A motor fault diagnosis method based on waveform analysis

Abstract

The application relates to the field of fault prediction and health management, and discloses a motor fault diagnosis method based on waveform analysis, which comprises the following steps: synchronously collecting electrical waveform signals and mechanical waveform signals in the operation process of a motor; performing adaptive wavelet filtering processing on the two signals based on frequency band energy distribution to eliminate background noise interference; respectively extracting statistical parameters of the filtered signals to construct a global one-dimensional feature vector; respectively calculating time-frequency energy matrices of the two signals and performing cross-multiplication mapping to generate a two-dimensional cross-modal coupling matrix representing electromechanical interaction characteristics; and inputting the matrix and the vector into a pre-trained double-flow convolutional neural network to perform multi-scale feature fusion and decision, and outputting a final fault diagnosis result. The application complements the advantages of implicit micro-modulation and explicit physical constraints, establishes a closed-loop adaptive evolution mechanism in combination with confidence evaluation, and improves the accuracy and robustness of diagnosis under harsh working conditions.

Description

Technical Field

[0001] This invention relates to the field of fault prediction and health management technology, specifically a motor fault diagnosis method based on waveform analysis. Background Technology

[0002] As a fundamental power source in industrial production, the operating status of electric motors is closely related to the continuity and safety of the production system. Monitoring the condition of electric motors and providing early warning of faults are standard practices to ensure the normal operation of equipment.

[0003] Currently, motor fault diagnosis mainly relies on signal analysis of single physical quantities, such as acquiring current signals from electrical circuits alone, or monitoring vibration signals from the motor surface alone. Since internal motor faults are usually accompanied by interactions between the electromagnetic system and the mechanical structure, signals from a single physical domain are insufficient to fully reflect the physical processes and modulation mechanisms of complex electromechanical faults.

[0004] Some technical solutions attempt to incorporate multi-sensor data for comprehensive diagnosis. However, in multimodal data processing, existing methods typically cascade and combine features extracted from different sensor channels without establishing a mathematical model for the deep cross-modulation relationship between electrical and mechanical signals in time-frequency energy distribution. In the signal acquisition and preprocessing stages, conventional filtering and denoising algorithms often employ a globally fixed threshold. This fixed truncation method, while eliminating system background noise, easily filters out weak high-frequency response components caused by early motor faults, resulting in the loss of crucial features.

[0005] Furthermore, when applying deep neural networks for fault classification, existing models primarily rely on data-driven methods to extract implicit features, failing to effectively integrate the network-extracted features with the actual physical and statistical constraints of the motor. This limits the model's generalization ability when facing complex and ever-changing field conditions. Simultaneously, after deployment in industrial settings, the network weights of existing diagnostic models often remain fixed, lacking a mechanism for closed-loop updates using newly added operational data. As the mechanical components of the motor wear down and the external operating environment changes, the diagnostic accuracy of static models will decrease over time. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a motor fault diagnosis method based on waveform analysis. This method solves the problems of existing diagnostic methods, such as the tendency to lose early weak fault features under strong background noise, difficulty in effectively extracting and characterizing the deep cross-modulation mechanism of electrical and mechanical signals, poor generalization ability of static diagnostic models lacking physical constraints under complex and variable operating conditions, and inability to achieve closed-loop adaptive updates throughout the entire life cycle.

[0007] To achieve the above objectives, this invention provides a motor fault diagnosis method based on waveform analysis. This method integrates electrical and mechanical multimodal signals, combines adaptive wavelet filtering with the construction of a two-dimensional cross-modal coupling matrix, and introduces a dual-stream convolutional neural network for deep feature mining and judgment, thereby achieving highly robust and accurate diagnosis of the motor's operating status.

[0008] Specifically, the present invention provides a motor fault diagnosis method based on waveform analysis, comprising the following steps: The electrical and mechanical waveform signals during motor operation are acquired synchronously. In specific implementation, an analysis sliding window length is set, and based on a unified clock trigger source, the discrete electrical and mechanical waveform signals are synchronously acquired within the analysis sliding window length at a global synchronous sampling frequency that satisfies the sampling theorem.

[0009] The electrical waveform signal and the mechanical waveform signal are respectively subjected to adaptive wavelet filtering based on frequency band energy distribution to obtain filtered electrical waveform signals and filtered mechanical waveform signals, respectively. Specifically, firstly, the global mean and standard deviation of the signal are calculated, abnormal sampling points are identified based on a set dispersion criterion, and replaced by linear interpolation of adjacent normal sampling points. After deducting the corresponding global DC bias component, zero-mean sequences are obtained. Subsequently, discrete wavelet decomposition is performed on each of the zero-mean sequences to obtain the corresponding low-frequency approximation coefficients and high-frequency detail coefficients of multiple decomposition layers.

