Online Fault Diagnosis Method and System for Blower Generators Based on Multidimensional Data Fusion

Abstract

This application discloses an online fault diagnosis method and system for blower motors based on multi-dimensional data fusion, relating to the field of fault diagnosis. First, it synchronously collects stator current, casing vibration, and speed data of the blower motor, and extracts signal envelopes and energy features respectively. Next, it fuses vibration and speed information to construct a load state tensor characterizing the operating conditions, and uses a lightweight time-series network to mine the mapping relationship between this tensor and current fluctuations, accurately reconstructing the load interference component. Subsequently, it uses an adaptive differential decoupling mechanism to accurately remove this interference component from the current signal, obtaining a pure fault residual unaffected by operating condition fluctuations, and performs spectral feature analysis in conjunction with the windowed average speed value. Based on this, it dynamically adjusts the alarm threshold according to the load state and performs logical judgment on the fault feature energy. This effectively eliminates the interference of varying load conditions on the fundamental current wave, achieving high-precision fault identification and robust diagnosis of blower motors in complex operating environments.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis, and more specifically, to a method and system for online fault diagnosis of blower motors based on multidimensional data fusion. Background Technology

[0002] As a core power unit in modern industrial production systems, blower motors are widely used in ventilation and material conveying in key fields such as metallurgy, power, chemical industry, and mining. Because they typically operate continuously for extended periods in environments characterized by high temperature, high humidity, and strong vibration, the stator windings, rotor structure, and supporting bearings within the motor are highly susceptible to various latent faults due to fatigue, wear, or electromagnetic stress. Traditional methods of periodic manual inspections or shutdown maintenance not only significantly increase enterprise maintenance costs but also fail to detect transient anomalies during motor operation. Sudden damage can lead to unplanned shutdowns or even serious safety accidents. Therefore, integrating multi-source sensing data such as stator current and casing vibration to construct an online fault diagnosis system based on multi-dimensional data fusion, enabling real-time, comprehensive monitoring of the motor's health status, has become a crucial path to ensure the continuous operation and intelligent management of modern industrial systems.

[0003] However, in real-world industrial applications, the operating conditions of blower motors are highly dynamic and non-stationary, posing a significant challenge to accurate diagnosis. While existing multi-dimensional monitoring solutions attempt to combine current and vibration signals, they still face significant technical bottlenecks in handling variable load interference. Specifically, when the blower adjusts damper opening and closing or switches loads due to process requirements, significant non-fault fluctuations occur in the stator current signal. This fundamental energy shift and spectral dispersion caused by sudden load changes are highly similar in feature space to the modulation components generated by typical faults such as rotor bar breakage and air gap eccentricity. Because existing diagnostic algorithms generally lack in-depth extraction of load state characteristics and adaptive decoupling capabilities for current signals, the system struggles to accurately isolate the residual characteristics caused by minor faults in complex dynamic operating environments, easily misjudging normal load adjustments as organic damage to the motor itself. This high false alarm rate not only weakens the credibility of the diagnostic system but also severely restricts the intelligent operation and maintenance level of blower units under complex operating conditions.

[0004] Therefore, there is an urgent need for an optimized online fault diagnosis method and system for blower motors. Summary of the Invention

[0005] This application is made in order to solve the above-mentioned technical problems.

[0006] According to one aspect of this application, an online fault diagnosis method for blower motors based on multidimensional data fusion is provided, comprising: S1: Synchronously collect stator phase current data, casing vibration data and real-time speed data of the blower motor, and extract the envelope and calculate the energy characteristics of the stator phase current data, casing vibration data and real-time speed data to obtain the current envelope sequence, vibration energy sequence and window average speed value. S2: Perform channel-dimensional stitching and sliding time window slicing on the vibration energy sequence and window-averaged rotational speed values ​​to obtain the load state tensor; S3: Reconstruct the load components based on a lightweight temporal network from the load state tensor to obtain the estimated load disturbance components; S4: Adaptive differential decoupling of the current envelope sequence based on the estimated load disturbance component is used to obtain a pure fault residual sequence; S5: Based on the window average rotational speed value, perform residual spectrum feature analysis and fault feature extraction on the pure fault residual sequence to obtain the fault feature energy value; S6: Determine the dynamic alarm threshold based on the load state tensor, and make a logical comparison and judgment on the fault characteristic energy value to obtain the final diagnosis result.

[0007] According to another aspect of this application, a blower motor fault online diagnosis system based on multi-dimensional data fusion is provided, comprising: The data acquisition and preprocessing module is used to synchronously acquire stator phase current data, casing vibration data and real-time speed data of the blower motor, and to perform envelope extraction and energy feature calculation on the stator phase current data, casing vibration data and real-time speed data to obtain current envelope sequence, vibration energy sequence and window average speed value. The load state tensor construction module is used to perform channel-dimensional stitching and sliding time window slicing on the vibration energy sequence and the window average rotation speed value to obtain the load state tensor. The load disturbance reconstruction module is used to reconstruct the load components of the load state tensor based on a lightweight temporal network to obtain the estimated load disturbance components. An adaptive differential decoupling module is used to perform adaptive differential decoupling on the current envelope sequence based on the estimated load disturbance component to obtain a clean fault residual sequence. The fault feature extraction module is used to perform residual spectrum feature analysis and fault feature extraction on the pure fault residual sequence based on the window average speed value to obtain the fault feature energy value. The dynamic threshold diagnosis decision module is used to determine the dynamic alarm threshold based on the load state tensor and to make a logical comparison and judgment on the fault characteristic energy value to obtain the final diagnosis result.

[0008] Compared with existing technologies, this application provides a method and system for online fault diagnosis of blower motors based on multi-dimensional data fusion. First, it synchronously collects stator current, casing vibration, and speed data of the blower motor, and extracts signal envelopes and energy features respectively. Next, it fuses vibration and speed information to construct a load state tensor characterizing the operating conditions, and uses a lightweight time-series network to mine the mapping relationship between this tensor and current fluctuations, accurately reconstructing the load interference component. Subsequently, it uses an adaptive differential decoupling mechanism to accurately remove this interference component from the current signal, obtaining a pure fault residual unaffected by operating condition fluctuations, and performs spectral feature analysis in conjunction with the windowed average speed value. Based on this, it dynamically adjusts the alarm threshold according to the load state and performs logical judgment on the fault feature energy. This effectively eliminates the interference of variable load conditions on the fundamental current wave, achieving high-precision fault identification and robust diagnosis of blower motors in complex operating environments. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of an online fault diagnosis method for blower motors based on multidimensional data fusion, according to an embodiment of this application.

[0011] Figure 2 This is a data flow diagram of the online fault diagnosis method for blower motors based on multi-dimensional data fusion according to an embodiment of this application.

[0012] Figure 3 This is a flowchart of sub-step S1 of the online fault diagnosis method for blower motors based on multi-dimensional data fusion according to an embodiment of this application.

[0013] Figure 4 This is a flowchart of sub-step S2 of the online fault diagnosis method for blower motors based on multi-dimensional data fusion according to an embodiment of this application.

[0014] Figure 5 This is a flowchart of sub-step S3 of the online fault diagnosis method for blower motors based on multi-dimensional data fusion according to an embodiment of this application.

[0015] Figure 6 This is a flowchart of sub-step S4 of the online fault diagnosis method for blower motors based on multi-dimensional data fusion according to an embodiment of this application.

[0016] Figure 7This is a flowchart of sub-step S5 of the online fault diagnosis method for blower motors based on multi-dimensional data fusion according to an embodiment of this application.

[0017] Figure 8 This is a flowchart of sub-step S6 of the online fault diagnosis method for blower motors based on multi-dimensional data fusion according to an embodiment of this application.

[0018] Figure 9 This is a block diagram of an online fault diagnosis system for blower motors based on multi-dimensional data fusion, according to an embodiment of this application. Detailed Implementation

[0019] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0020] To address the problems mentioned above, this application proposes an online fault diagnosis method for blower motors based on multidimensional data fusion. Figure 1 This is a flowchart of an online fault diagnosis method for blower motors based on multidimensional data fusion, according to an embodiment of this application. Figure 2 This is a data flow diagram of the online fault diagnosis method for blower motors based on multi-dimensional data fusion according to an embodiment of this application. Figure 1 and Figure 2 As shown, the online fault diagnosis method for blower motors based on multi-dimensional data fusion includes the following steps: S1, synchronously collecting stator phase current data, casing vibration data, and real-time speed data of the blower motor, and performing envelope extraction and energy feature calculation on the stator phase current data, casing vibration data, and real-time speed data to obtain current envelope sequence, vibration energy sequence, and window average speed value; S2, performing channel dimension splicing and sliding time window slicing on the vibration energy sequence and window average speed value to obtain load state tensor; S3, reconstructing the load state tensor based on lightweight time-series network load components to obtain estimated load disturbance components; S4, performing adaptive differential decoupling on the current envelope sequence based on the estimated load disturbance components to obtain a clean fault residual sequence; S5, performing residual spectrum feature analysis and fault feature extraction on the clean fault residual sequence based on window average speed value to obtain fault feature energy value; S6, determining the dynamic alarm threshold based on the load state tensor, and performing logical comparison and judgment on the fault feature energy value to obtain the final diagnosis result.

