A fault diagnosis ar display method and system of a high-frequency shaker

Abstract

The application discloses a kind of high-frequency shaking machine fault diagnosis AR display method and system, method includes: real-time acquisition high-frequency shaking machine vibration waveform, current harmonic, temperature field distribution and phase difference time series data, and fusion generates multidimensional state characteristic matrix;Multidimensional state characteristic matrix is decomposed in time-frequency domain, and output candidate fault type and its corresponding time offset;Candidate fault type, time offset and multidimensional state characteristic matrix are jointly input into adaptive fuzzy neural network, and output early diagnosis result;Based on early diagnosis result, generate from fault source to associated component progressive diffusion path graph and key maintenance decision sequence;Using augmented reality technology, progressive diffusion path graph and maintenance decision sequence are mapped to the three-dimensional visual model of real shaking machine, and generate AR fault indication layer.Utilize the embodiment of the application, early fault accurate diagnosis can be realized and augmented reality visualization is assisted, and the intelligentization and intuitive level of equipment operation and maintenance is improved.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, and in particular to a fault diagnosis AR display method and system for a high-frequency shaking machine. Background Technology

[0002] High-frequency vibratory rollers are core equipment in the forming and processing of thin materials in modern industries such as papermaking and textiles. Their operational stability directly affects product quality and production efficiency. Due to the harsh working environment and long-term exposure to high-frequency alternating loads, critical components of vibratory rollers, such as bearings, crank connecting rods, and spring plates, are prone to fatigue damage or even sudden failures. Traditional fault diagnosis methods mainly rely on periodic shutdown inspections, single-channel vibration threshold alarms, or manual experience judgment, which suffer from prominent problems such as poor real-time performance, limited information dimensions, and difficulty in locating deep-seated fault roots. In recent years, a few industrial systems have introduced data-driven diagnostic methods, but these often analyze multi-source sensor data independently, ignoring the inherent spatiotemporal coupling characteristics between vibration, current, temperature, and phase, resulting in low accuracy in identifying early, weak, and complex faults. Furthermore, existing diagnostic results are mostly presented in text or two-dimensional charts, making it difficult to intuitively and comprehensively guide on-site maintenance personnel to quickly locate fault propagation paths and maintenance access points. Summary of the Invention

[0003] The purpose of this invention is to provide an AR display method and system for fault diagnosis of a high-frequency shaking machine, so as to overcome the shortcomings of the prior art, realize accurate early fault diagnosis and supplement it with augmented reality visualization, and improve the intelligence and intuitiveness of equipment operation and maintenance.

[0004] One embodiment of this application provides an AR display method for fault diagnosis of a high-frequency rocking machine, the method comprising: The vibration waveform, current harmonics, temperature field distribution and phase difference time series data of the high-frequency rocking machine are collected in real time by a multi-source sensor array and fused to generate a spatiotemporally synchronized multidimensional state feature matrix. The multidimensional state feature matrix is ​​decomposed in the time-frequency domain to extract fault-sensitive feature vectors. The dynamic time warping algorithm is used to perform similarity matching with the historical fault mode library to output candidate fault types and their corresponding time offsets. The candidate fault type, the time offset, and the multidimensional state feature matrix are input into a pre-trained adaptive fuzzy neural network. Through fuzzy rule reasoning and membership calculation, an early diagnostic result containing the final fault type, confidence level, and remaining effective lifetime is output. Based on the early diagnosis results, the fault propagation directed graph model is invoked to perform causal chain reasoning, generating a hierarchical diffusion path diagram from the fault source to the associated components and a key maintenance decision sequence. Augmented reality technology is used to map the hierarchical diffusion path map and maintenance decision sequence to a three-dimensional visualization model of the actual shaking machine, generating an AR fault indication layer aligned with the spatial position of the actual equipment. This AR fault indication layer is then overlaid on the real-time monitoring screen, and alarm information and maintenance instructions for the faulty parts are dynamically updated.

[0005] Another embodiment of this application provides an AR display system for fault diagnosis of a high-frequency rocking machine, the system comprising: The acquisition module is used to acquire the vibration waveform, current harmonics, temperature field distribution and phase difference time series data of the high-frequency rocking machine in real time through a multi-source sensor array, and fuse them to generate a spatiotemporally synchronized multi-dimensional state feature matrix. The decomposition module is used to perform time-frequency domain decomposition on the multidimensional state feature matrix, extract fault-sensitive feature vectors, and perform similarity matching with the historical fault mode library using a dynamic time warping algorithm to output candidate fault types and their corresponding time offsets. The diagnostic module is used to input the candidate fault type, the time offset, and the multidimensional state feature matrix into a pre-trained adaptive fuzzy neural network, and output an early diagnostic result containing the final fault type, confidence level, and remaining effective lifetime through fuzzy rule reasoning and membership degree calculation. The reasoning module is used to perform causal chain reasoning by calling the fault propagation directed graph model based on the early diagnosis results, and to generate a hierarchical diffusion path diagram from the fault source to the associated components and a key maintenance decision sequence. The display module is used to map the hierarchical diffusion path diagram and maintenance decision sequence to a three-dimensional visualization model of the actual shaking machine using augmented reality technology, generate an AR fault indication layer aligned with the spatial position of the actual equipment, and overlay the AR fault indication layer on the real-time monitoring screen, and dynamically update the alarm information and maintenance instructions of the fault location.

[0006] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0007] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0008] Compared with existing technologies, the present invention provides an AR display method for fault diagnosis of a high-frequency shaking machine, which can realize accurate early fault diagnosis and be supplemented by augmented reality visualization, thereby improving the intelligence and intuitiveness of equipment operation and maintenance. Attached Figure Description

[0009] Figure 1 Hardware structure block diagram of a computer terminal for an AR display method for fault diagnosis of a high-frequency shaking machine provided in an embodiment of the present invention; Figure 2 A schematic flowchart of an AR display method for fault diagnosis of a high-frequency rocking machine provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an AR display system for fault diagnosis of a high-frequency shaking machine provided in an embodiment of the present invention. Detailed Implementation

[0010] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] This invention first provides an AR display method for fault diagnosis of a high-frequency shaking machine. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0012] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for an AR display method for fault diagnosis of a high-frequency shaking machine, provided in an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0013] See Figure 2 The present invention provides an AR display method for fault diagnosis of a high-frequency shaking machine, which may include the following steps: S201 uses a multi-source sensor array to collect in real time the vibration waveform, current harmonics, temperature field distribution and phase difference time series data of the high-frequency shaking machine, and fuses them to generate a spatiotemporally synchronized multi-dimensional state feature matrix. Specifically, accelerometers, current transformers, thermocouple arrays, and Hall sensors can be deployed at key measuring points of the shaking machine to simultaneously collect vibration waveforms, current harmonics, temperature field distribution, and phase difference time series data, generating multi-source heterogeneous time series datasets. The core of this step is to precisely deploy multiple types of sensors on the key force-bearing and operating components of the high-frequency shaking machine. By synchronously collecting multi-dimensional physical signals, raw data reflecting the equipment's operating status is obtained, providing a complete multi-source data foundation for subsequent fault diagnosis. The specific implementation method is as follows: High-frequency vibrating machines are industrial equipment that rely on high-frequency reciprocating vibration to complete operations. The core operating components include the base, drive motor, bearing housing, vibrating spindle, and vibrating worktable. These components are high-risk areas for failure, so these locations are identified as key measuring points. Sensors are deployed around these key measuring points to ensure that signal acquisition covers all core state dimensions of the equipment.