[0010] In the innovative filtering mechanism of this invention, to overcome the defect of feature loss caused by traditional fixed wavelet thresholds, an adaptive shrinkage threshold for the high-frequency detail coefficients of each decomposition layer is determined based on the frequency band energy distribution. First, the energy proportion of the high-frequency detail coefficients of each decomposition layer is calculated. Extract the median of the absolute values ​​of high-frequency detail coefficients. And combined with the total number of sequence sampling points Calculate the adaptive shrinkage threshold specific to each of the decomposition layers. Its computational logic can be expressed as: ; in, This is the scaling factor. The design causes the adaptive shrinkage threshold to decrease as the energy percentage increases, thereby preserving high-energy, high-frequency band signals rich in fault characteristics while filtering out unwanted noise to the greatest extent. This adaptive shrinkage threshold is used to shrink the corresponding high-frequency detail coefficients and perform wavelet reconstruction, resulting in filtered electrical and mechanical waveform signals.

[0011] Statistical parameters of the filtered electrical waveform signal and the filtered mechanical waveform signal are extracted respectively to construct a global one-dimensional feature vector, and a two-dimensional cross-modal coupling matrix is ​​constructed based on the time-frequency energy matrix of the filtered electrical waveform signal and the filtered mechanical waveform signal.

[0012] On the one hand, the time-domain and frequency-domain statistical parameters of the filtered signal are calculated separately, and after splicing, dimensionality reduction is performed by principal component analysis. Principal components whose cumulative variance contribution rate meets the set threshold are selected to construct the low-dimensional and uncorrelated global one-dimensional feature vector.

[0013] On the other hand, discrete wavelet packet decomposition is performed on the filtered signal to obtain multiple frequency band signal subspaces. The energy accumulation value is calculated within the analysis sliding window, and the electrical time-frequency energy matrix is ​​constructed accordingly. and mechanical time-frequency energy matrix To effectively extract the characteristics of electromechanical composite faults, the transposes of the electrical time-frequency energy matrix and the mechanical time-frequency energy matrix are cross-multiplied and mapped to generate the two-dimensional cross-modal coupling matrix. : ; This coupling matrix can quantitatively characterize the cooperative correlation of transient energy modulation and highlight the implicit interaction features between modes.

[0014] The two-dimensional cross-modal coupling matrix and the global one-dimensional feature vector are input into a pre-trained two-stream convolutional neural network for feature fusion and classification decision, and the fault diagnosis result of the motor is output to realize the fault prediction and health management of the motor.

[0015] The two-dimensional cross-modal coupling matrix is ​​used as a two-dimensional mesh tensor input through convolutional layers in the network. Multiple parallel convolutional kernels with different receptive field sizes are used to perform multi-scale two-dimensional convolution operations, obtaining feature maps of different scales. After max pooling and flattening, a one-dimensional vector of local depth features is generated. Subsequently, this one-dimensional vector of local depth features is concatenated and fused with the aforementioned one-dimensional global feature vector to form a hybrid feature vector. This hybrid feature vector is input into a fully connected network and mapped through a normalized exponential function to a decision vector containing probabilities of various diagnostic states. The state corresponding to the highest probability is selected as the fault diagnosis result, and the probability distribution of the decision vector is used as the pre-diagnosis and health management data for the motor.

[0016] Finally, the maximum probability value is compared with the set alarm judgment threshold. If the maximum probability value is greater than or equal to the alarm judgment threshold, a fault alarm response procedure is triggered, and the relevant signal waveform, feature vector, and cross-modal coupling matrix are stored as samples in the PHM database for periodic closed-loop updates of the network to support the fault prediction and health management mechanism of the motor. If the value is less than the alarm judgment threshold, the motor is determined to be in normal operating condition. This invention achieves a deep integration of physical-driven feature analysis and data-driven deep learning, significantly improving the fault diagnosis accuracy and robustness of motors under complex operating conditions.

[0017] This invention provides a motor fault diagnosis method based on waveform analysis. It has the following beneficial effects: 1. This invention overcomes the limitation of traditional fixed denoising thresholds, which easily obliterate key high-frequency information, by introducing an adaptive wavelet contraction mechanism based on frequency band energy distribution. This design dynamically links threshold attenuation with the proportion of frequency band energy, effectively filtering out complex industrial background noise while better preserving easily submerged weak fault transient impacts, significantly improving the sensitivity of early anomaly detection from the signal source.