[0021] In the above-mentioned online fault diagnosis method for blower motors based on multi-dimensional data fusion, step S1 involves simultaneously collecting stator phase current data, casing vibration data, and real-time speed data of the blower motor. Envelope extraction and energy feature calculation are then performed on the stator phase current data, casing vibration data, and real-time speed data to obtain a current envelope sequence, a vibration energy sequence, and a window-averaged speed value. It should be understood that because the operating conditions of blower motors in actual industrial scenarios are extremely complex, a single physical quantity cannot fully reflect the motor's health status, and directly processing the raw high-frequency sampled data would lead to excessive computational load and introduce a large amount of redundant noise. Therefore, this application simultaneously collects stator phase current data, casing vibration data, and real-time speed data of the blower motor and performs targeted envelope and energy feature extraction to construct a multi-dimensional panoramic view encompassing electromagnetic characteristics, mechanical vibration, and operating conditions, and transforms the raw high-dimensional time-series signal into a low-dimensional feature sequence that better characterizes the essence of the fault. This effectively solves the problem of unstructured data sources in the spatiotemporal dimensions, significantly reduces the data throughput of subsequent fusion algorithms, and retains key dynamic fault information, laying a solid data foundation for accurate diagnosis under varying load conditions.

[0022] In particular, in one specific embodiment, Figure 3 This is a flowchart of sub-step S1 of the online fault diagnosis method for blower motors based on multi-dimensional data fusion according to an embodiment of this application. Figure 3 As shown, step S1 includes: S11, using the time base of the stator phase current data as a reference, interpolating and resampling the casing vibration data and real-time speed data, and performing sliding window truncation to obtain a synchronous data frame containing stator current frames, casing vibration frames, and speed frames; S12, extracting the current envelope of the stator current frame based on Hilbert transform to obtain a current envelope sequence; S13, extracting vibration energy features of the casing vibration frame based on short-time root mean square to obtain a vibration energy sequence; S14, statistically averaging the sampling points within the speed frame to obtain a windowed average speed value.

[0023] Specifically, in step S11, using the time reference of the stator phase current data, interpolation resampling and sliding window truncation are performed on the casing vibration data and real-time speed data to obtain a synchronous data frame containing stator current frames, casing vibration frames, and speed frames. It should be understood that since different types of sensors are often controlled by independent data acquisition channels, their hardware clock sources have slight drift, and the sampling frequencies of the accelerometer and the current transformer are not consistent, direct combination will lead to phase misalignment of multi-source data on the time axis. Therefore, this application uses the sampling time of the stator phase current data as a reference to perform linear interpolation resampling and sliding window truncation on the casing vibration data and real-time speed data, thereby forcibly unifying the time reference of heterogeneous data and dividing it into synchronous data frames with fixed lengths and strict time alignment. This completely eliminates the timing asynchronous error between multi-source data, ensuring that current fluctuations, mechanical vibrations, and speed changes within the same time slice have a strict physical causal correspondence, thus avoiding the failure of fusion model training due to time misalignment.

[0024] Specifically, in one possible embodiment, after reading the original data stream, the sampling time sequence of the stator phase current data is first locked as the global master clock sequence. For the chassis vibration data with a high sampling rate, the linear weights between adjacent original sampling points are calculated based on the master clock time points to synthesize a new vibration amplitude that strictly corresponds to the current sampling time. For real-time speed data with a low sampling rate or non-uniform sampling, step interpolation or linear interpolation is used to complete the speed values ​​at intermediate times, ensuring that all data sources have the same length and time index. After resampling, a sliding window with a preset length (e.g., 1024 sampling points) and overlap rate (e.g., 50% overlap rate) is moved along the time axis. At each window position, current data, resampled vibration data, and speed data are simultaneously captured and combined to generate a synchronous data frame. This synchronous data frame contains stator current waveform segments, aligned chassis vibration waveform segments, and corresponding speed curve segments within the same time period.

[0025] Specifically, in step S12, the stator current frame is extracted using a Hilbert transform-based current envelope to obtain a current envelope sequence. It should be understood that when a blower motor experiences a fault or load fluctuation, its characteristic information mainly manifests as modulation of the fundamental amplitude of the stator current. The high-frequency power frequency carrier component in the original current signal masks these weak low-frequency modulation features. Therefore, this application further employs Hilbert transform technology to construct an analytical signal of the stator current and calculate its magnitude, thereby removing the high-frequency carrier component and accurately extracting the current envelope sequence reflecting the instantaneous changes in current amplitude. This significantly improves the signal-to-noise ratio of fault characteristics, clearly revealing the amplitude fluctuations originally hidden under the sinusoidal carrier wave. This allows subsequent algorithms to directly focus on the dynamic changes in current amplitude caused by broken bars, eccentricity, or sudden load changes, thereby reducing data dimensionality and improving diagnostic sensitivity.

[0026] Specifically, in one possible embodiment, a stator current frame segment from the synchronization data frame is first received and subjected to a Fast Fourier Transform (FFT) to convert it to the frequency domain. In the frequency domain, negative frequency components are set to zero, and positive frequency components are multiplied by a specific coefficient. Then, an inverse Fourier transform is performed to return to the time domain, thereby constructing a complex analytic signal with the original current signal as its real part and a 90-degree phase-shifted signal as its imaginary part. Next, the magnitude of this complex analytic signal is calculated point-by-point at each sampling moment to obtain an instantaneous amplitude curve that varies with time. To further compress the data volume and eliminate high-frequency noise, anti-aliasing low-pass filtering and downsampling operations are performed on the amplitude curve, ultimately outputting a smooth current envelope sequence. This sequence eliminates power frequency oscillation interference and retains only amplitude fluctuation information closely related to the motor's health status and load conditions.

[0027] Specifically, step S13 involves extracting vibration energy features from the casing vibration frame based on short-time root mean square (RMS) values ​​to obtain a vibration energy sequence. It should be understood that since casing vibration signals typically contain a large amount of high-frequency clutter caused by airflow turbulence and random friction, directly analyzing the original waveform makes it difficult to capture the macroscopic energy fluctuation trends caused by changes in load conditions. Therefore, this application further divides the casing vibration frame into several short-time subframes and calculates their RMS values ​​one by one to perform energy integration and smoothing on the high-frequency vibration signal, extracting a vibration energy sequence characterizing the intensity of mechanical vibration. This effectively filters out transient impact noise and irrelevant high-frequency details, transforming the complex vibration waveform into a low-frequency energy curve reflecting changes in mechanical load intensity, thereby providing stable and physically meaningful mechanical domain feature inputs for the subsequent construction of the load state tensor.

[0028] Specifically, in one possible embodiment, the captured chassis vibration frame data is further subdivided into multiple continuous and non-overlapping small time segments, i.e., short-time subframes, on the time axis. For each short-time subframe, all vibration acceleration sampling points within it are traversed. First, the amplitude of each sampling point is squared to quantify its instantaneous power. Then, the arithmetic mean of all squared values ​​within the subframe is calculated, representing the average power over that time period. Finally, the square root of this average power is taken to obtain the root mean square value of the subframe. By processing all subframes sequentially in chronological order, a series of calculated root mean square values ​​are connected to form a vibration energy sequence reflecting the evolution of vibration energy over time. This sequence macroscopically displays the fluctuations in the mechanical vibration intensity of the motor within the current window.

[0029] Specifically, in step S14, the sampling points within the speed frame are statistically averaged to obtain the window average speed value. It should be understood that the speed signal may fluctuate within a very short time due to measurement quantization errors or small torque pulsations in the motor, and subsequent fault characteristic frequency localization requires a stable operating condition reference point. Therefore, this application further performs an arithmetic average of all sampling point values ​​within the speed frame to eliminate the influence of instantaneous speed fluctuations and obtain the window average speed value within that time window. This provides a high-confidence scalar benchmark for subsequent calculation of fault characteristic frequencies, avoiding frequency localization deviations caused by using instantaneous speed, thereby ensuring the accuracy of spectrum analysis.