[0014] Accelerometers are used to collect vibration waveform signals from the equipment. They are deployed at three locations: the bottom of the vibrating workbench, the outside of the bearing housing, and the end cover of the drive motor. These three locations can respectively reflect the vibration of the working load, the vibration of the transmission components, and the vibration state of the power source. The sensor range is set to ±50mm, and the sampling frequency is set to 10kHz. It can accurately capture the vibration impact and fluctuation signals during high-frequency shaking. The collected vibration waveform is output in the form of a time-series voltage signal, which completely records the change law of vibration amplitude over time.

[0015] The current transformer is deployed in the power supply circuit of the motor driving the vibrating machine. Its core function is to collect the current harmonic signal of the motor operation. The range is set from 0 to 50A (at least 50A). It can collect the time-domain waveform of the motor's operating current in real time, extract each harmonic component through waveform decomposition, and reflect the current distortion state caused by abnormal load or winding fault. The collected data includes the amplitude and phase information of the fundamental wave and each harmonic.

[0016] The thermocouple array adopts a multi-point distributed layout, deployed in four areas: the vibrating main shaft, bearing housing, motor housing, and base. Each area has 3 to 5 thermocouple measuring points, forming a temperature field acquisition array. The thermocouples are of type K, with a temperature measurement range of -20 degrees Celsius to 200 degrees Celsius. It can simultaneously collect temperature values ​​from multiple measuring points, and construct the overall temperature field distribution of the equipment through multi-point data, reflecting temperature changes caused by faults such as increased component friction and abnormal heat dissipation.

[0017] The Hall sensor is deployed at the output shaft end of the drive motor. Its core function is to collect the phase difference time sequence data between the motor rotor and the vibrating main shaft. By detecting changes in the magnetic field, it accurately records the rotation phase of the shaft. The phase detection accuracy reaches 0.1 degrees. It can capture the phase offset between different shafts in the transmission system in real time and reflect fault characteristics such as loose transmission mechanism and coaxiality deviation.

[0018] All sensors are connected to a synchronous acquisition terminal, which has a built-in hardware synchronization clock. The synchronization error is controlled within 1 millisecond, ensuring that the four types of sensors start and stop acquisition at the same time. The acquisition duration is set to be 10 consecutive seconds. After acquisition, the four types of data—vibration waveform, current harmonics, temperature field distribution, and phase difference—are integrated according to the acquisition time and measurement point location. The four types of data have different formats, units, and sampling frequencies, forming a multi-source heterogeneous time-series dataset containing multiple types, multiple measurement points, and multiple dimensions of signals. The dataset has no signal loss and no acquisition interruption, and completely retains the original state information of the equipment operation.

[0019] Timestamp alignment is performed on multi-source heterogeneous time-series datasets, and an interpolation algorithm is used to unify the sampling frequency of each sensor, eliminate data acquisition delay deviation, and generate time-synchronized multi-channel data streams. The core of this step is to solve the problems of asynchronous time and inconsistent sampling frequencies of data from multiple sensors. By calibrating timestamps and unifying frequencies, time delay deviations are eliminated, and heterogeneous data is transformed into a time-aligned standardized data stream, ensuring the accuracy of subsequent feature extraction. The specific implementation method is as follows: In multi-source heterogeneous time-series datasets, the raw data from different sensors all have their own acquisition timestamps, with the timestamp format being year-month-day-hour-minute-second-millisecond. Due to differences in sensor hardware response speed and transmission paths, there will be slight acquisition delay deviations, with the maximum deviation reaching 5 milliseconds. This deviation will cause vibration, current, temperature, and phase data to not accurately correspond to the device status at the same moment. Therefore, timestamp alignment processing is carried out first.

[0020] Timestamp alignment uses the time axis of the accelerometer as the reference clock. This sensor has the highest sampling frequency and the fastest signal response, making it suitable as a reference. The timestamps of the current transformer, thermocouple array, and Hall sensor are compared with the reference timestamp one by one to identify the time delay deviation. A time axis translation correction method is used to adjust the lagging or leading timestamps to the reference time point. After correction, the timestamps of all data are completely unified, and the same time tag corresponds to the same equipment operating status. The time alignment error is controlled within 0.1 milliseconds.

[0021] After completing the timestamp alignment, the sampling frequencies of different sensors are unified. The sampling frequency of the accelerometer is 10kHz, the current transformer is 5kHz, the thermocouple array is 1kHz, and the Hall sensor is 2kHz. The difference in sampling frequency will result in different numbers of data points within the same time length, making direct splicing and fusion impossible. Therefore, a linear interpolation algorithm is used to unify the frequency.

[0022] The core of the linear interpolation algorithm is to calculate the missing data points at the target sampling frequency based on the known sampling point values ​​through a linear relationship. In this study, all sensor data were unified to a sampling frequency of 1kHz, which can preserve the device status characteristics while reducing the amount of data processing. High-frequency data is downsampled, and uniformly distributed data points are selected and retained; low-frequency data is upsampled, and the interpolation point values ​​are calculated through the linear relationship between adjacent known points. The interpolation process is free of signal distortion and feature loss.

[0023] After timestamp alignment and sampling frequency unification, the time axes of the four types of sensor data are completely consistent, the number of data points is the same, and the time delay deviation is completely eliminated. The four types of data—vibration, current, temperature, and phase—are treated as independent channels and combined according to the measurement point order to form a multi-channel data stream containing multiple measurement points, multiple dimensions, and complete time synchronization. The data stream format is standardized and can be directly used for subsequent modal feature extraction.

[0024] Based on time-synchronized multi-channel data streams, peak values ​​and frequency components are extracted from vibration waveforms, distortion rate is calculated for current harmonics, spatial grid interpolation is performed on the temperature field, relative angles are calculated for phase differences, and feature sequences of each mode are generated. The core of this step is to extract the core state features of each physical quantity from the time-synchronized data stream, transform the original waveforms and numerical signals into quantifiable feature sequences, and construct feature data for four modes: vibration, current, temperature, and phase. The specific implementation method is as follows: Vibration modal feature extraction focuses on the vibration waveform data stream, which reflects the mechanical vibration state of the equipment. Two key features are extracted: vibration peak value and frequency components. The vibration peak value refers to the maximum amplitude of the vibration per unit time, representing the intensity of the vibration. By traversing each set of vibration waveform data, the maximum amplitude is selected. In the example, the vibration peak value at the vibrating workbench measuring point is 12.5 mm, at the bearing housing measuring point it is 8.3 mm, and at the motor end measuring point it is 5.6 mm. Frequency components are extracted using Fast Fourier Transform (FFT), converting the time-domain vibration waveform into a frequency-domain signal to identify the dominant frequency and harmonic frequencies. In the example, the dominant vibration frequency is 50 Hz, corresponding to the equipment's operating frequency, and the second harmonic is 100 Hz, reflecting component loosening. The vibration peak value and frequency components at each moment are arranged along the time axis to generate a vibration modal feature sequence.