[0018] 2. To address the challenge of electromechanical coupling in motor fault evolution, this solution proposes constructing a two-dimensional cross-modal coupling matrix through the cross-multiplication of electrical and mechanical time-frequency energy matrices. This matrix visualizes the transient modulation process of electromagnetic field distribution by mechanical deformation in the data space, effectively amplifying the implicit interaction characteristics that are difficult to detect under a single mode, and providing a reliable basis for identifying complex concurrent faults with strong concealment.

[0019] 3. In the diagnostic and judgment stage, this invention constructs a dual-stream network architecture that integrates physical statistical priors and deep features. The global macroscopic features extracted through principal component dimensionality reduction are cascaded with multi-scale local coupled textures automatically mined by convolutional networks. This complementary advantage of shallow physical constraints and deep implicit features effectively curbs the "overfitting" tendency of pure data-driven models under variable frequency and load conditions, reducing the false alarm and false negative rates in actual industrial settings.

[0020] 4. This invention breaks through the limitations of traditional static diagnostic models and constructs a complete closed loop for fault prediction and health management. By using the probability distribution of network output as pre-diagnosis and health management data, and combining it with the PHM database for periodic edge sample backflow and adaptive iteration, the system can not only accurately identify current faults, but also provide in-depth data support for predicting the degradation trend of equipment throughout its entire life cycle and for maintenance decisions. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a schematic diagram of the hardware layout of the multimodal signal synchronous acquisition and system of the present invention; Figure 3 This is a detailed flowchart of the adaptive wavelet filtering process based on frequency band energy distribution of the present invention; Figure 4 This is a schematic diagram illustrating the principle of constructing the two-dimensional cross-modal coupling matrix of the present invention; Figure 5 This is a schematic diagram of the dual-stream convolutional neural network architecture and feature fusion of the present invention; Figure 6 This is a flowchart of the intelligent response and model closed-loop adaptive evolution logic of the present invention. Detailed Implementation

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

[0023] Please see the appendix Figure 1 -Appendix Figure 6 This invention provides a motor fault diagnosis method based on waveform analysis, comprising the following steps: Synchronously acquire electrical and mechanical waveform signals during motor operation; The electrical waveform signal and the mechanical waveform signal are respectively subjected to adaptive wavelet filtering based on frequency band energy distribution to obtain the filtered electrical waveform signal and the filtered mechanical waveform signal, respectively. Statistical parameters of the filtered electrical waveform signal and the filtered mechanical waveform signal are extracted respectively to construct a global one-dimensional feature vector; the time-frequency energy matrix of the filtered electrical waveform signal and the filtered mechanical waveform signal are calculated respectively, and the time-frequency energy matrix of the electrical waveform signal and the transpose of the time-frequency energy matrix of the mechanical waveform signal are cross-multiplied and mapped to generate a two-dimensional cross-modal coupling matrix for characterizing electromechanical interaction features; The two-dimensional cross-modal coupling matrix and the global one-dimensional feature vector are input into a pre-trained two-stream convolutional neural network for feature fusion and classification decision, and the fault diagnosis result of the motor is output to realize the fault prediction and health management of the motor.

[0024] In implementing the above diagnostic methods, obtaining raw data that accurately and rigorously reflects the physical state of the motor is the foundation for the entire technical solution. Specifically, high-precision electrical sensors and mechanical vibration sensors are respectively configured in the motor's drive circuit and on the bearings or housing surfaces to achieve data coverage of electromagnetic and mechanical deformation states. For example, the electrical sensors can be Hall current / voltage sensors, and the mechanical vibration sensors can be piezoelectric accelerometers or triaxial vibration sensors.

[0025] Electromechanical coupling faults during motor operation are often accompanied by extremely high-frequency transient energy exchange. Even a slight time phase difference in the acquisition of multi-source signals can directly lead to a severe physical misalignment in the subsequently constructed cross-modal coupling matrix, causing the neural network to extract false coupling features. To address this, the system sets a fixed analysis sliding window length and uses a unified high-precision clock trigger source to coordinate the main frequency pulses of the multi-channel data acquisition modules, thereby ensuring absolute alignment of the electrical and mechanical channels on the physical time scale.

[0026] Within the specified analysis sliding window length, the system synchronously acquires the discrete electrical waveform signal and the mechanical waveform signal at a global synchronous sampling frequency that satisfies the sampling theorem. To ensure complete preservation of high-frequency details of the signals, a global synchronous sampling frequency is set... Strictly governed by the Nyquist-Shannon sampling theorem, its computational logic satisfies the following relationship: ; in, The global synchronization sampling frequency set for the system. This refers to the highest effective characteristic frequency component contained in the motor's operation and its potential fault-induced response. Setting this frequency parameter avoids frequency domain aliasing from the hardware level.