[0030] Specifically, in one possible embodiment, speed frame data from a synchronization data frame is acquired. This data contains a series of instantaneous speed values ​​sampled continuously within a time window. First, outlier detection is performed on the values ​​in this sequence to remove extreme points caused by sensor interference. Then, all remaining valid speed sampling points are summed and divided by the total number of sampling points to calculate the arithmetic mean. This arithmetic mean serves as the average speed value for the current time window, representing the macroscopic operating speed of the motor during this period, and is used as a key parameter input for calculating the characteristic frequency of rotor bar breakage or eccentricity faults in subsequent steps.

[0031] In the above-mentioned online fault diagnosis method for blower motors based on multi-dimensional data fusion, step S2 involves channel-dimensional splicing and sliding time window slicing of the vibration energy sequence and the window average speed value to obtain the load state tensor. It should be understood that, due to the deep electromechanical coupling between the vibration energy of the blower casing and the real-time speed during variable operating conditions, a single-dimensional time-domain feature cannot accurately depict the dynamic evolution of the motor under transient load switching. Therefore, this application further processes the vibration energy sequence and window average speed value through channel-dimensional splicing and sliding time window slicing to construct a spatiotemporal feature representation that can characterize the coupling characteristics of multi-dimensional physical fields, providing a load state tensor that conforms to tensor operation rules for subsequent nonlinear models. In this way, the limitations of single-sensor perception can be effectively compensated for through the complementary fusion of multi-source features, giving the interference signal caused by load fluctuations a traceable reference dimension, significantly improving the spatiotemporal alignment accuracy and state characterization capability of fault features in the time-series analysis process.

[0032] In particular, in one specific embodiment, Figure 4 This is a flowchart of sub-step S2 of the online fault diagnosis method for blower motors based on multi-dimensional data fusion according to an embodiment of this application. Figure 4 As shown, step S2 includes: S21, using the length of the vibration energy sequence as a reference, expanding the window average rotational speed value by time axis broadcasting to obtain a rotational speed expansion sequence of the same length as the vibration energy sequence; S22, stacking the vibration energy sequence and the rotational speed expansion sequence along the channel dimension to obtain a composite feature matrix; S23, slicing the composite feature matrix based on a sliding window with a preset step size and stacking and reconstructing it along the batch dimension to obtain the load state tensor.

[0033] Specifically, in step S21, the average rotational speed value of the window is expanded by time-axis broadcasting using the length of the vibration energy sequence as a benchmark to obtain a rotational speed expansion sequence of the same length as the vibration energy sequence. It should be understood that since the vibration energy sequence is a dynamic time-domain feature that changes over time, while the average rotational speed value of the window is a static scalar quantity obtained statistically within a specific time span, their dimensional structures are fundamentally different, making point-to-point feature fusion impossible in time-series calculations. Therefore, this application further uses the length of the vibration energy sequence as a benchmark to expand the average rotational speed value of the window by time-axis broadcasting, thereby expanding the static rotational speed mean into a dynamic rotational speed feature stream to obtain the rotational speed expansion sequence, thus achieving time-step alignment of heterogeneous physical quantities at the data structure level. This ensures that each vibration energy sampling point has a corresponding real-time operating speed as a background reference, eliminating information asymmetry in the feature space and providing a complete and homogeneous data foundation for constructing a composite feature matrix that can accurately describe the dynamic process of variable loads.

[0034] Specifically, in one possible embodiment, the total number of valid samples of the vibration energy sequence within a preset sampling window is first determined, serving as the length index for spatial expansion. Next, the average rotational speed scalar calculated within this window is read, and an array broadcast algorithm is used to fill this single value with an equivalent value in the new time vector space. In the specific operation process, an empty memory region with the same dimension as the vibration energy sequence is first allocated, and then the average rotational speed value is sequentially written to all time nodes in this region. Next, the generated sequence is subjected to integrity verification to ensure that the value in each time slice is strictly equal to the original window's average rotational speed. Finally, the generated equivalent sequence is defined as the rotational speed expansion sequence, thereby completing the conversion from static operating parameters to time-varying feature channels, enabling subsequent feature stacking operations to be executed smoothly within a unified tensor dimension.

[0035] Specifically, in step S22, the vibration energy sequence and the speed extension sequence are stacked in channel dimensions to obtain a composite feature matrix. It should be understood that since the casing vibration energy reflects the force response of the mechanical structure, while the speed characteristics determine the operating frequency distribution of the motor, simple independent analysis of the two in the feature space is insufficient to reveal the cross-influence of load mutations on electromagnetic characteristics. Therefore, this application further stacks the vibration energy sequence and the speed extension sequence in channel dimensions to deeply integrate physical information of different dimensions, constructing a multi-channel composite feature matrix. In this way, by increasing the feature channel density, the vibration intensity in the mechanical domain and the operating speed in the operating condition domain are forcibly coupled into the same matrix structure, enhancing the accuracy of the comprehensive description of complex operating conditions. This provides a high-dimensional composite feature matrix input with sufficient information density for subsequent lightweight time-series network identification of nonlinear load interference components.

[0036] Specifically, in one possible embodiment, vibration energy sequences and rotational speed extension sequences under the same time base are extracted and used as the raw materials for dual-channel input. First, the two sequences are mapped to column vector structures and juxtaposed horizontally according to a preset physical attribute order. Next, a matrix synthesis operation is performed, combining the two originally independent feature vectors into a two-dimensional data array with two columns, such that each row of the matrix corresponds to multiple physical attribute values ​​at the same sampling time. Subsequently, the data array is standardized and verified to confirm the consistency of length in the time dimension and the logical rationality of the numerical distribution of each channel. Finally, the two-dimensional data array with completed multi-dimensional feature alignment is determined as a composite feature matrix, thereby realizing a structured evolution from single-dimensional physical quantities to multi-dimensional correlated feature maps, providing a continuous and structured data carrier for subsequent slice reconstruction.

[0037] Specifically, in step S23, the composite feature matrix is ​​sliced ​​and reconstructed along the batch dimension using a sliding window with a preset step size to obtain the load state tensor. It should be understood that since the composite feature matrix contains a lengthy original feature stream, directly inputting it into the model would lead to low training efficiency and difficulty in capturing the short-term temporal evolution features of load changes. Furthermore, deep learning frameworks typically require the input features to have a three-dimensional tensor structure. Therefore, this application further slices and reconstructs the composite feature matrix along the batch dimension using a sliding window with a preset step size, thereby transforming the continuous feature stream into overlapping sample blocks with fixed time lengths to obtain the load state tensor, thus achieving data augmentation and structured transformation of the feature sequence. This effectively extracts the local dynamic patterns of load fluctuations and normalizes heterogeneous data streams into a tensor format suitable for neural network operations, significantly enhancing the sensitivity of the diagnostic algorithm to capturing complex operating trajectories and optimizing the allocation efficiency of computing resources.

[0038] Specifically, in one possible embodiment, a composite feature matrix is ​​obtained, and the width of the sliding window and the overlap step size between adjacent windows are set according to the input specifications of the deep learning model. For example, the window length is set to 128 sampling points, and the overlap step size is 64 sampling points, i.e., a 50% overlap rate. First, the sampling window is placed at the beginning of the matrix, and a local feature matrix is ​​extracted as the first temporal sample based on the window width. Then, the window is slid in the positive direction of the time axis according to the preset step size, and the slicing operation is performed cyclically to obtain continuous local feature sample blocks. Next, all the extracted two-dimensional local feature blocks are sequentially concatenated in the newly added batch dimension according to the extraction order. Subsequently, the generated collection is shape transformed to reconstruct a three-dimensional matrix structure that meets the input requirements of the deep neural network. Finally, the generated three-dimensional data entity containing batch, time step, and feature channel dimensions is defined as a load state tensor, thereby completing the final evolution from the original composite feature stream to a high-performance computing tensor.