[0025] Current mode feature extraction focuses on the current harmonic data stream, with the core calculation of the total harmonic distortion (THD) rate. This parameter is a key indicator of the degree of distortion in the current waveform; a higher THD rate indicates a more pronounced motor fault. The calculation method involves first extracting the effective values ​​of each harmonic in the current signal, then calculating the square root of the sum of the squares of the effective values ​​of each harmonic, dividing by the effective value of the fundamental current, and finally converting it into a percentage value. In the example, the THD rate is 3.2% during normal operation and rises to 8.7% under abnormal conditions. The current harmonic distortion rates at each moment are arranged along the time axis to generate a current mode feature sequence.

[0026] Temperature modal feature extraction is performed on the temperature data of the thermocouple array. The original temperature data consists of discrete measurement point values. A continuous temperature field distribution needs to be constructed through spatial grid interpolation. A cubic spline interpolation algorithm is used to divide the key area of ​​the equipment into a spatial grid with a spacing of 1 mm. Based on the temperature values ​​of the discrete measurement points, the temperature value of each grid node is calculated to form continuous temperature field data. The highest temperature, lowest temperature, and average temperature at each moment are extracted. In the example, the highest temperature in the main axis area is 65 degrees Celsius, and the average temperature is 58 degrees Celsius. These features are arranged along the time axis to generate a temperature modal feature sequence.

[0027] Phase mode feature extraction is based on the phase data of the Hall sensor. The core calculation is the relative phase difference angle between the motor rotor shaft and the vibration main shaft. The relative phase difference is obtained by performing a difference calculation on the real-time phase values ​​of the two shaft systems. The angle accuracy is 0.1 degrees. During normal operation, the phase difference is stable at about 2.5 degrees. During transmission failure, it will shift to more than 8 degrees. The relative phase difference angle at each moment is arranged on the time axis to generate a phase mode feature sequence.

[0028] Feature sequences of four modes—vibration, current, temperature, and phase—were extracted. Each sequence was of consistent length and aligned with the time axis, fully quantifying the core states of the equipment's mechanical vibration, electrical operation, thermal state, and transmission accuracy, providing standardized input for subsequent feature fusion.

[0029] The modal feature sequences are spliced ​​together along the time axis to construct a spatiotemporally synchronized high-dimensional feature matrix. Principal component analysis is then used for dimensionality reduction and fusion to finally generate a multidimensional state feature matrix.

[0030] The core of this step is to concatenate and fuse the independent feature sequences of the four modalities, eliminate feature redundancy and reduce data dimensionality through principal component analysis, and generate a concise multidimensional state feature matrix that contains all effective information, providing high-quality input for subsequent fault diagnosis. The specific implementation method is as follows: Each modal feature sequence contains multiple measurement points and multiple feature parameters. The vibration mode has 6 features (peak value and frequency component at 3 measurement points), the current mode has 1 feature (harmonic distortion rate), the temperature mode has 3 features (maximum temperature, minimum temperature, and average temperature), and the phase mode has 1 feature (phase difference), for a total of 11 feature parameters. These feature parameters are horizontally concatenated along the time axis, with each time point corresponding to one row of feature data and each column corresponding to one feature parameter, constructing a spatiotemporally synchronized high-dimensional feature matrix. The number of rows in the matrix is ​​the total number of sampling points (10,000 rows in this example), and the number of columns is 11. Each element in the matrix is ​​a quantified state feature value, fully preserving the multi-dimensional operating state information of the equipment.

[0031] High-dimensional feature matrices contain some redundant features, and there are correlations between features, such as the correlation between vibration peak and temperature. Redundant features increase the computational load of subsequent algorithms and reduce diagnostic efficiency. Therefore, principal component analysis is used for dimensionality reduction and fusion. This method is an unsupervised data dimensionality reduction algorithm that can transform high-dimensional features into low-dimensional independent principal components while retaining the maximum amount of information.

[0032] First, the high-dimensional feature matrix is ​​standardized by mapping all feature values ​​to the interval between 0 and 1 to eliminate dimensional differences. Next, the covariance matrix of the feature matrix is ​​calculated, and the eigenvalues ​​and eigenvectors of the covariance matrix are extracted. Then, principal components are selected based on the contribution rate of the eigenvalues. The cumulative contribution rate threshold is set to 95%, and the top four principal components are selected. These four principal components can retain more than 95% of the state information of the original 11 features. Finally, the original high-dimensional feature matrix is ​​projected onto the principal component space to obtain the dimensionality-reduced low-dimensional feature matrix.

[0033] The number of rows in the matrix after dimensionality reduction and fusion remains unchanged at 10,000, while the number of columns is reduced to 4. Each row in the matrix represents the comprehensive state of the device at a given time, and each column represents an independent principal component feature. This eliminates redundancy and correlation of the original features, significantly reduces the amount of data, and fully preserves fault-sensitive information. This matrix is ​​the final multidimensional state feature matrix, which can be directly input into the subsequent time-frequency domain decomposition and fault matching modules to support the fault diagnosis process of high-frequency shaking machines.

[0034] S202, perform time-frequency domain decomposition on the multidimensional state feature matrix, extract fault-sensitive feature vectors, and use dynamic time warping algorithm to perform similarity matching with historical fault mode library, output candidate fault types and their corresponding time offsets. Specifically, the multidimensional state feature matrix can be decomposed into discrete wavelet packet decomposition, the signal can be decomposed into multiple frequency band components, the energy proportion and singular entropy of each frequency band can be calculated, and a time-frequency domain feature set can be generated. The core of this step is to achieve joint time-frequency domain analysis of multidimensional state signals through discrete wavelet packet decomposition. This decomposes complex equipment state signals into components of different frequency bands, and then quantifies the characteristic information of each frequency band using energy proportion and singular entropy. This constructs a time-frequency domain feature set that accurately reflects fault characteristics. The specific implementation method is as follows: The multidimensional state feature matrix integrates spatiotemporal synchronous data of vibration, current, temperature, and phase difference. Rows represent time sampling points, and columns represent state features from different sensors, encompassing comprehensive state information of the high-frequency rocking machine. Discrete wavelet packet decomposition is a signal processing method that considers both time and frequency domain analysis. It can simultaneously perform fine decomposition of both high-frequency and low-frequency components of the signal, adapting to the non-stationary characteristics of high-frequency rocking machine fault signals. The decomposition process uses the Daubechies4 wavelet basis function, which possesses compact support and orthogonality, reducing energy leakage during decomposition and improving frequency band resolution accuracy.

[0035] The decomposition layer is set to 5 layers, which can decompose the original signal into 32 equal-width frequency band components, covering the frequency range from 0 to 5000 Hz, completely covering the characteristic frequency range of high-frequency shaking machine faults. After decomposition, each frequency band component corresponds to an independent frequency range, which can separate signal segments with different characteristics such as normal operation signals, abnormal vibration components, and current distortion components, avoiding mutual interference between different characteristic signals.