[0027] After the initial analog-to-digital conversion, the system discretizes the continuous analog signal within each given analysis sliding window period. The extracted discrete electrical waveform signal sequence is defined as follows: The synchronously extracted discrete mechanical waveform signal sequence is defined as Among them, symbols This represents the index of discrete sampling points, and its value range is... ,symbol This represents the total number of sampling points within the window, determined by both the length of the analysis sliding window and the global synchronous sampling frequency.

[0028] The obtained sequence and Having the same data dimensions and completely consistent occurrence timestamps establishes a reliable original data source for subsequent separation of background noise and construction of deep time-frequency energy interaction mapping.

[0029] In this embodiment, since the electrical and mechanical waveform signals acquired in the industrial field inevitably contain random pulse interference and equipment background noise, direct feature extraction would severely degrade diagnostic accuracy. Therefore, this system performs rigorous adaptive filtering processing on the synchronously acquired electrical and mechanical waveform signals respectively.

[0030] In this invention, the first step in the filtering process is to screen and reset the time-domain discreteness of the original data sequence. This is done using discrete acquisition sequences from any modal channel. For example, among which The system first calculates the global mean of the sequence. with standard deviation Its mathematical expression is as follows: ; ; in, This represents the total number of sampling points within the analysis sliding window.

[0031] Based on the established dispersion criterion, the system automatically identifies and replaces abnormal sampling points caused by abnormal disturbances. Specifically, it calculates the deviation of each sampling point in the sequence from the corresponding global mean. absolute error : ; When the absolute error Greater than the corresponding standard deviation preset multiple When, that is, satisfied Determine the corresponding sampling point These are abnormal sampling points. (Symbol) A preset multiple constant for the dispersion based on the system's fault tolerance limits. In specific engineering implementations, this is based on the Laida criterion (…). (Guideline), this constant The preferred value range is set between 2.5 and 3.5, and in this embodiment, it is preferably 3, so as to balance the rigor of outlier removal and the fidelity of normal waveform.

[0032] For sampling points identified as abnormal, they are replaced by linear interpolation of their adjacent normal sampling points to reconstruct a local smooth waveform. The update rules are as follows: ; After identifying and replacing abnormal sampling points, the corresponding global DC bias component (i.e., subtracting the global mean) is subtracted from the updated entire sequence. This process yields zero-mean sequences for the corresponding electrical and mechanical waveform signals, respectively. This operation completely eliminates the physical interference of baseline drift on subsequent frequency band energy calculations.

[0033] After obtaining the zero-mean sequences, discrete wavelet decomposition is performed on each zero-mean sequence using selected orthogonal wavelet basis functions. Through multi-scale decomposition operations, the original broadband signal is decoupled to different frequency bands, yielding corresponding low-frequency approximation coefficient sets and multiple decomposition layers. High-frequency detail coefficient .in, Indicates the decomposition level index. This indicates the position index of the high-frequency coefficients within the current decomposition layer.

[0034] Existing technologies typically employ a globally uniform threshold to forcibly truncate high-frequency coefficients for noise reduction. This mechanism indiscriminately eliminates weak high-frequency transient impacts associated with early faults. To overcome this limitation, this embodiment dynamically determines the adaptive shrinkage threshold of the high-frequency detail coefficients of each decomposition layer based on the frequency band energy distribution.

[0035] First, the system independently calculates the energy percentage of the high-frequency detail coefficients for each of the decomposition layers. Normalize the sum of squares of all high-frequency detail coefficients in the current layer to the global sum of high-frequency energies across all decomposition layers: ; Among them, symbols Indicates the first Layer The absolute amplitude of each high-frequency detail coefficient.

[0036] Subsequently, the median of the absolute values ​​of the high-frequency detail coefficients of each of the decomposition layers is extracted. ,Right now Combined with the total number of sampling points of the corresponding zero-mean sequence. With the corresponding energy percentage Calculate the adaptive shrinkage threshold specific to each of the decomposition layers. :

[0037] In this formula, The scaling factor is preferably between 1.5 and 3.0, used to control the rate at which the threshold decays with the energy percentage; the constant 0.6745 is a robust scaling factor for evaluating the intensity of background white noise under the standard normal distribution.

[0038] The above nonlinear calculation logic enables the adaptive shrinkage threshold. With the energy percentage The threshold decreases exponentially with increasing energy. In frequency bands rich in fault excitation energy, the threshold is automatically lowered to preserve key features; while in pure noise frequency bands with low energy content, the threshold is automatically raised to perform deep filtering.