[0039] In the above-mentioned online fault diagnosis method for blower motors based on multi-dimensional data fusion, step S3 involves reconstructing the load state tensor using a lightweight time-series network to obtain estimated load disturbance components. It should be understood that during operation, blower motors experience severe stator current fluctuations due to damper opening and closing or load switching. These non-fault fluctuations have a high degree of similarity in time-frequency characteristics to modulation components caused by physical damage such as rotor bar breakage, making accurate identification difficult with single-dimensional electrical parameter monitoring. Therefore, this application further reconstructs the load state tensor using a lightweight time-series network to deeply explore the electromechanical coupling mapping relationship between mechanical vibration energy, real-time speed, and electromagnetic torque disturbance, and predicts the current fluctuation trajectory generated by load regulation behavior from a cross-physics field perspective. This allows for the construction of estimated load disturbance components with physical causal logic, providing core data support for subsequent high-precision adaptive differential decoupling, effectively eliminating the masking effect of dynamic load disturbances on fault characteristics, and significantly improving the robustness of the diagnostic system under varying operating conditions.

[0040] In particular, in one specific embodiment, Figure 5 This is a flowchart of sub-step S3 of the online fault diagnosis method for blower motors based on multi-dimensional data fusion according to an embodiment of this application. Figure 5 As shown, step S3 includes: S31, using a one-dimensional convolutional neural network to extract local spatiotemporal features from the load state tensor to extract high-dimensional hidden layer features with implicit nonlinear dynamic features; S32, inputting the high-dimensional hidden layer features into a regression decoding network containing gated recurrent units to obtain block prediction tensors; S33, reconstructing the block prediction tensors by overlapping and adding sequences to obtain estimated load disturbance components.

[0041] Specifically, in step S31, a one-dimensional convolutional neural network is used to extract local spatiotemporal features from the load state tensor to extract high-dimensional hidden layer features containing implicit nonlinear dynamic characteristics. It should be understood that since the load state tensor integrates multi-dimensional nonstationary information such as high-frequency vibration and real-time rotational speed, the mechanical impact and aerodynamic turbulence laws it contains exhibit significant nonlinear dynamic characteristics. Directly using simple linear regression is insufficient to capture the deep correlation between these characteristics and the electromagnetic response. Therefore, this application further utilizes a one-dimensional convolutional neural network to extract local spatiotemporal features from the load state tensor. This allows multiple convolutional kernels with specific sensing fields to perform sliding scans along the time axis, capturing transient features of load changes at different scales, and constructing high-dimensional hidden layer features that characterize the essential properties of the load. This effectively filters out random measurement noise and redundant background information in the original tensor, transforming low-level sensor responses into high-level abstract feature maps. This provides nonlinear feature inputs with strong representational capabilities for subsequent time-dependent modeling, thus laying a precise physical foundation for cross-modal feature transformation.

[0042] Specifically, in one possible embodiment, a step-by-step convolution operation is first performed along the time dimension of the tensor using a one-dimensional convolutional kernel of a preset size (e.g., a kernel size of 3×1, a stride of 1, and 32 kernel channels). By calculating the multiplication and summation of the kernel weights and the local window data, a primary feature map reflecting the abrupt changes in mechanical vibration energy and the slope of rotational speed fluctuations is extracted. Next, a nonlinear activation function (e.g., ReLU) is applied immediately after the convolutional layer output to simulate the dynamic response curve under load abrupt changes by increasing the nonlinearity of the feature space. Subsequently, max pooling (with a pooling kernel size of 2×1) is applied to downsample the feature map, reducing the data dimensionality while preserving key dynamic features and enhancing robustness to time shifts. Finally, the results of alternating multi-layer convolution and pooling are flattened and reorganized to generate a high-dimensional hidden layer feature containing rich spatiotemporal context information, thus completing the transformation from raw load-sensing data to a high-dimensional feature representation.

[0043] Specifically, in step S32, the high-dimensional hidden layer features are input into a regression decoding network containing gated recurrent units to obtain a block prediction tensor. It should be understood that since the change in load state is a continuous temporal process with inertia and hysteresis, and the performance of load disturbance components in the current signal depends not only on the current mechanical state but also on the depth of the previous load evolution trajectory, this application further inputs the high-dimensional hidden layer features into a regression decoding network containing gated recurrent units. This utilizes a gating mechanism to adaptively retain or forget long-term and short-term load historical information, achieving end-to-end regression prediction of current fluctuation components under non-stationary operating conditions. This accurately characterizes the dynamic evolution logic in the electromechanical coupling system, overcomes the limitation of traditional regression models in handling time correlation, and ensures that the predicted current fluctuation segments are logically highly consistent with the load switching process, thereby significantly improving the real-time performance and dynamic response accuracy of load disturbance prediction and providing data assurance for obtaining high-confidence decoupling signals.

[0044] Specifically, in one possible embodiment, the lightweight temporal network requires offline pre-training before deployment, employing the Adam optimizer with an initial learning rate of 0.001, a mean squared error loss function, a training batch size of 64, and 100 iterations to ensure the fitting accuracy of the electromechanical mapping relationship. During prediction, the high-dimensional hidden layer sequence is first fed sequentially into a gated recurrent unit layer with 128 hidden layer nodes. A reset gate signal is calculated to determine the degree of ignoring historical load states, while an update gate signal controls the proportion of new input features retained at the current moment. Next, a linear mapping layer is appended after each hidden layer node to project the hidden state vector with temporal memory onto the current envelope prediction space. Subsequently, a regression output layer generates a current amplitude estimate corresponding to the current time window based on the evolution of the hidden states. Finally, all batches of generated local prediction values ​​are aggregated to form a set of block prediction tensors distributed along the time axis.

[0045] Specifically, step S33 involves reconstructing the block prediction tensor using an overlapping and summing sequence to obtain the estimated load interference component. It should be understood that, to accommodate the parallel computation requirements of neural networks, load prediction is performed on independent sliding window slices. Directly physically splicing these predicted blocks would lead to discontinuous step transitions or phase misalignments at the boundaries of adjacent windows. Therefore, this application further reconstructs the block prediction tensor using an overlapping and summing sequence to utilize the overlap information between windows to perform cross-window weighted fusion and smooth transition, restoring the discrete, fragmented prediction results to a continuous and single-valued time-domain waveform on the global time axis, thereby obtaining the estimated load interference component. This effectively eliminates the edge effects and numerical abrupt changes introduced by the block processing, ensuring that the generated interference reference signal possesses the smoothness and temporal coherence inherent in physical signals, and ensuring that the decoupled fault residual sequence is not mixed with pseudo-features caused by algorithm splicing, providing a clean and coherent interference elimination benchmark for achieving highly sensitive weak fault feature extraction.

[0046] Specifically, in one possible embodiment, firstly, the starting timestamp index of each prediction block in the global time series is determined based on the sliding step size and window width set during the preceding slicing. Next, each block is mapped to a full-length pre-allocated memory buffer according to the index order, ensuring that adjacent blocks have numerical overlap in overlapping regions. Subsequently, a weighted average calculation is performed on each overlapping time node; by calculating the number of overlaps and taking the average, minor amplitude differences between prediction values ​​from different windows are eliminated. Finally, normalization and smoothing processing of the global sequence is performed, and linear extrapolation is applied to edge regions to output an estimated load interference component of equal length to the original current signal and possessing continuous physical meaning. This prepares a high-precision reference waveform for subsequent adaptive differential decoupling operations in the current envelope sequence.

[0047] In the above-mentioned online fault diagnosis method for blower motors based on multi-dimensional data fusion, step S4 involves adaptive differential decoupling of the current envelope sequence based on the estimated load interference component to obtain a pure fault residual sequence. It should be understood that, in actual operation, the measured stator current envelope signal of a blower motor is a mixture of fundamental frequency changes caused by load fluctuations and weak modulation characteristics caused by potential faults. Directly analyzing the original envelope is easily masked by significant load changes, leading to misjudgment. Therefore, this application further implements adaptive differential decoupling processing on the measured current envelope sequence based on the estimated load interference component predicted by a deep learning model, thereby constructing a mathematical separation mechanism that can dynamically offset non-fault load fluctuations. This effectively removes the load interference component with the largest energy proportion from the signal, significantly improving the signal-to-noise ratio of weak fault features in the remaining signal. This provides a pure fault residual sequence containing only fault information as a basis for subsequent spectrum analysis, ensuring that the diagnostic system still has extremely high sensitivity under varying load conditions.

[0048] In particular, in one specific embodiment, Figure 6 This is a flowchart of sub-step S4 of the online fault diagnosis method for blower motors based on multi-dimensional data fusion according to an embodiment of this application. Figure 6 As shown, step S4 includes: S41, calculating the optimal amplitude ratio between the current envelope sequence and the estimated load disturbance component to obtain the adaptive correction coefficient; S42, performing linear correction on the estimated load disturbance component based on the adaptive correction coefficient, and calculating the absolute value of the difference between it and the current envelope sequence to obtain the original residual sequence; S43, performing noise suppression processing on the original residual sequence to obtain the clean fault residual sequence.