[0036] Energy percentage refers to the proportion of energy of a single frequency band component to the total signal energy, reflecting the intensity distribution of the signal in that frequency band. It is calculated by summing the squares of the signal amplitudes of the frequency band components to obtain the energy of a single frequency band, then dividing by the sum of the energies of all frequency bands to get the percentage value. A higher energy percentage indicates that the frequency band contains richer state information. Singular entropy is a quantitative indicator based on matrix singular value decomposition (SVD). It describes the complexity and irregularity of a frequency band signal. Under fault conditions, signal complexity increases, and the singular entropy value increases significantly. The calculation first constructs a trajectory matrix for the frequency band components, then performs singular value decomposition and normalization, and finally calculates the singular entropy value using the entropy formula.

[0037] The energy proportions of the 32 frequency band components are sequentially combined with the singular entropy to form a time-frequency domain feature set containing 64 parameters. This feature set contains both temporal variations in the time domain and frequency distribution information in the frequency domain, which can capture the weak features of early faults in high-frequency rocking machines. In the example, the energy proportion of the mid-to-high frequency band corresponding to a certain bearing wear fault reaches 0.32, and the singular entropy reaches 2.15, which is much higher than the corresponding value in the normal state, providing complete basic data for subsequent fault feature screening.

[0038] Based on the time-frequency domain feature set, the mutual information criterion is used to select the feature parameters with the highest correlation to the fault, eliminate redundant features, and generate a fault-sensitive feature vector. The core of this step is to quantify the correlation between time-frequency domain features and fault types using the mutual information criterion, eliminate redundant features with low correlation and overlapping information, retain the core parameters most sensitive to faults, and improve diagnostic accuracy while compressing feature dimensions. The specific implementation method is as follows: The mutual information criterion is an indicator that measures the degree of nonlinear correlation between two variables. It can accurately reflect the dependence of time-frequency domain characteristic parameters on the fault type of a high-frequency rocking machine. The larger the mutual information value, the stronger the ability of the characteristic parameter to distinguish faults, and it is the core criterion for determining fault-sensitive characteristics. The calculation uses historical fault labels as the target variable and each characteristic parameter in the time-frequency domain feature set as the input variable, calculating the mutual information value one by one. The value ranges from 0 to 1, where 0 represents no correlation and 1 represents complete correlation.

[0039] A mutual information threshold of 0.25 was set. This threshold, verified through extensive testing with numerous fault samples, effectively distinguishes between sensitive and redundant features. Feature parameters with mutual information values ​​greater than or equal to 0.25 are classified as fault-sensitive features, while those less than 0.25 are classified as redundant and directly removed. Simultaneously, the Pearson correlation coefficient between retained features is calculated, and highly overlapping features with correlation coefficients greater than 0.9 are eliminated to avoid feature duplication that reduces diagnostic efficiency and further simplifies feature dimensions.

[0040] After screening, the original 64-dimensional time-frequency domain features were simplified to 18-dimensional core features, including parameters strongly correlated with faults such as the proportion of low-frequency vibration energy, the singular entropy of mid-frequency current, the temperature gradient characteristics in the high-temperature region, and the phase difference deviation characteristics. These parameters can quickly respond to typical faults such as bearing loosening, coil aging, and transmission imbalance. The screened core features are arranged in a fixed order to form a one-dimensional fault-sensitive feature vector. The vector is concise and has high information density. In the example, the fault-sensitive feature vector contains key parameters such as the proportion of vibration peak frequency energy, total harmonic distortion rate, and singular entropy of the temperature field, which can be directly used as the query sequence for subsequent fault matching, greatly improving the efficiency of subsequent matching.

[0041] The fault-sensitive feature vector is used as a query sequence and dynamically time-normalized and matched with the template sequences in the historical fault mode library one by one. The normalized path distance and the optimal alignment path are calculated to generate a similarity score sequence. The core of this step is to use a dynamic time warping algorithm to solve the problem of temporal offset of fault features. It performs flexible matching between the current fault-sensitive feature vector and the historical fault template, quantifies the similarity through warping distance, and generates an objective similarity score sequence. The specific implementation method is as follows: The historical fault mode library is a feature database that stores various typical faults of high-frequency rocking machines. It includes 12 common faults such as bearing wear, motor instability, transmission looseness, and coil overheating. Each fault type corresponds to a standardized template sequence. The template sequence is the average feature vector after multiple samplings of the same type of fault, and has stable fault feature representativeness.

[0042] Dynamic time warping is the core algorithm for flexible matching of time series sequences. It allows the sequence to locally scale on the time axis, solving the matching deviation problem between the current fault features and historical templates caused by time offset, and adapting to real-world scenarios where the fault occurrence time is uncertain. The extracted fault-sensitive feature vector is used as the query sequence and matched sequentially with each template sequence in the historical fault pattern library. During the matching process, a two-dimensional distance matrix is ​​constructed, where the matrix elements are the Euclidean distances between corresponding feature points of the query sequence and the template sequence. Then, dynamic programming is used to find the optimal alignment path from the starting point to the ending point of the matrix; this path is the warped path with the minimum cumulative distance.

[0043] The normalized path distance is the cumulative Euclidean distance along the optimal alignment path. The smaller the distance value, the higher the similarity between the query sequence and the template sequence, and the closer the fault types are. The normalized path distance is inversely normalized and mapped to a similarity score from 0 to 100. A higher score indicates higher similarity. In the example, the normalized path distance between the query sequence and the bearing wear template is 12.36, corresponding to a similarity score of 87.6, while the normalized path distance between the query sequence and the motor instability template is 35.72, corresponding to a similarity score of 52.3.

[0044] The similarity scores of the query sequence and all historical fault templates are combined in sequence to form a similarity score sequence. Each score in the sequence corresponds to the degree of matching of a type of historical fault, which fully reflects the similarity relationship between the current fault and various typical faults, and provides a quantitative basis for candidate fault screening.

[0045] Based on the similarity score sequence, select the top-scoring candidate fault types, extract the time offset corresponding to each candidate fault from the optimal alignment path, and finally output the candidate fault types and their time offsets.

[0046] The core of this step is to filter candidate faults with high matching degree based on similarity score, extract temporal offset information from the alignment path, and finally output standardized candidate fault results to provide accurate input for subsequent fuzzy neural network diagnosis. The specific implementation method is as follows: The similarity score sequence includes the matching scores of the current fault and all historical faults. The top 3 fault types with the highest scores are selected as candidate fault types. This number can cover the high-probability fault types, while avoiding too many candidates that would increase the complexity of subsequent diagnosis. The screening rule is to remove low-match faults with scores below 60 and only retain candidate results with high confidence. In the example, the selected candidate fault types are bearing wear, transmission looseness, and support seat looseness, with corresponding similarity scores of 87.6, 78.2, and 71.5, respectively.

[0047] Time offset refers to the misalignment difference between the current fault feature sequence and the historical fault template sequence on the time axis. It reflects the relative temporal deviation of the fault occurrence. It is extracted by finding the maximum alignment offset steps between the feature points of the query sequence and the template sequence in the optimal alignment path, and then converting this into the actual time offset in milliseconds, combined with the sensor sampling frequency. This parameter reflects the lag or advance of the fault occurrence, providing a temporal basis for fault tracing. In the example, the time offset corresponding to the bearing wear candidate fault is 150 milliseconds, the time offset corresponding to transmission loosening is 80 milliseconds, and the time offset corresponding to support loosening is 210 milliseconds.