[0039] Finally, the adaptive shrinkage threshold specific to each layer is utilized. For the high-frequency detail coefficients of each of the corresponding decomposition layers Soft thresholding is performed. The processed high-frequency detail coefficients and the corresponding low-frequency approximation coefficients are then reconstructed using inverse wavelet reconstruction. The output time series are the filtered electrical waveform signal and the filtered mechanical waveform signal, thus laying a data foundation with an extremely high signal-to-noise ratio for subsequent cross-modal correlation operations.

[0040] In this embodiment, after rigorous denoising preprocessing of the multi-source signals, the system performs multi-dimensional analysis of the data. The present invention extracts statistical parameters from the filtered electrical waveform signal and the filtered mechanical waveform signal respectively to construct a global one-dimensional feature vector, and constructs a two-dimensional cross-modal coupling matrix based on the time-frequency energy matrices of the filtered electrical waveform signal and the filtered mechanical waveform signal.

[0041] Specifically, for the filtered electrical waveform signal and the filtered mechanical waveform signal, the system independently calculates time-domain statistical parameters and frequency-domain parameters reflecting the signal fluctuation pattern. The time-domain parameters include root mean square (RMS), peak power factor (GPF), skewness, and kurtosis; the frequency-domain parameters include centroid frequency and frequency variance. The physical parameters extracted from the dual channels are concatenated and spliced ​​according to a predetermined sequence to construct an initial global high-dimensional feature vector.

[0042] Since directly inputting high-dimensional parameters with linear correlation into subsequent networks can lead to the curse of dimensionality and computational redundancy, the system performs principal component analysis (PCA) to reduce the dimensionality of the initial global high-dimensional feature vector. By solving for the eigenvalues ​​and eigenvectors of the initial high-dimensional feature covariance matrix, the principal directions are extracted and sorted in descending order of eigenvalue magnitude.

[0043] Let the total dimension of the initial features be... The th in descending order The feature values ​​are Before system calculation Cumulative variance contribution rate of each principal component : ; Selecting the cumulative variance contribution rate The minimum number of principal components that satisfy a set threshold (preferably 85%–95%, and specifically 90% in this embodiment). The corresponding projection components are extracted, and finally, a low-dimensional and uncorrelated global one-dimensional feature vector is constructed, denoted as . This feature vector achieves a high degree of condensation of the overall macroscopic operating indicators of the motor.

[0044] In this invention, to deeply explore the implicit physical correlation between the motor's operating state in the electrical and mechanical domains, the system simultaneously performs deep coupling analysis of time-frequency energy. Discrete wavelet packet decomposition is performed on both the filtered electrical waveform signal and the filtered mechanical waveform signal. Unlike conventional decomposition, which only expands the low-frequency approximation portion, wavelet packet decomposition performs symmetrical orthogonal partitioning of the high-frequency and low-frequency bands of the signal, thereby obtaining multiple frequency band signal subspaces with equal bandwidth resolution across the entire frequency band.

[0045] The system sets an analysis sliding window and calculates the sum of squares of the absolute values ​​of the amplitudes of all discrete sampling points in each frequency band signal subspace within the analysis sliding window, thereby obtaining the energy accumulation value of the corresponding frequency band. The energy accumulation values ​​of each frequency band are arranged row-wise according to the node order from low to high frequency, constructing the electrical time-frequency energy matrix corresponding to the filtered electrical waveform signal. And the mechanical time-frequency energy torque corresponding to the filtered mechanical waveform signal. These two matrices constitute the fundamental physical representation of the time-frequency energy matrix.

[0046] Subsequently, the system will use the electrical time-frequency energy matrix With the mechanical time-frequency energy matrix transpose matrix Perform cross-multiplication mapping. The operational rules are expressed as follows: ;

[0047] The generated matrix is ​​the two-dimensional cross-modal coupling moment. In this coupling matrix, the row and column indices correspond to specific frequency bands in the electrical and mechanical domains, respectively, and the value of each element in the matrix directly reflects the product interaction strength of the energy in these two cross-domain frequency bands.

[0048] When complex electromechanical anomalies occur inside the motor, such as air gap eccentricity or bearing spalling, its weak modulation frequency components will simultaneously excite local frequency band energy resonance in both the electrical windings and the mechanical rotor. Through the aforementioned cross-modal mapping operation, this scheme transforms independent one-dimensional sequences into a two-dimensional plane reflecting the cooperative correlation of transient energy modulation, accurately pinpointing the weak mutual modulation mechanism between signals. This provides a highly in-depth analytical framework for subsequent neural networks to avoid superficial features and directly pinpoint deep physical hazards.

[0049] In this embodiment, after completing the construction and mapping of the multidimensional feature space, the system inputs the two-dimensional cross-modal coupling matrix and the global one-dimensional feature vector into a pre-trained two-stream convolutional neural network for feature fusion and classification decision, and outputs the fault diagnosis result of the motor.