[0049] Specifically, in step S41, the optimal amplitude ratio between the current envelope sequence and the estimated load disturbance component is calculated to obtain an adaptive correction coefficient. It should be understood that since the deep learning model primarily learns the nonlinear mapping trend and waveform shape of current fluctuations caused by load changes during inference, the numerical amplitude output by the model may have an overall proportional deviation from the current measured amplitude due to differences in normalization parameters or sensor gain drift. Therefore, this application further employs least squares fitting to calculate the optimal amplitude ratio between the current envelope sequence and the estimated load disturbance component, thereby obtaining an adaptive correction coefficient that can quantify the difference between the two on the energy scale. This eliminates the systematic error introduced by the incomplete matching between the model prediction scale and the actual physical dimensions, ensuring that subsequent differential operations are performed on the same energy level, and avoiding the inability to completely cancel load disturbances or the introduction of artificial computational residuals due to amplitude mismatch.

[0050] Specifically, in one possible embodiment, the current envelope sequence is first used as the target vector, and the estimated load disturbance component is used as the reference vector. Next, an optimization function is constructed to minimize the difference between the two. This function aims to find a scalar factor such that the Euclidean distance or mean square error between the reference vector multiplied by this factor and the target vector is minimized. Subsequently, the optimization problem is solved using numerical methods (such as gradient descent with an iteration step size of 0.01) to calculate the optimal scaling factor within the time window. This calculation process is performed independently within each sliding window to accommodate possible changes in gain characteristics over different time periods (such as amplitude fluctuations within ±5% due to sensor temperature drift). Finally, this calculated scalar factor is determined as an adaptive correction coefficient and used for amplitude correction of the estimated components in subsequent steps.

[0051] Specifically, in step S42, the estimated load disturbance component is linearly corrected based on adaptive correction coefficients, and the absolute value of the difference between it and the current envelope sequence is calculated to obtain the original residual sequence. It should be understood that simple waveform subtraction may produce alternating positive and negative values, while fault characteristics in energy analysis focus on the magnitude of deviation from the normal reference, and correction coefficients for aligning the amplitude have already been obtained. Therefore, this application further utilizes adaptive correction coefficients to perform multiplicative linear correction on the estimated load disturbance component, and calculates the absolute value of the difference between the corrected sequence and the measured current envelope sequence. This physically simulates the process of subtracting the ideal healthy current from the total current signal, and uniformly maps all fluctuations deviating from the ideal waveform to non-negative deviations. In this way, an original residual sequence reflecting the degree of current signal distortion can be generated. The magnitude of the values ​​in this sequence directly characterizes whether there is any abnormal energy injection besides load fluctuations at the current moment, achieving preliminary explicitness of fault characteristics.

[0052] Specifically, in one possible embodiment, the calculated adaptive correction coefficients are first read and multiplied by the value of each sampling point in the estimated load interference component sequence to generate a corrected interference sequence that remains unchanged in waveform trend but is aligned with the measured signal in amplitude. Next, the measured current envelope sequence is aligned point-by-point with this corrected interference sequence, and a subtraction operation is performed at each time step to obtain a difference sequence with positive and negative signs. Subsequently, the absolute value of all values ​​in this difference sequence is taken, flipping negative deviations to positive amplitudes to ensure that all fluctuations deviating from the reference are recorded as positive energy contributions. Finally, the processed non-negative sequence is output as the original residual sequence.

[0053] Specifically, step S43 involves noise suppression processing of the original residual sequence to obtain a clean fault residual sequence. It should be understood that high-frequency random noise is inevitably introduced during data acquisition, transmission, and computation, and differential operations may generate non-physical high-frequency spikes or computational artifacts at local signal edges. These interference components affect the accuracy of subsequent frequency domain feature extraction. Therefore, this application further selects a smoothing filter with edge-preserving characteristics to perform noise suppression processing on the original residual sequence, thereby filtering out high-frequency and irregular random jitter components in the sequence while retaining the modulation waveforms of specific frequencies caused by faults such as rotor bar breakage or eccentricity, thus obtaining a clean fault residual sequence. This significantly improves the smoothness and signal-to-noise ratio of the residual signal, prevents high-frequency computational noise from forming false peaks in spectral analysis, and ensures that the finally extracted fault feature energy values ​​truly reflect the health status of the motor.

[0054] Specifically, in one possible embodiment, a sliding smoothing window of fixed length is first defined, such as a median filter or a five-point cubic smoothing filter with a window length of 5 sampling points. Next, this window is slid in steps along the time axis of the original residual sequence. At each window position, the smoothed estimate of the center time is calculated using the data within the window through a local polynomial fitting or sorting selection algorithm, replacing the original coarse value. This process effectively smooths out sharp random abrupt changes and difference calculation artifacts, while preserving the original modulation periodicity trend of the signal caused by the fault. Subsequently, the smoothing operation is performed at all points throughout the entire sequence. Finally, the filtered sequence is output as a clean fault residual sequence for use by the subsequent spectrum analysis module.

[0055] While the method described in the aforementioned embodiments involves low computational cost and is suitable for operating conditions with relatively smooth load changes, in high-dynamic scenarios with drastic load fluctuations, when calculating the optimal amplitude ratio between the current envelope sequence and the estimated load disturbance component to obtain the adaptive correction coefficient, if the adaptive correction coefficient is a single scalar coefficient, it is equivalent to assuming that there is only a simple linear gain relationship between the mechanical load model and the current response. As an idealized waveform generated by a deep learning model, although the estimated load disturbance component has the same shape as the current fluctuation caused by the load, it may have slight deviations from the actual physical current response in terms of phase lag and local frequency response characteristics. Therefore, if there is residual error when subtracting the above-mentioned disturbance component from the measured current envelope sequence (for example, ghost peaks appearing after subtraction due to phase misalignment), it may interfere with the accurate identification of the fault frequency.

[0056] In other words, when using methods such as least squares to calculate a single scalar coefficient, there is an assumption that the transfer function from the mechanical load to the stator current is a flat gain across the entire frequency band (pure resistive element). However, in a physical entity, the blower is a system that includes inertia (inductance, rotational inertia). The load change is transmitted to the current with time lag and low-pass filtering effect. A single scalar is difficult to completely eliminate the phase difference, and high-frequency interference components caused by phase misalignment may remain in the residual.

[0057] Specifically, in another embodiment, step S4 includes: constructing an adaptive transfer function convolution kernel vector, which is used to simultaneously compensate for amplitude differences, phase lag, and waveform distortion; constructing the estimated load disturbance components into a Toplitz data matrix and introducing a Tikhonov regularization term to construct an optimization objective function; solving the optimization objective function based on the current envelope sequence and the Toplitz data matrix to obtain the optimal convolution kernel vector; performing convolution filtering on the estimated load disturbance components based on the optimal convolution kernel vector, and then calculating the difference with the current envelope sequence to obtain a clean fault residual sequence.

[0058] Specifically, firstly, an adaptive transfer function convolution kernel vector is constructed to replace the single scalar coefficients. This kernel is a short-time finite impulse response filter kernel capable of describing the dynamic response characteristics from the load to the current. That is, it assumes the actual current disturbance components... The load components are predicted by the model. After passing through some unknown physical transfer function Convolutional, i.e. ,in:

[0059] It is a length of The convolution kernel vector can simultaneously compensate for amplitude differences, phase lag, and waveform distortion. Specifically, the convolution kernel length... This determines the maximum time lag range that the system can compensate for. If If the value is too small, it cannot cover the physical delay (i.e., electromechanical time constant) in the blower motor's response to load changes; if If the value is too large, it will introduce too many parameters to be estimated, leading to a surge in computation and increasing the risk of overfitting. Specifically, in one possible implementation, the kernel length... Based on the sampling frequency of the blower motor With the system's maximum expected response delay time To determine, that is, to satisfy For a typical industrial blower motor, its electromechanical response delay is typically between 10ms and 50ms. For example, when the sampling frequency is 2kHz, the convolution kernel length... The optimal number of sampling points is 50 to 100. This ensures that the convolution kernel has a sufficient temporal span to capture the complete transient response waveform while maintaining real-time computation.