[0048] The selected candidate fault types and their corresponding time offsets are sorted from high to low according to their similarity scores to form standardized output results. The results clearly indicate the name, similarity score, and time offset of each candidate fault type. In the example, the output results are bearing wear, 150 milliseconds, transmission looseness, 80 milliseconds, and support seat looseness, 210 milliseconds. This result is directly input into the subsequent adaptive fuzzy neural network to complete the transition from feature matching to accurate diagnosis, providing reliable basic data for early fault diagnosis.

[0049] S203, the candidate fault type, the time offset, and the multidimensional state feature matrix are input into a pre-trained adaptive fuzzy neural network. Through fuzzy rule reasoning and membership calculation, an early diagnosis result containing the final fault type, confidence level, and remaining effective lifespan is output. Specifically, candidate fault types can be encoded into fuzzy label vectors, time offsets can be normalized into time delay features, and these features can be concatenated with a multidimensional state feature matrix to generate a fused diagnostic feature vector. The core of this step is to complete the standardized encoding and feature fusion of multi-source diagnostic information, unifying discrete fault types, numerical time offsets, and high-dimensional state features into a continuous and computable fusion vector, providing standardized input for the adaptive fuzzy neural network. The specific implementation method is as follows: The candidate fault types are discrete classification results obtained by matching the historical fault pattern library. They include various common fault types of high-frequency vibrating machines, such as bearing loosening, rotor imbalance, base loosening, and coil aging. To adapt to the numerical calculation requirements of neural networks, fuzzy one-hot encoding is used to convert them into fuzzy label vectors. The encoding rule is that each fault type corresponds to one dimension of the vector. The dimension value is assigned using fuzzy membership degree instead of hard 0 or 1, and the value range is from 0 to 1. The value represents the initial confidence of the candidate fault type. In the example, the matched candidate fault types are bearing loosening, rotor imbalance, and coil aging, and the corresponding fuzzy label vectors are [0.92, 0.75, 0.36]. The vector dimension is consistent with the number of candidate fault types. Fuzzy assignment can retain the uncertainty of the correlation between faults and improve the diagnostic accuracy.

[0050] The time offset is a time deviation value obtained by the dynamic time warping algorithm, reflecting the degree of time misalignment between the current fault feature and the historical fault template. The original value ranges from -10 to 10 seconds. Since the units and numerical ranges differ significantly, the min-max normalization method is used to convert it into a time delay feature in the range of 0 to 1. The normalization formula is that the normalized time delay feature equals the original time offset minus the minimum offset, and then divided by the difference between the maximum and minimum offset. In the example, the original time offset is 2.5 seconds, and the normalized time delay feature is 0.625. This processing can eliminate the interference of numerical scale on neural network training.

[0051] The multidimensional state feature matrix is ​​a 64-dimensional high-dimensional feature matrix containing vibration, current, temperature, and phase difference. The fuzzy label vector, normalized time delay feature, and multidimensional state feature matrix are concatenated along the feature dimensions in a fixed order: fuzzy label vector first, time delay feature in the middle, and multidimensional state feature matrix last. During the concatenation process, the timestamps of all features are fully synchronized without any temporal misalignment. Finally, a fused diagnostic feature vector with unified dimensions and complete information is generated. In the example, the fused vector has a dimension of 68, covering fault classification, time series offset, and full-dimensional information of equipment status. It can be directly input into the neural network for diagnostic inference.

[0052] The fused diagnostic feature vector is input into the preamble of the adaptive fuzzy neural network, and the activation intensity of each input variable to the fuzzy rule is calculated by the Gaussian membership function to generate the rule activation intensity vector. The core of this step is to transform the input features into activation signals for fuzzy rules through fuzzification, and then use a predecessor network to complete fuzzy membership calculation and rule matching, providing a strength basis for subsequent fuzzy inference. The specific implementation method is as follows: The adaptive fuzzy neural network is an adaptive diagnostic model that integrates fuzzy logic and neural networks. It consists of a pre-processor network and a post-processor network. The pre-processor network is responsible for fuzzification and rule activation. Internally, it contains 32 pre-defined fuzzy rules for diagnosing high-frequency shaking machine faults. Each rule corresponds to a relationship between a combination of fault features and the fault type. After the fused diagnostic feature vector is input into the pre-processor network, the network assigns an independent Gaussian membership function to each input feature dimension. The Gaussian membership function is a continuous function describing the degree to which an input variable belongs to a fuzzy set. The function form is determined by the center value c_i and the width value σ_i. The center value c_i is the mean of the fuzzy set, and the width value σ_i controls the steepness of the function. Both the center value and the width value are automatically optimized through the early training of the neural network and do not require manual setting.

[0053] For each input variable in the fusion diagnostic feature vector, the corresponding membership value is calculated by substituting it into the Gaussian membership function. The value ranges from 0 to 1. The larger the value, the more the variable conforms to the characteristics of the corresponding fuzzy set. After calculating the membership of all input variables, the activation strength of a single fuzzy rule is calculated using the algebraic product method. That is, the overall activation strength of a rule is obtained by multiplying the membership values ​​of all input variables in a rule. In the example, for a rule corresponding to bearing loosening fault, the membership of the input variables are 0.95, 0.88, and 0.91, respectively. The activation strength is 0.95 multiplied by 0.88 multiplied by 0.91, which is 0.76.

[0054] The activation intensity of 32 fuzzy rules is calculated sequentially. Weak activation rules with an activation intensity lower than 0.1 are removed, and the activation intensity values ​​of the valid rules are retained. They are arranged in the order of preset rules to form a one-dimensional rule activation intensity vector. In the example, the rule activation intensity vector is [0.76, 0.63, 0.58, 0.42, ...]. The length of the vector is consistent with the number of valid fuzzy rules. This vector directly reflects the triggering degree of each fault diagnosis rule and is the core input of fuzzy inference.

[0055] Based on the rule-based activation intensity vector, fuzzy inference is performed in the consequent network, and the centroid method is used to defuzzify the data, output the confidence score of each fault type, and generate the fault probability distribution. The core of this step is to perform fuzzy logic reasoning through the consequent network, transforming the fuzzy rule activation intensity into precise fault confidence, forming a standardized fault probability distribution, and clarifying the final fault type. The specific implementation method is as follows: The consequent network is the inference output layer of the adaptive fuzzy neural network. It is fully connected to the antecedent network. After receiving the rule activation intensity vector, it performs inference operations using the Mamdani fuzzy inference method. The inference logic is to weight and superimpose the rule activation intensity with the corresponding consequent output. The consequent output is a preset standardized score for each fault type. The initial value is obtained by training with historical fault data. During the inference process, the weights are adaptively adjusted to fit the nonlinear correlation characteristics of the fault features of the high-frequency shaking machine.

[0056] The output of fuzzy inference is in the form of a fuzzy set, which cannot be directly used for fault determination. It needs to be defuzzified using the centroid method. The centroid method is a classic method that uses the centroid position of the fuzzy set as the precise output value. The calculation logic is to multiply all values ​​in the fault score interval by their corresponding membership degrees, sum them up, and then divide by the sum of all membership degree values ​​to obtain the precise confidence score of a single fault type. The score range is from 0 to 1, and the higher the score, the greater the probability of the occurrence of the fault type.

