal features are encoded using a temperature-adaptive Transformer encoder. S4, Acoustic-Vibration Co-functional Feature Learning: Computational acoustic signals With vibration signal The coherence function is used to evaluate the frequency domain correlation, and the phase information of the cross power spectral density is extracted to analyze the phase synchronization characteristics. Then, a nonlinear coupling model of acoustic and vibration signals is established through bispectral analysis. S5, Feature Fusion: Feature fusion is performed using a multi-scale time-frequency fusion module, and the feature importance evaluation module within the multi-scale time-frequency fusion module evaluates mutual information. and normalized variance The importance score of each feature is calculated. The adaptive weight allocation module in the multi-scale time-frequency fusion module uses the softmax function to calculate the adaptive weights for the importance scores. The multi-level feature fusion module in the multi-scale time-frequency fusion module performs feature-level fusion and decision-level fusion respectively. If the feature fusion is completed, the process proceeds to the next step. If it is not completed, the process returns to step S2. S6. Intelligent decision-making module classifies and identifies device status: An ensemble classifier consisting of Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting Tree (GBDT) is constructed. The ensemble weights are determined based on the accuracy of each classifier on the validation set. The Monte Carlo Dropout method is used to estimate the prediction uncertainty. Finally, the device status monitoring results are output.

2. The method for monitoring the condition of mechanical equipment based on multi-source information fusion according to claim 1, characterized in that: The specific steps of preprocessing in step S1 are as follows: S11. Calculate acoustic signals using the data synchronization alignment module. With vibration signal cross-correlation function By searching The maximum value determines the optimal delay. : ； ； Time alignment of acoustic signals: ; S12, Signal Quality Assessment: Within each time window, with a length of T_win and an overlap rate of 50%, the following three indices are calculated for both acoustic and vibration signals: (1) Signal-to-noise ratio (SNR): First, calculate the main signal energy as signal power Ps based on the operating frequency band of the equipment, and then estimate the noise power Pn based on the average energy of the non-operating frequency band. The SNR is taken as 10×log10(Ps / Pn). When the acoustic signal SNR ≥ 10 dB and the vibration signal SNR ≥ 12 dB, it is considered qualified. (2) Integrity Index (CI): This represents the proportion of valid sampling points to the total number of sampling points. A CI below 0.95 indicates incomplete sampling or packet loss. (3) Consistency Index (COI): Represents the stability of signal fluctuations. It is calculated by subtracting the ratio of variance to the square of the mean from 1. The result is limited to between 0 and 1. When COI ≥ 0.6, the signal is considered stable. When all three indicators meet the above conditions, the signal quality is considered qualified; if not, denoising or resampling operations are performed, and signal windows that fail twice consecutively are automatically downweighted during subsequent feature fusion. S13, Temperature Adaptive Correction: A sensor temperature response model is established to linearly compensate for the temperature sensitivity of acoustic and vibration signals; the specific method is as follows: (1) Under constant operating conditions, signals were collected under multiple temperature conditions, and the signal amplitude of each group was statistically analyzed; (2) By least squares fitting, the linear relationship between the signal amplitude and temperature is obtained; (3) Define temperature compensation coefficients βa (acoustic) and βv (vibration). During real-time compensation, the correction coefficient is calculated based on the difference between the current temperature T and the reference temperature Tref. ； When the temperature change exceeds ±10℃, the compensation coefficient can be calibrated separately in different temperature ranges; After the above processing, a temperature-corrected multimodal signal is obtained, which is used for subsequent VMD parameter optimization and feature extraction.

3. The method for monitoring the condition of mechanical equipment based on multi-source information fusion according to claim 1, characterized in that: Step S2 of the Crowned Porcupine optimization algorithm specifically includes the following sub-steps: S21. Initialize the Crowned Hog pig population, initializing the population size to... The maximum number of iterations in the Crown Porcupine pig population Each individual represents a set of VMD parameter combinations; S22. Define the fitness function as a weighted combination of envelope entropy and correlation coefficient; S23. Dynamically adjust the search parameters and modal number based on the current temperature information. The search scope is Punishment factor The search scope is ; S24, Crowned Porcupine Defensive Behavior and Location Update; S25. Update the population position and evaluate fitness, iterate to find the best option until convergence.

4. The method for monitoring the condition of mechanical equipment based on multi-source information fusion according to claim 1, characterized in that: In step S3, the temperature adjustment factor of the temperature-adaptive Transformer encoder is calculated based on the difference between the current average temperature and the reference temperature. ; Temperature regulation coefficient , , The average temperature within the current time window. This is a reference temperature.

5. The method for monitoring the condition of mechanical equipment based on multi-source information fusion according to claim 1, characterized in that: The encoding process described in step S3 is as follows: S31. Map the input multimodal feature matrix to a unified high-dimensional feature space through a linear transformation to obtain the initial feature embedding sequence; S32. Calculate the temperature adjustment factor based on the difference between the average temperature Tcur and the reference temperature Tref within the current time window. γ is the temperature adjustment coefficient, which is used to dynamically adjust the distribution of attention weights in the Transformer. In the self-attention mechanism encoding process, the temperature adjustment factor is introduced into the multi-head attention module to perform weighted correction on the query, key and value vectors. The temperature-adjusted attention mechanism strengthens the suppression of temperature-sensitive features and enhances robust features. S33. Perform layer normalization on the attention output and extract deep nonlinear features through a feedforward fully connected network; In the output stage, a temperature-compensated attention layer is introduced to perform weighted fusion of acoustic, vibration, and temperature modal features. The softmax function is used to calculate the adaptive weights of each modality to achieve multimodal feature fusion based on temperature perception. Finally, the fused feature vector encoded by temperature adaptive Transformer is output to provide input for subsequent acoustic-vibration collaborative feature learning and multi-scale time-frequency fusion modules.

6. The method for monitoring the condition of mechanical equipment based on multi-source information fusion according to claim 1, characterized in that: In step S4, the coherence function is calculated using the Welch method, with the Hanning window selected as the window function, a window length of 1024 points, and an overlap rate of 50%.

7. The method for monitoring the condition of mechanical equipment based on multi-source information fusion according to claim 1, characterized in that: The weights of the ensemble classifiers in step S6 are determined based on the accuracy of each classifier on the validation set. The weight of the Support Vector Machine (SVM) is set to 0.4, the weight of the Random Forest (RF) is set to 0.3, and the weight of the Gradient Boosting Tree (GBDT) is set to 0.3.

Images