[0060] To transform the convolution operation into matrix multiplication for solving, the estimated load disturbance components of the input sequence are constructed as a Toplitz data matrix. Specifically, assume the estimated load disturbance component sequence within the current sliding window is: ,in This is the window length. This is used to perform the convolution operation. Transform into matrix multiplication form The constructed Toplitz matrix It is a dimension A matrix. Each column of this matrix is ​​the input sequence. A time-delayed copy. The specific construction is as follows:

[0061] Correspondingly, the measured current envelope vector Extract as To maintain dimensional alignment. This allows for matrix processing, leveraging the parallel computing instruction sets of modern processors (such as the BLAS library) to accelerate the optimal solution process. Furthermore, to prevent overfitting (i.e., to prevent...), To accommodate the extreme oscillations caused by high-frequency noise, a Tikhonov regularization term is introduced, aiming to minimize the sum of squared prediction errors while constraining the energy of the convolution kernel: in It is the measured current envelope vector, obtained by arranging the current envelope sequence vertically in chronological order. This is a regularization intensity hyperparameter used to control the smoothness of the convolution kernel and prevent overfitting to measurement noise. In particular, in one possible embodiment, it is adaptively determined using the L-curve method or generalized cross-validation (GCV). The optimal value is found. However, in online diagnostic scenarios with extremely high real-time requirements, an empirical fixed value strategy is preferred. The value range is usually set within to The value should be between 0.01 and 0.01 (e.g., 0.01). This range is determined based on the signal-to-noise ratio (SNR) of the current signal. When there is strong electromagnetic interference at the scene, the value should be increased appropriately. Values ​​are used to enhance the algorithm's noise robustness. Let be the objective function to be minimized. The global optimal solution to this convex optimization problem can be obtained directly using analytical methods.

[0062] Specifically, a matrix is ​​first constructed based on the estimated load disturbance components. The optimal convolution kernel is then calculated as follows:

[0063] And the pure fault residual sequence is calculated as follows: .

[0064] In other words, by solving for the optimal convolution kernel, the system automatically learns the minute time delay (determined by the peak position of the convolution kernel) and frequency response attenuation (determined by the spectrum of the convolution kernel) from the load prediction to the actual current under the current operating conditions, based on the dynamic time-varying characteristics of electromechanical coupling. This allows the convolution kernel to automatically compensate for the subtle millisecond-level time difference between the deep learning model output and the sensor's measured data. Even if the predicted load waveform is slightly harder or softer than the actual current waveform, the filtering effect of the convolution kernel can shape it into a form that most closely resembles the real interference. In this way, the background noise in the pure fault residual sequence is suppressed to the greatest extent, and the weak high-frequency ripples caused by the fault (such as rotor bar breakage characteristics) are no longer masked by the residual edges of large load fluctuations.

[0065] In the above-mentioned online fault diagnosis method for blower motors based on multi-dimensional data fusion, step S5 involves performing residual spectrum feature analysis and fault feature extraction on the pure fault residual sequence based on the window average speed value to obtain the fault feature energy value. It should be understood that although the decoupled pure fault residual sequence eliminates load interference, it is still a discrete waveform in the time domain, making it impossible to directly distinguish between fluctuations caused by random noise and periodic modulations caused by specific fault mechanisms. Furthermore, the variable frequency operation characteristics of the blower motor cause the fault feature frequency to dynamically drift with the speed. Therefore, this application further uses the real-time acquired window average speed value as the frequency benchmark, maps the time-domain residual signal to the frequency domain, and locks a specific fault frequency band for energy quantification. This transforms the abstract time-domain fluctuations into a fault severity index with clear physical meaning, namely, the fault feature energy value. This overcomes the failure problem of the fixed-frequency detection method under variable speed conditions and achieves accurate quantitative assessment of specific fault types.

[0066] In particular, in one specific embodiment, Figure 7 This is a flowchart of sub-step S5 of the online fault diagnosis method for blower motors based on multi-dimensional data fusion according to an embodiment of this application. Figure 7 As shown, step S5 includes: S51, performing frequency domain transformation and power spectral density estimation on the pure fault residual sequence to obtain the residual power spectrum; S52, based on the physical parameters of the blower motor, performing fault characteristic frequency localization based on the window average speed value to obtain the target fault center frequency; S53, based on the target fault center frequency and the preset tolerance bandwidth, performing narrowband energy accumulation integration on the residual power spectrum to obtain the fault characteristic energy value.

[0067] Specifically, step S51 involves performing frequency domain transformation and power spectral density estimation on the pure fault residual sequence to obtain the residual power spectrum. It should be understood that since the pure fault residual sequence exhibits a waveform with amplitude varying over time in the time domain, the subtle fault features are often masked by broadband background noise, and the time-domain signal is difficult to intuitively reveal the distribution pattern of signal energy at different frequency components. Therefore, this application further employs signal processing techniques to transform the time-domain residual sequence to the frequency domain and calculate its power spectral density to highlight the periodic fault modulation components while suppressing non-periodic random noise interference, thus obtaining the residual power spectrum. This significantly improves the signal-to-noise ratio of the fault features, providing a clear and high-resolution energy distribution map for subsequent accurate location and extraction of fault features in complex spectra.

[0068] Specifically, in one possible embodiment, a Hanning or Hamming window function matching the length of the residual sequence is first selected and multiplied point-by-point with the time-domain residual sequence to suppress sidelobe leakage effects in spectral analysis. Next, a discrete Fourier transform is performed on the windowed sequence to calculate the complex spectral coefficients at each frequency point. Subsequently, the squares of the moduli of these complex coefficients are calculated and normalized by dividing by the frequency resolution and the sequence length, thus obtaining a series of values ​​characterizing the power intensity of each frequency component. Arranging these values ​​in frequency order constitutes the residual power spectrum reflecting the frequency domain energy distribution of the residual signal.

[0069] Specifically, in step S52, based on the physical parameters of the blower motor, the window average speed value is used to locate the fault characteristic frequency based on the speed to obtain the target fault center frequency. It should be understood that since the fault characteristic frequency of an induction motor (such as the sideband frequency caused by a broken rotor bar) has a strict kinematic coupling relationship with the motor's real-time operating speed and slip, this characteristic frequency changes constantly in scenarios where the blower frequently adjusts its airflow. Therefore, this application further combines the inherent physical structural parameters of the blower motor and uses the real-time calculated window average speed value to deduce the current target fault center frequency, thereby establishing a dynamic frequency tracking mechanism. This ensures that the diagnostic algorithm always focuses on the correct frequency band where a fault may occur at the current speed, avoiding frequency alignment deviations caused by speed changes, thus achieving full-speed-domain fault locking for the variable frequency drive motor.

[0070] Specifically, in one possible embodiment, pre-stored motor physical parameters, including the number of pole pairs, the number of rotor bars, and the rated power supply frequency, are first read. Simultaneously, the average rotational speed within the current time window is acquired and converted to synchronous speed-frequency. Next, the current slip rate is calculated based on the difference between the synchronous speed and the actual rotational speed. Subsequently, the slip rate and the fundamental power supply frequency are substituted into a specific fault characteristic frequency calculation model, such as calculating the positions of the sideband components on both sides of the fundamental frequency for rotor bar breakage faults. Finally, the precise Hertz value under the current operating condition is calculated and marked as the target fault center frequency that requires focused monitoring.

[0071] Specifically, in step S53, based on the target fault center frequency and a preset tolerance bandwidth, the residual power spectrum is integrated using narrowband energy accumulation to obtain the fault characteristic energy value. It should be understood that due to small fluctuations in rotational speed, limitations in spectral resolution, and the spectral leakage effect of the algorithm itself in actual measurements, the actual fault energy is often not concentrated at a single frequency point, but rather dispersed in a narrowband region near the center frequency. Therefore, this application further uses the calculated target fault center frequency as a benchmark, combined with a preset tolerance bandwidth, to define an energy integration interval, and accumulates the residual power spectrum within this interval to capture the fault characteristic energy value leaked from the fault signal to adjacent frequency points. This improves the robustness of feature extraction, prevents the omission of crucial fault energy due to small errors in frequency positioning, and thus obtains a stable and accurate fault characterization quantity.

[0072] Specifically, in one possible embodiment, the discrete index position of the target fault center frequency in the residual power spectrum is first determined based on the system's sampling rate and spectral resolution. Next, according to a pre-set frequency tolerance bandwidth (e.g., ±2Hz of the target fault center frequency or ±3% of the center frequency), the boundary indices on both sides of the center frequency are calculated, thus defining a specific frequency band. Subsequently, all discrete frequency points within this frequency band are traversed, their corresponding power spectral density values ​​are read, and summed. The sum obtained represents the total energy intensity caused by the specific fault under the current operating speed, and is output as the final fault characteristic energy value.