[0057] After calculating the confidence scores for each candidate fault type, the scores are normalized. The normalization method is to divide the confidence score of a single fault by the sum of the confidence scores of all faults, so that the sum of the confidence scores of all fault types is 1, forming a standardized fault probability distribution. In the example, after calculation and normalization, the fault probability distribution is: bearing looseness 0.85, rotor imbalance 0.12, and coil aging 0.03. This distribution clearly shows the probability of occurrence of each fault, among which bearing looseness with the highest confidence score is the core fault type initially determined.

[0058] By combining the failure probability distribution and the current operating time, the Weibull proportional hazards model is used to predict the remaining effective lifetime, and the final output includes the final failure type, confidence level and remaining effective lifetime as an early diagnostic result.

[0059] The core of this step is to predict the remaining lifespan based on failure probability and equipment operating data, integrate failure classification, confidence level, and lifespan prediction results to form a complete early diagnostic conclusion. The specific implementation method is as follows: The Weibull proportional hazards model is a classic model applicable to the prediction of the life of rotating machinery. It is composed of a basic Weibull distribution and a proportional hazards function. It can accurately predict the remaining effective life by combining covariates such as failure probability and operating time. The shape parameter m and scale parameter η of the basic Weibull distribution in the model are obtained by training with historical full life cycle failure data of high-frequency shaking machines. The shape parameter m is 2.3, which reflects the increasing trend of failure occurrence rate. The scale parameter η is 1800 hours, which represents the average failure life of the equipment.

[0060] The proportional hazards function uses the highest confidence level in the failure probability distribution as the core covariate. The higher the confidence level, the larger the risk proportional coefficient and the shorter the remaining effective life. Substituting the current cumulative running time of the equipment and the highest confidence level of the failure into the model, the instantaneous failure rate of the equipment is calculated. Then, the remaining effective life is derived through integration. The remaining effective life refers to the length of time the equipment can operate safely under the current failure state, in hours. In the example, the current running time of the equipment is 1260 hours, the confidence level of the bearing loosening failure is 0.85, and the remaining effective life is calculated by the model to be 115 hours.

[0061] By integrating fault diagnosis and life prediction results, the fault type with the highest confidence in the fault probability distribution is defined as the final fault type, and the corresponding confidence level is defined as the core confidence level. Combined with the predicted remaining effective life, an early diagnosis result is generated. The result includes three core pieces of information: final fault type, core confidence level, and remaining effective life. The diagnosis time and feature source are also labeled. In the example, the early diagnosis result is that the final fault type is bearing loosening, with a confidence level of 0.85 and a remaining effective life of 115 hours. This result can directly support subsequent fault propagation reasoning and maintenance decision-making, realizing early and accurate diagnosis of high-frequency vibrating machine faults.

[0062] S204. Based on the early diagnosis results, the fault propagation directed graph model is invoked to perform causal chain reasoning, generating a hierarchical diffusion path diagram from the fault source to the associated components and a key maintenance decision sequence. Specifically, the final fault type and confidence level in the early diagnosis results can be analyzed, and the corresponding fault propagation directed graph model can be loaded from the fault knowledge base. The nodes of this model represent components, and the edges represent causal relationships, resulting in a weighted directed graph instance. The core of this step is to extract key information from early diagnostic results, match and call the propagation model in the fault knowledge base, and transform the abstract fault into a reasonable directed graph structure, providing a standardized model foundation for subsequent fault propagation analysis. The specific implementation method is as follows: The early diagnostic results are structured data containing the final fault type, fault confidence level, and remaining effective life. The final fault type is a classification identifier characterizing the abnormal state of the high-frequency shaking machine; common types include loose rolling bearings, bent shafts, frame resonance, and eccentric wear of the motor. The confidence level is a value between 0 and 1, representing the reliability of the fault type determination; the closer the value is to 1, the higher the reliability of the determination. The parsing process is completed by a structured data parsing module. The module automatically extracts the text identifier of the final fault type and the confidence level value, removes irrelevant data such as remaining effective life, and retains the core parameters used for fault propagation inference. In the example, the parsed final fault type is loose rolling bearings, with a confidence level of 0.92, which indicates that this fault determination result has high reliability.

[0063] The fault knowledge base is a standardized database that stores the propagation rules of all types of faults in high-frequency rocking machines. It pre-stores a directed graph model for each type of fault. The model is represented by a directed graph topology. The nodes correspond to the physical components of the high-frequency rocking machine, including core components such as rolling bearings, shafts, frames, drive motors, fastening bolts, and transmission supports. Each node is labeled with basic information such as component name, three-dimensional spatial coordinates, material properties, and operating parameters. Edges represent the causal relationship between components. The weight of the edge represents the probability and degree of impact of the fault propagating from the upstream component to the downstream component. The weight ranges from 0 to 1. The higher the value, the greater the probability of fault propagation and the more significant the impact.

[0064] Based on the final fault type obtained from the analysis, the fault knowledge base automatically matches the corresponding directed graph model for fault propagation. Taking a loose rolling bearing fault as an example, the matched model nodes include the rolling bearing, shaft, fastening bolt, transmission support, frame, and drive motor. The edge connections are: rolling bearing points to shaft, shaft points to frame, rolling bearing points to fastening bolt, fastening bolt points to transmission support, and frame points to drive motor. The edge weights are set to 0.95, 0.82, 0.88, 0.75, and 0.63, respectively, based on historical fault propagation statistics. After loading, the node information, edge connections, and weight parameters are integrated to generate a weighted directed graph instance. This instance fully preserves the topological relationships and influence strength of fault propagation, providing an accurate model carrier for subsequent path reasoning.

[0065] Starting with the component corresponding to the fault type, perform a breadth-first search in the weighted directed graph instance, calculate the causal influence strength and propagation depth of each reachable node, and generate a set of fault propagation paths. The core of this step is to traverse the fault propagation path using breadth-first search, quantify the impact of the fault on each component and the propagation level, form a complete set of fault propagation paths, and clarify all paths from the source of the fault to the entire machine. The specific implementation method is as follows: The starting node is the fault source component of the high-frequency rocking machine corresponding to the fault type, directly determined by early diagnostic results. In the example, the starting node corresponding to the loose rolling bearing fault is the rolling bearing component. This node is the source of fault propagation and also the starting point of breadth-first search. Breadth-first search is a classic algorithm for hierarchical traversal of directed graphs. It traverses all reachable nodes in hierarchical order from near to far, completely covering the entire path of fault propagation from the source outwards, avoiding omission of potential propagation branches. The algorithm's traversal order strictly follows the hierarchical progression rule, first traversing the first-level nodes directly connected to the starting node, then traversing the second-level nodes connected to the first-level nodes, and so on downwards until there are no new reachable nodes.