[0050] In this invention, the dual-stream convolutional neural network has two parallel data ingestion channels and internally includes convolutional layers for processing grid mapping data. The system directly inputs the two-dimensional cross-modal coupling matrix as a two-dimensional grid tensor into the convolutional layers of the dual-stream convolutional neural network.

[0051] The coupling characteristics induced by complex motor faults often exhibit physical phenomena of varying scales and shapes on the frequency band mapping plane. A single-sized receptive field cannot simultaneously accommodate large-scale macroscopic energy accumulation and localized minute frequency band abrupt changes. Therefore, the system employs multiple parallel convolution kernels with different receptive field sizes to perform multi-scale two-dimensional convolution operations on the two-dimensional mesh tensor.

[0052] In a preferred embodiment, the system constructs a convolutional layer with three parallel branches, with corresponding receptive field sizes set to 3×3, 5×5, and 7×7, respectively, to simultaneously capture local high-frequency modulation textures and macroscopic low-frequency energy contours.

[0053] Let the input two-dimensional mesh tensor be... , No. The convolution kernel weight matrix configured for each parallel convolution branch is as follows: The corresponding bias term is The multi-scale two-dimensional convolution operation satisfies the following mathematical relationship: ; Among them, symbols This represents a two-dimensional convolution operation. This represents an activation function that introduces nonlinear characteristics. For the corresponding number Feature maps extracted after convolutional operations on each branch. Through the above parallel mechanism, the network can simultaneously extract low-frequency coarse-grained contours and high-frequency fine-grained textures representing cross-domain energy interactions at the same layer, obtaining feature maps of multiple different scales.

[0054] To eliminate spatial redundancy in the feature maps at different scales and to impart translation invariance to the model for small signal shifts, the system applies local max pooling to the feature maps at each scale. Within a defined two-dimensional pooling sliding window, the maximum activation response value within the local receptive field is extracted, achieving spatial dimensionality reduction and compression of the feature tensor.

[0055] After completing the pooling operation, the system concatenates the pooling outputs of each parallel branch along the channel dimension and performs flattening. This operation unfolds the high-dimensional 3D feature tensor into a one-dimensional sequence, generating a one-dimensional vector of local depth features, denoted as . This vector highly condenses the deep implicit features under the synergistic effect of electromechanical modulation.

[0056] In this invention, the dual-stream convolutional neural network further includes a fully connected network component. Considering that deep features extracted under purely data-driven conditions are prone to becoming detached from physical reality when dealing with complex variable frequency and load conditions, the system implements a cross-modal, cross-layer feature cascading and fusion mechanism. The extracted local deep feature one-dimensional vector is then... The global one-dimensional feature vector extracted in the pre-stage Perform cascade fusion to form a hybrid feature vector Its cascade expression is: ; This hybrid feature vector complements the advantages of explicit physical statistical constraints at the macroscopic level and implicit mapping laws at the microscopic level, greatly enhancing the completeness of the representation space.

[0057] Subsequently, the system will use the hybrid feature vector The input is fed into the fully connected network, where the weight matrices of the neurons in the fully connected layers are used for high-order nonlinear spatial mapping. At the network classification terminal, the data is mapped via a normalized exponential function to a decision vector containing probabilities for various diagnostic states.

[0058] Suppose the total number of motor diagnostic status categories defined in the network is . For the first In this diagnostic state, the unnormalized logical value output by the end of the fully connected network is... Then its corresponding normalized probability value The calculation rules are as follows: ; in, Through this operation, the output of the fully connected network is mapped to a decision vector whose sum of probability values ​​is strictly 1.

[0059] Finally, the system iterates through the decision vector to find the state category with the highest probability as the fault diagnosis result for the motor. Simultaneously, the system extracts the probability distribution trend of each state in the decision vector, using it as pre-diagnosis and health management data to assess the degree of motor degradation. This processing method not only outputs a definite fault classification but also retains the confidence tendency for evolution to other fault states, thereby achieving high-confidence end-to-end decision-making from multi-source heterogeneous waveforms to specific health state labels.

[0060] In this embodiment, after completing the feature fusion and classification decision of the dual-stream convolutional neural network and outputting the fault diagnosis result of the motor, the system further establishes a set of intelligent response and closed-loop evolution mechanism based on confidence evaluation to realize closed-loop control and model adaptive iteration in industrial field.

[0061] In this invention, the system extracts the maximum probability value from the judgment vector output at the network endpoint and rigorously compares it with a set alarm judgment threshold. Let the extracted maximum probability value be... The alarm judgment threshold set by the system is The alarm detection threshold. As a rigorous mathematical evaluation boundary, it quantifies the lower confidence limit of the network output class label in practical engineering applications.