[0073] In the above-mentioned online fault diagnosis method for blower motors based on multi-dimensional data fusion, step S6 involves determining a dynamic alarm threshold based on the load state tensor and performing logical comparison and judgment on the fault feature energy values ​​to obtain the final diagnostic result. It should be understood that, due to the extremely complex variable load conditions faced by blower motors in actual industrial applications, a fixed fault alarm threshold is difficult to balance between sensitivity and false alarm rate. Furthermore, even with differential decoupling, a small amount of nonlinear computational noise may still remain in the residual sequence during extremely drastic load fluctuations. Therefore, this application further utilizes a load state tensor containing historical vibration and speed information to assess the severity of the current operating conditions, and dynamically adjusts the judgment criteria based on this, performing adaptive logical judgment on the extracted fault feature energy values ​​to obtain the final diagnostic result. This ensures that the system maintains high sensitivity to capture early, weak faults when operating conditions are stable, while automatically raising the threshold to suppress false alarms caused by transient disturbances during drastic changes in operating conditions, thereby significantly improving the robustness and reliability of the diagnostic system across the entire operating range.

[0074] In particular, in one specific embodiment, Figure 8 This is a flowchart of sub-step S6 of the online fault diagnosis method for blower motors based on multi-dimensional data fusion according to an embodiment of this application. Figure 8 As shown, step S6 includes: S61, calculating the norm of the vibration component in the load state tensor to quantify the working condition intensity, and combining it with a preset benchmark threshold to perform nonlinear correction on the working condition intensity to obtain a dynamic alarm threshold; S62, performing instantaneous state logic comparison between the fault characteristic energy value and the dynamic alarm threshold to obtain an instantaneous fault flag; S63, performing confidence statistics and smoothing filtering on the instantaneous fault flag to output the final diagnostic result.

[0075] Specifically, in step S61, the norm of the vibration component in the load state tensor is calculated to quantify the working condition intensity, and the working condition intensity is nonlinearly corrected by combining it with a preset benchmark threshold to obtain a dynamic alarm threshold. It should be understood that since the load state tensor is a high-dimensional data structure, directly using it for adjusting a scalar threshold lacks a clear physical correspondence. However, the energy intensity of the vibration signal is directly related to the severity of airflow turbulence and mechanical impact, making it the best indicator for measuring background noise. Therefore, this application further extracts vibration channel data from the load state tensor, calculates its norm or statistical variance to generate a quantized scalar, and uses a nonlinear function to map this working condition intensity scalar to the incremental part of the threshold, superimposing it onto the preset benchmark threshold to obtain the dynamic alarm threshold. This establishes an adaptive mechanism where "the more severe the working condition, the higher the tolerance," allowing the alarm threshold curve to closely follow the envelope of load fluctuations, accurately excluding broadband noise caused by damper operation from the fault judgment range, and preventing false alarms triggered by increased background noise.

[0076] Specifically, in one possible embodiment, a feature slice representing the chassis vibration signal is first extracted from the load state tensor. This slice contains the vibration amplitude values ​​of all sampling points within the current time window. Next, Frobenius norm or root mean square calculation is performed on this vibration slice data to compress the multidimensional waveform data into a single value, representing the current operating condition intensity or turbulence disturbance level. Subsequently, a pre-calibrated baseline threshold (e.g., set to three times the statistical average of the characteristic energy of the motor under rated load and healthy conditions) is read, and the calculated operating condition intensity value is multiplied by a sensitivity adjustment coefficient. To simulate the nonlinear effect of noise on the threshold, a nonlinear transformation, such as square root or exponential operation, is performed on the adjusted intensity value, and the transformation result is added as a floating increment to the baseline threshold. Finally, the dynamic alarm threshold at that moment is output, which is numerically higher than the baseline value and monotonically increases with increasing vibration intensity.

[0077] Specifically, step S62 involves performing an instantaneous state logic comparison between the fault characteristic energy value and the dynamic alarm threshold to obtain an instantaneous fault flag. It should be understood that since both the fault characteristic energy value and the dynamic alarm threshold are analog quantities that change continuously with time, explicit binary decision-making is necessary to convert the continuous physical characteristics into discrete state logic that can be processed by a computer. Therefore, this application further compares the real-time calculated fault characteristic energy value with the dynamic alarm threshold that is currently in effect at each independent time step, and outputs a high-level or low-level instantaneous fault flag based on the comparison result. This simplifies the complex assessment of fault severity to an instantaneous "present" or "absent" logical state, providing standardized Boolean input data for subsequent statistical analysis and anti-jitter processing. This achieves a crucial transformation from the analog signal domain to the digital logic domain, giving the diagnostic process clear triggering conditions.

[0078] Specifically, in one possible embodiment, the fault characteristic energy value is first used as the comparison number, and the dynamic alarm threshold is used as the comparison benchmark, both of which are strictly synchronized on the time axis. Next, comparator logic operations are performed: if the current fault characteristic energy value is strictly greater than the dynamic alarm threshold, it indicates that the energy contained in the residual signal has exceeded the allowable noise range under the current operating conditions. In this case, the comparator outputs logic 1, representing that a potential fault has been detected in the current frame; conversely, if the characteristic energy value is less than or equal to the threshold, it indicates that the signal fluctuation is within the allowable range, and the comparator outputs logic 0, representing that the current frame is normal. This comparison process is repeated in each processing cycle, thereby generating a sequence of instantaneous fault flags that continuously changes over time. This sequence records the original decision state of each frame of data.

[0079] Specifically, step S63 involves performing confidence statistics and smoothing filtering on the instantaneous fault flags to output the final diagnostic result. It should be understood that due to electromagnetic interference or computational boundary effects, sporadic, discontinuous jump pulses may appear in the instantaneous fault flag sequence. Directly relying on a single frame result for alarm would lead to frequent system output flickering and a lack of stability. Therefore, this application further establishes a historical state buffer queue with a fixed length, performs sliding window statistics on the instantaneous fault flags, calculates the frequency percentage of fault states, and only locks the final diagnostic result when the percentage exceeds a high confidence setting value. This introduces a debouncing mechanism similar to a hysteresis comparator, effectively filtering out isolated spikes caused by random noise, ensuring that an alarm is triggered only when the fault characteristics persist and stabilize over a period of time, thereby outputting a smooth, reliable, and industrially stable final diagnostic result.

[0080] Specifically, in one possible embodiment, a first-in, first-out (FIFO) circular buffer is first established, with a capacity set to several frames (e.g., ten consecutive time windows). Whenever a new transient fault flag is generated, it is pushed into the head of the buffer, while the oldest flag is pushed out. Next, all storage bits in the buffer are traversed, the total number of logic 1s is counted, and its percentage in the total buffer capacity is calculated. This percentage is then compared to a preset confidence threshold (e.g., 80%). Only when the percentage of fault flags in the buffer exceeds this threshold is the algorithm diagnosed as a fault and the final diagnostic state is set to fault; otherwise, even if individual fault flags exist, the system maintains a normal output state. Finally, the smoothed diagnostic conclusion is encoded into a control command and sent to the host computer or controller.

[0081] In summary, the online fault diagnosis method for blower motors based on multi-dimensional data fusion, as described in this application, is explained. First, it synchronously collects stator current, casing vibration, and speed data of the blower motor, and extracts signal envelopes and energy features respectively. Next, it fuses vibration and speed information to construct a load state tensor characterizing the operating conditions, and uses a lightweight time-series network to mine the mapping relationship between this tensor and current fluctuations, accurately reconstructing the load interference component. Subsequently, it utilizes an adaptive differential decoupling mechanism to accurately remove this interference component from the current signal, obtaining a pure fault residual unaffected by operating condition fluctuations, and performs spectral feature analysis in conjunction with the windowed average speed value. Based on this, the alarm threshold is dynamically adjusted according to the load state, and logical judgments are made on the fault feature energy. This effectively eliminates the interference of varying load conditions on the fundamental current wave, achieving high-precision fault identification and robust diagnosis of the blower motor in complex operating environments.