[0066] Causal impact strength is a core indicator for quantifying the impact of a fault on downstream components. It is calculated as the product of the weights of all edges along the path from the starting node to the target node. This value directly reflects the severity of the impact after the fault propagates to the component; a higher value indicates a greater impact. Propagation depth is an indicator of the fault propagation hierarchy. The starting node is at depth 0, and directly connected downstream nodes are at depth 1. The depth value increases by 1 for each level of propagation downwards. A larger depth value indicates a longer propagation distance and a wider impact range. In the example, the starting node rolling bearing has a depth of 0 and a causal influence strength of 1.0; the directly connected shaft and fastening bolt have a depth of 1, with the shaft having a causal influence strength of 0.95 and the fastening bolt having a causal influence strength of 0.88; the frame connected to the shaft has a depth of 2 and a causal influence strength of 0.95 × 0.82 = 0.779; the transmission support connected to the fastening bolt has a depth of 2 and a causal influence strength of 0.88 × 0.75 = 0.66; the drive motor connected to the frame has a depth of 3 and a causal influence strength of 0.779 × 0.63 = 0.491.

[0067] During the traversal, each complete fault propagation path is recorded synchronously. Each path starts from the initial node, and the connections between downstream nodes are recorded in order of propagation depth. The generated fault propagation paths in the example include two core paths: rolling bearing-shaft-frame-drive motor and rolling bearing-fastening bolt-transmission support. All traversed paths, the causal influence strength of each node, and the propagation depth are integrated to form a fault propagation path set. This set fully presents the entire path and quantified impact of the fault propagating from its source to related components of the entire machine, providing data support for subsequent maintenance decisions.

[0068] Based on the failure probability and maintenance cost of each node in the fault propagation path set, a dynamic programming algorithm is used to generate the optimal maintenance decision sequence, including the detection order and replacement priority, and to generate the critical maintenance decision sequence. The core of this step is to combine failure probability and maintenance cost, optimize the maintenance strategy through dynamic programming algorithm, generate the optimal maintenance decision sequence that balances efficiency and cost, and clarify the order of maintenance operations and the priority of component replacement. The specific implementation method is as follows: The node failure probability is a core parameter for determining the risk of component failure. It is obtained by multiplying the confidence level of the early diagnosis results by the causal influence strength of the node, and takes a value from 0 to 1. The higher the value, the greater the risk of the component failing. In the example, the failure probability of the rolling bearing node is 0.92×1.0=0.92, the shaft node is 0.92×0.95=0.874, the fastening bolt node is 0.92×0.88=0.81, the frame node is 0.92×0.779=0.717, the transmission support node is 0.92×0.66=0.607, and the drive motor node is 0.92×0.491=0.452. Maintenance cost is a comprehensive indicator that quantifies the cost and time of maintenance operations. It includes three dimensions: maintenance time, parts cost, and downtime loss. The values ​​of the three dimensions are standardized and then weighted and summed to obtain a single maintenance cost value. The lower the value, the more economical and efficient the maintenance operation is. In the example, the maintenance cost of rolling bearing is 1.2, shaft is 1.8, fastening bolt is 1.5, frame is 3.5, transmission support is 2.8, and drive motor is 4.2.

[0069] Dynamic programming is an optimization algorithm for finding optimal decisions. It aims to minimize total maintenance cost and maximize troubleshooting efficiency. Each node in the fault propagation path is treated as a decision stage, with fault probability and maintenance cost as state variables. The algorithm iteratively calculates the optimal decision value for each node, ultimately obtaining the globally optimal maintenance decision sequence. The core logic of the algorithm is to prioritize components with high fault probability and low maintenance cost, performing detection operations first, and then determining replacement priority based on the fault condition, thus avoiding resource waste and downtime caused by blind repairs.

[0070] After iterative calculation using a dynamic programming algorithm, the generated optimal maintenance decision sequence includes two core components: inspection order and replacement priority. The inspection order is arranged from high to low failure probability and from low to high maintenance cost. In the example, the inspection order is rolling bearing, fastening bolt, shaft, transmission support, frame, and drive motor. The replacement priority is set according to the degree of failure risk: components with a failure probability higher than 0.8 are classified as first-level replacement priority, those between 0.6 and 0.8 as second-level, and those below 0.6 as third-level. In the example, rolling bearing, shaft, and fastening bolt are first-level replacement priority, transmission support and frame are second-level, and drive motor is third-level. Integrating the inspection order, replacement priority, and maintenance operation suggestions generates a critical maintenance decision sequence. This sequence provides standardized operational guidance for on-site maintenance, improving maintenance efficiency and accuracy.

[0071] The node coordinates and connections in the fault propagation path set are visualized to generate a hierarchical propagation path diagram from the fault source to the associated components, and finally outputs the hierarchical propagation path diagram and the key maintenance decision sequence.

[0072] The core of this step is to transform the abstract fault propagation path and decision information into a visual graph, generating an intuitive hierarchical propagation path diagram, and integrating it with the maintenance decision sequence for output, providing intuitive visual support for fault diagnosis and maintenance. The specific implementation method is as follows: Node coordinates are the positional parameters of each component of the high-frequency shaking machine in three-dimensional space. They are derived from the node attributes of a weighted directed graph instance and are represented using a three-dimensional Cartesian coordinate system, including x, y, and z axis values, precisely corresponding to the actual installation position of the component on the shaking machine. Connection relationships represent the path direction of fault propagation, indicated by the edges of the directed graph. Arrows represent the direction of fault propagation, pointing from the upstream faulty component to the downstream affected component. Visualization rendering uses a 3D graphics rendering engine to draw identification icons for each component according to the node coordinates. The icon size is proportional to the actual size of the component. Node colors are rendered according to the intensity of causal influence: red for intensity above 0.8, orange for 0.6 to 0.8, yellow for 0.4 to 0.6, and blue for below 0.4. The colors visually distinguish the severity of the component's impact from the fault.

[0073] The step-by-step propagation path diagram starts with the fault source component and draws the fault propagation path step by step according to the propagation depth. The thickness of the arrow lines is positively correlated with the weight of the edge; the greater the weight, the thicker the line, clearly showing the differences in the intensity of fault propagation. The diagram also simultaneously labels key information such as the name of each component, fault probability, propagation depth, and maintenance priority, intuitively showing the complete path of the fault's step-by-step propagation from the source to related components. In the example, the path diagram starts with the red rolling bearing icon, drawing arrows pointing to the orange shaft and orange fastening bolts, then extending to the yellow frame and yellow transmission support, and finally to the blue drive motor, completely showing the step-by-step propagation path and the degree of impact of the fault.

[0074] After visualization, the hierarchical propagation path diagram and key maintenance decision sequence are standardized and integrated. The output format includes both visual images and text data. The visual images intuitively display the fault propagation path, while the text data clearly lists information such as the detection sequence, replacement priority, maintenance operation suggestions, and maintenance costs. The final output of the hierarchical propagation path diagram and key maintenance decision sequence achieves a visualized presentation of fault propagation and standardized output of maintenance decisions. It can be directly used to guide on-site maintenance personnel, significantly improving the efficiency of fault diagnosis and maintenance of high-frequency vibration machines and reducing the risk of misoperation.

[0075] S205, using augmented reality technology, the hierarchical diffusion path diagram and maintenance decision sequence are mapped to a three-dimensional visualization model of the actual shaking machine, generating an AR fault indication layer aligned with the spatial position of the actual equipment, and overlaying the AR fault indication layer onto the real-time monitoring screen, and dynamically updating the alarm information and maintenance guidance of the fault location.