[0062] The system employs a deterministic logical branching mechanism. If the maximum probability value... Greater than or equal to the alarm determination threshold That is, satisfying the inequality conditions At this point, the network determines that the currently identified abnormal state has a very high degree of reliability. The system immediately triggers the fault alarm response procedure, sending a deterministic intervention command to the field control bus to execute mandatory safety actions such as reducing motor frequency and load or cutting off the drive circuit, thus preventing further spread of equipment deterioration from its physical source.

[0063] Simultaneously with triggering the fault alarm response procedure, the system performs structured retention of high-value data. The original input that triggered the alarm determination is bundled with key intermediate mappings, that is, the corresponding electrical waveform signal, the mechanical waveform signal, the extracted global one-dimensional feature vector, and the constructed two-dimensional cross-modal coupling matrix are stored as incremental samples in a PHM database dedicated to full lifecycle tracking.

[0064] The composite data accumulated in the database forms a complete case library containing low-level time series and high-level logic, specifically used for the periodic closed-loop updates of the dual-stream convolutional neural network to support the motor's fault prediction and health management mechanism. As data from harsh on-site operating conditions accumulates, the system periodically uses these real-world edge samples to fine-tune the network's weight parameters. This data-driven, closed-loop self-evolutionary mechanism endows the diagnostic system with lifelong learning capabilities in the face of long-term mechanical wear and aging and significant environmental drift, completely breaking the limitation of traditional static models whose accuracy rapidly declines after actual deployment.

[0065] Conversely, if the value of the maximum probability is... Less than the alarm judgment threshold That is, the conditions are met. If the current classification result does not reach a sufficient confidence level to trigger intervention, the system determines that the motor is in normal operation. Under this condition, the equipment maintains its current monitoring polling operation without taking any dimensionality reduction or shutdown intervention. This hard constraint effectively intercepts false alarms caused by transient random interference.

[0066] Thus, the method proposed in this invention completes the entire technical chain from multimodal high-frequency synchronous acquisition, cross-domain coupling matrix construction, deep feature fusion and judgment to closed-loop adaptive evolution of diagnostic models.

[0067] It should be understood that the above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the technical details of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A motor fault diagnosis method based on waveform analysis, characterized in that, Includes the following steps: Synchronously acquire electrical and mechanical waveform signals during motor operation; The electrical waveform signal and the mechanical waveform signal are respectively subjected to adaptive wavelet filtering based on frequency band energy distribution to obtain the filtered electrical waveform signal and the filtered mechanical waveform signal, respectively. Statistical parameters of the filtered electrical waveform signal and the filtered mechanical waveform signal are extracted respectively to construct a global one-dimensional feature vector; The time-frequency energy matrices of the filtered electrical waveform signal and the filtered mechanical waveform signal are calculated respectively, and the time-frequency energy matrix of the electrical waveform signal and the transpose of the time-frequency energy matrix of the mechanical waveform signal are cross-multiplied and mapped to generate a two-dimensional cross-modal coupling matrix for characterizing electromechanical interaction features. The two-dimensional cross-modal coupling matrix and the global one-dimensional feature vector are input into a pre-trained two-stream convolutional neural network for feature fusion and classification decision, and the fault diagnosis result of the motor is output to realize the fault prediction and health management of the motor.

2. The motor fault diagnosis method based on waveform analysis according to claim 1, characterized in that, Synchronously acquire electrical and mechanical waveform signals during motor operation, including: The analysis sliding window length is set, and based on a unified clock trigger source, within the analysis sliding window length, the discrete electrical waveform signal and the mechanical waveform signal are synchronously acquired at a global synchronous sampling frequency that satisfies the sampling theorem.

3. The motor fault diagnosis method based on waveform analysis according to claim 1, characterized in that, The electrical waveform signal and the mechanical waveform signal are respectively subjected to adaptive wavelet filtering based on frequency band energy distribution to obtain filtered electrical waveform signals and filtered mechanical waveform signals, respectively, including: Calculate the global mean and standard deviation of the electrical waveform signal and the mechanical waveform signal, identify and replace abnormal sampling points based on the set dispersion criterion, and subtract the corresponding global DC bias component to obtain the corresponding zero mean sequence respectively; Discrete wavelet decomposition is performed on each of the zero-mean sequences to obtain the corresponding low-frequency approximation coefficients and high-frequency detail coefficients of multiple decomposition layers. The adaptive shrinkage threshold of the high-frequency detail coefficients of each decomposition layer is determined based on the frequency band energy distribution; After shrinking the high-frequency detail coefficients of each decomposition layer using the adaptive shrinking threshold, wavelet reconstruction is performed with the corresponding low-frequency approximation coefficients to obtain the filtered electrical waveform signal and the filtered mechanical waveform signal, respectively.