[0082] Figure 9 This is a block diagram of an online fault diagnosis system for blower motors based on multi-dimensional data fusion, according to an embodiment of this application. Figure 9 As shown, the blower motor fault online diagnosis system 100 based on multi-dimensional data fusion according to an embodiment of this application includes: a data acquisition and preprocessing module 110, used to simultaneously acquire stator phase current data, casing vibration data, and real-time speed data of the blower motor, and to perform envelope extraction and energy feature calculation on the stator phase current data, casing vibration data, and real-time speed data to obtain a current envelope sequence, a vibration energy sequence, and a window average speed value; a load state tensor construction module 120, used to perform channel dimension splicing and sliding time window slicing on the vibration energy sequence and the window average speed value to obtain a load state tensor; and a load interference reconstruction module 120. Block 130 is used to reconstruct the load components of the load state tensor based on a lightweight time-series network to obtain the estimated load disturbance components; Adaptive differential decoupling module 140 is used to perform adaptive differential decoupling on the current envelope sequence based on the estimated load disturbance components to obtain a clean fault residual sequence; Fault feature extraction module 150 is used to perform residual spectrum feature analysis and fault feature extraction on the clean fault residual sequence based on the window average speed value to obtain the fault feature energy value; Dynamic threshold diagnosis decision module 160 is used to determine the dynamic alarm threshold based on the load state tensor and perform logical comparison and judgment on the fault feature energy value to obtain the final diagnosis result.

[0083] As described above, the blower motor fault online diagnosis system 100 based on multi-dimensional data fusion according to the embodiments of this application can be implemented in various wireless terminals, such as servers with a blower motor fault online diagnosis algorithm based on multi-dimensional data fusion. In one possible implementation, the blower motor fault online diagnosis system 100 based on multi-dimensional data fusion according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the blower motor fault online diagnosis system 100 based on multi-dimensional data fusion can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the blower motor fault online diagnosis system 100 based on multi-dimensional data fusion can also be one of many hardware modules of the wireless terminal.

[0084] Alternatively, in another example, the blower motor fault online diagnosis system 100 based on multidimensional data fusion and the wireless terminal can also be separate devices, and the blower motor fault online diagnosis system 100 based on multidimensional data fusion can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with the agreed data format.

[0085] Here, those skilled in the art will understand that the specific operations of each step in the above-described online fault diagnosis system for blower motors based on multi-dimensional data fusion have been referenced above. Figures 1 to 8The online fault diagnosis method for blower motors based on multidimensional data fusion has been described in detail, and therefore, its repeated description will be omitted.

Claims

1. A method for online fault diagnosis of blower motors based on multi-dimensional data fusion, characterized in that, include: S1: Synchronously collect stator phase current data, casing vibration data and real-time speed data of the blower motor, and extract the envelope and calculate the energy characteristics of the stator phase current data, casing vibration data and real-time speed data to obtain the current envelope sequence, vibration energy sequence and window average speed value. S2: Perform channel-dimensional stitching and sliding time window slicing on the vibration energy sequence and window-averaged rotational speed values ​​to obtain the load state tensor; S3: Reconstruct the load components based on a lightweight temporal network from the load state tensor to obtain the estimated load disturbance components; S4: Adaptive differential decoupling of the current envelope sequence based on the estimated load disturbance component is used to obtain a pure fault residual sequence; S5: Based on the window average rotational speed value, perform residual spectrum feature analysis and fault feature extraction on the pure fault residual sequence to obtain the fault feature energy value; S6: Determine the dynamic alarm threshold based on the load state tensor, and make a logical comparison and judgment on the fault characteristic energy value to obtain the final diagnosis result.

2. The online fault diagnosis method for blower motors based on multi-dimensional data fusion according to claim 1, characterized in that, Step S1 includes: Using the time base of the stator phase current data as a reference, the casing vibration data and real-time speed data are interpolated, resampled, and truncated by a sliding window to obtain a synchronous data frame containing stator current frame, casing vibration frame, and speed frame. The current envelope sequence is obtained by extracting the current envelope from the stator current frame based on Hilbert transform. Vibration energy features based on short-time root mean square are extracted from the casing vibration frames to obtain the vibration energy sequence; The sampling points within the rotation speed frame are statistically averaged to obtain the window average rotation speed value.

3. The online fault diagnosis method for blower motors based on multi-dimensional data fusion according to claim 1, characterized in that, Step S2 includes: Using the length of the vibration energy sequence as a benchmark, the window average rotational speed value is extended by time-axis broadcasting to obtain a rotational speed extension sequence of the same length as the vibration energy sequence. The vibration energy sequence and the speed extension sequence are stacked by channel dimensions to obtain a composite feature matrix; The composite feature matrix is ​​sliced ​​and reconstructed by stacking along the batch dimension using a sliding window with a preset step size to obtain the load state tensor.

4. The online fault diagnosis method for blower motors based on multi-dimensional data fusion according to claim 1, characterized in that, Step S3 includes: A one-dimensional convolutional neural network is used to extract local spatiotemporal features of the load state tensor in order to extract high-dimensional hidden layer features of implicit nonlinear dynamic features. High-dimensional hidden layer features are input into a regression decoding network containing gated recurrent units to obtain block prediction tensors; The block prediction tensor is reconstructed by overlapping and summing sequences to obtain the estimated load disturbance components.

5. The online fault diagnosis method for blower motors based on multi-dimensional data fusion according to claim 1, characterized in that, Step S4 includes: The optimal amplitude ratio between the current envelope sequence and the estimated load disturbance component is calculated to obtain the adaptive correction coefficient; The estimated load disturbance component is linearly corrected based on the adaptive correction coefficient, and the absolute value of the difference between it and the current envelope sequence is calculated to obtain the original residual sequence. The original residual sequence is subjected to noise suppression processing to obtain a clean fault residual sequence.

6. The online fault diagnosis method for blower motors based on multi-dimensional data fusion according to claim 1, characterized in that, Step S4 includes: Construct an adaptive transfer function convolution kernel vector, which is used to simultaneously compensate for amplitude differences, phase lag, and waveform distortion. The estimated load disturbance components are constructed as Toplitz data matrices, and Tikhonov regularization terms are introduced to construct the optimization objective function; Based on the current envelope sequence and the Toplitz data matrix, the optimization objective function is solved to obtain the optimal convolution kernel vector; After performing convolution filtering on the estimated load disturbance components based on the optimal convolution kernel vector, the difference between the estimated load disturbance components and the current envelope sequence is calculated to obtain the pure fault residual sequence.

7. The online fault diagnosis method for blower motors based on multi-dimensional data fusion according to claim 1, characterized in that, Step S5 includes: Frequency domain transformation and power spectral density estimation are performed on the clean fault residual sequence to obtain the residual power spectrum; Based on the physical parameters of the blower motor, the fault characteristic frequency is located based on the average speed value of the window to obtain the target fault center frequency. Based on the target fault center frequency and the preset tolerance bandwidth, the residual power spectrum is integrated by narrowband energy accumulation to obtain the fault characteristic energy value.

8. The online fault diagnosis method for blower motors based on multi-dimensional data fusion according to claim 1, characterized in that, Step S6 includes: The norm of the vibration component in the load state tensor is calculated to quantify the working condition intensity, and the working condition intensity is nonlinearly corrected by a preset benchmark threshold to obtain a dynamic alarm threshold. The instantaneous state logic comparison between the fault characteristic energy value and the dynamic alarm threshold is performed to obtain the instantaneous fault flag; The confidence level of the instantaneous fault flag is statistically analyzed and smoothed by filtering to determine the final diagnostic result.

9. A blower motor fault online diagnosis system based on multi-dimensional data fusion, characterized in that, include: The data acquisition and preprocessing module is used to synchronously acquire stator phase current data, casing vibration data and real-time speed data of the blower motor, and to perform envelope extraction and energy feature calculation on the stator phase current data, casing vibration data and real-time speed data to obtain current envelope sequence, vibration energy sequence and window average speed value. The load state tensor construction module is used to perform channel-dimensional stitching and sliding time window slicing on the vibration energy sequence and the window average rotation speed value to obtain the load state tensor. The load disturbance reconstruction module is used to reconstruct the load components of the load state tensor based on a lightweight temporal network to obtain the estimated load disturbance components. An adaptive differential decoupling module is used to perform adaptive differential decoupling on the current envelope sequence based on the estimated load disturbance component to obtain a clean fault residual sequence. The fault feature extraction module is used to perform residual spectrum feature analysis and fault feature extraction on the pure fault residual sequence based on the window average speed value to obtain the fault feature energy value. The dynamic threshold diagnosis decision module is used to determine the dynamic alarm threshold based on the load state tensor and to make a logical comparison and judgment on the fault characteristic energy value to obtain the final diagnosis result.

Images