4. The motor fault diagnosis method based on waveform analysis according to claim 3, characterized in that, The adaptive shrinkage threshold for the high-frequency detail coefficients of each decomposition layer is determined based on the frequency band energy distribution, including: Calculate the energy percentage of the high-frequency detail coefficients of each of the decomposition layers; The median of the absolute values ​​of the high-frequency detail coefficients of each decomposition layer is extracted. Combined with the total number of sampling points of the corresponding zero-mean sequence and the corresponding energy ratio, the adaptive shrinkage threshold specific to each decomposition layer is calculated, such that the adaptive shrinkage threshold decreases as the energy ratio increases.

5. The motor fault diagnosis method based on waveform analysis according to claim 1, characterized in that, Statistical parameters of the filtered electrical waveform signal and the filtered mechanical waveform signal are extracted respectively to construct a global one-dimensional feature vector, including: The time-domain statistical parameters and frequency-domain parameters of the filtered electrical waveform signal and the filtered mechanical waveform signal are calculated respectively, and then concatenated to construct an initial global high-dimensional feature vector; Principal component analysis is performed on the initial global high-dimensional feature vector to reduce its dimensionality. Principal components whose cumulative variance contribution rate meets a set threshold are selected to construct the low-dimensional and uncorrelated global one-dimensional feature vector.

6. The motor fault diagnosis method based on waveform analysis according to claim 1, characterized in that, Calculate the time-frequency energy matrices of the filtered electrical waveform signal and the filtered mechanical waveform signal, respectively, including: Discrete wavelet packet decomposition is performed on the filtered electrical waveform signal and the filtered mechanical waveform signal respectively to obtain multiple frequency band signal subspaces; An analysis sliding window is set, and within the analysis sliding window, the energy accumulation value of each frequency band signal subspace is calculated respectively. The electrical time-frequency energy matrix corresponding to the filtered electrical waveform signal and the mechanical time-frequency energy matrix corresponding to the filtered mechanical waveform signal are constructed as the time-frequency energy matrix.

7. The motor fault diagnosis method based on waveform analysis according to claim 1, characterized in that, The two-stream convolutional neural network includes convolutional layers; the process of processing the two-dimensional cross-modal coupling matrix through the convolutional layers includes: The two-dimensional cross-modal coupling matrix is ​​input as a two-dimensional mesh tensor into the convolutional layer of the two-stream convolutional neural network; Multiple parallel convolution kernels with different receptive field sizes are used to perform multi-scale two-dimensional convolution operations on the two-dimensional grid tensor to obtain feature maps of multiple different scales; Local max pooling is applied to the feature maps of different scales, and the pooling outputs are concatenated and flattened according to the channel dimension to generate a one-dimensional vector of local depth features.

8. The motor fault diagnosis method based on waveform analysis according to claim 7, characterized in that, The dual-stream convolutional neural network further includes a fully connected network; the step of performing feature fusion and classification decision, and outputting the fault diagnosis result of the motor, includes: The local depth feature one-dimensional vector and the global one-dimensional feature vector are concatenated and fused to form a hybrid feature vector; The hybrid feature vector is input into the fully connected network and mapped to a judgment vector containing the probabilities of various diagnostic states via a normalized exponential function. The state corresponding to the highest probability is selected as the fault diagnosis result, and the probability distribution of the judgment vector is used as the pre-diagnosis and health management data of the motor.

9. The motor fault diagnosis method based on waveform analysis according to claim 8, characterized in that, After outputting the fault diagnosis result of the motor, the method further includes: The value of the maximum probability is compared with the set alarm judgment threshold; If the value of the maximum probability is greater than or equal to the alarm judgment threshold, the fault alarm response program is triggered, and the corresponding electrical waveform signal, the mechanical waveform signal, the extracted global one-dimensional feature vector, and the constructed two-dimensional cross-modal coupling matrix that triggered the alarm are stored as samples in the PHM database for periodic closed-loop updates of the network to support the fault prediction and health management mechanism of the motor. If the value of the maximum probability is less than the alarm determination threshold, then the motor is determined to be in normal operating condition.

10. The motor fault diagnosis method based on waveform analysis according to claim 3, characterized in that, Identify and replace outlier sampling points based on a set dispersion criterion, including: Calculate the absolute error of each sampling point in the electrical waveform signal and the mechanical waveform signal from the corresponding global mean; When the absolute error is greater than a preset multiple of the corresponding standard deviation, the corresponding sampling point is determined to be an abnormal sampling point, and the abnormal sampling point is replaced by linear interpolation of its adjacent normal sampling points.

Images