[0076] Specifically, a three-dimensional digital model of each component of the shaking machine can be established, and the model coordinate system can be spatially registered with the world coordinate system of the real equipment through a calibration algorithm to generate a spatially aligned three-dimensional visualization model. The core of this step is to complete the three-dimensional digital reconstruction of the high-frequency shaking machine and accurately match the virtual and real spaces, providing a unified spatial benchmark for subsequent AR layer overlay and ensuring that the virtual fault information completely coincides with the real equipment location. The specific implementation method is as follows: The construction of the 3D digital model is based on the actual physical structure of the high-frequency shaking machine. The equipment is disassembled into independent components such as the base, drive motor, shaking table, rolling bearings, transmission linkages, sensor mounting bases, wiring terminals, and cooling air ducts. These components are meticulously modeled at a 1:1 scale, with modeling accuracy controlled within 0.1 mm, fully restoring the external dimensions, installation position, connection relationships, and spatial contours of each component. During the modeling process, an independent index identifier for each component is retained, recording parameters such as the geometric center coordinates, boundary range, and spatial orientation of each component, forming a modular 3D model library that can be independently accessed, facilitating accurate mapping of subsequent fault information.

[0077] The calibration algorithm used is the hand-eye calibration algorithm, which is the core technology for achieving virtual-real space registration. It establishes a coordinate transformation relationship between the AR display device, the 3D digital model, and the actual shaking machine, eliminating positional deviations between the model space and the real world. During the calibration process, coplanar calibration targets are placed at locations such as the base, motor end cover, and worktable corners of the actual shaking machine. The center coordinates of these targets serve as spatial reference points. There are no fewer than nine such targets, evenly distributed at different heights and orientations of the equipment, ensuring that the calibration covers the entire equipment space.

[0078] The system acquires images of the calibration target on a real device using an AR display device, extracts the pixel coordinates of the target's center, and simultaneously reads the model coordinates of the corresponding reference point in the 3D digital model. These coordinates are then used in a hand-eye calibration algorithm to calculate a rotation matrix and a translation vector. The rotation matrix corrects the model's spatial orientation, and the translation vector adjusts its spatial position, transforming the local coordinate system of the 3D digital model to the world coordinate system of the real device. After registration, the spatial alignment accuracy is verified through multi-point calibration, with the spatial deviation of the calibration points controlled within 0.5 mm. This ensures that the position of any component on the model perfectly corresponds to the real device, ultimately generating a spatially aligned 3D visualization model. This model can adjust its display angle in real time according to the viewing perspective, maintaining visual consistency with the real device.

[0079] Map the nodes and edges in the hierarchical diffusion path diagram to the spatial locations of the corresponding components in the 3D visualization model to generate an AR fault indication layer with color coding and arrow indicators. The core of this step is to transform the abstract fault propagation logic into a visual AR icon. Through spatial mapping and visual encoding, the fault location, propagation path, and scope of impact are intuitively presented on the 3D model. The specific implementation method is as follows: In the hierarchical diffusion path diagram, nodes correspond to actual components of the high-frequency shaking machine. Each node carries information such as fault type, confidence level, and propagation depth. Edges represent the causal propagation relationship of the fault, and the weight of the edge represents the propagation risk level. The mapping process uses a spatially aligned 3D visualization model as a carrier. Based on the component index corresponding to the node, the node information is accurately projected onto the geometric center position of the corresponding component in the model. The mapping accuracy is consistent with the 3D modeling accuracy, ensuring that the node identifier fits the component surface without offset or misalignment.

[0080] Color coding is the core visual rule for distinguishing fault states. Fault source nodes are marked in bright red, representing the initial location of the fault. First-level diffusion-related components are marked in orange, representing components directly affected by the fault. Second-level and above diffusion-related components are marked in yellow, representing indirectly related components. Normal components have no color marking to maintain the original texture of the model. The transparency of different colors is uniformly set to 70%, which clearly distinguishes fault states without obscuring the structure of the components themselves.

[0081] The fault propagation path is indicated by directional arrows. The starting point of the arrow is the fault source node, and the ending point is the affected related components. The thickness of the arrow corresponds to the risk weight of the fault propagation; the higher the weight, the thicker the arrow, which intuitively reflects the intensity and direction of the fault spread. At the same time, key data such as fault confidence and remaining effective life are displayed next to the node identifier. The font size of the data automatically adapts to the viewing distance, ensuring clear viewing at close range and prominent identification at long distance.

[0082] All color markings, arrow paths, and text labels are integrated to form an AR fault indication layer. The layer is rendered in vector format and can be freely scaled and rotated without distortion. The layer is bound to the 3D visualization model to achieve synchronous movement and transformation, providing standardized visual materials for subsequent AR image overlay.

[0083] Real-time monitoring images are captured by augmented reality display devices, and the device attitude is tracked in real time using visual inertial odometry. The AR fault indication layer is rendered on the screen in a semi-transparent overlay manner to ensure that the layer changes synchronously with the viewing angle and generate an AR overlay image. The core of this step is to achieve real-time fusion of the virtual AR layer and the real device image. Pose tracking ensures seamless synchronization between the virtual and real images, allowing maintenance personnel to intuitively see fault indication information in their real-world view. The specific implementation method is as follows: The augmented reality display device uses lightweight head-mounted glasses or industrial tablets. The device has a built-in high-definition image acquisition sensor to capture real-time monitoring images of the scene where the high-frequency shaking machine is located. The frame rate of the image acquisition is no less than 30 frames per second, and the resolution is no less than 1080P, ensuring that the real image is clear, smooth, and without lag or blur.

[0084] Visual inertial odometry (VIO) is the core technology for real-time device attitude tracking. It integrates visual image features with data from the inertial measurement unit (IMU) to calculate the position, translation, rotation angle, and other attitude parameters of the AR display device in real time. Tracking accuracy is controlled within 0.1 mm, and attitude update latency is less than 50 milliseconds, effectively avoiding image drift and misalignment issues. During tracking, texture features and angular contours of the real device surface are extracted as visual reference points. Combined with acceleration and angular velocity data from the IMU, the device attitude is continuously corrected, ensuring that the AR fault indication layer accurately matches the real device regardless of the viewing angle from the front, side, or top.

[0085] The rendering process employs a semi-transparent overlay method, with the transparency parameter of the AR fault indicator layer set to 0.3. This ensures that the fault indicators are clearly visible while allowing the structural details of the actual equipment to be seen through the layer, without obstructing the view during maintenance operations. The layer rendering follows the visual rule of objects appearing larger when closer and smaller when farther away, automatically adjusting the size of the indicators with the viewing distance and the display orientation with the viewing angle, perfectly matching the spatial perspective of the actual equipment.

[0086] The virtual layer rendered in real time is merged frame by frame with the real monitoring images collected to generate a coherent and accurate AR overlay image. The fault location and propagation path in the image are completely aligned with the real equipment, providing maintenance personnel with an intuitive virtual-real fusion diagnostic view.

[0087] The alarm information of the fault location and the maintenance guidance text in the AR overlay screen are dynamically updated according to the maintenance decision sequence. The maintenance operation is guided by voice and text prompts, and finally a dynamically updated AR fault diagnosis and maintenance guidance mechanism is formed.

[0088] The core of this step is to transform maintenance decisions into real-time interactive AR guidance, dynamically updating prompts based on the fault status, and standardizing maintenance operations through multimodal guidance, forming a closed-loop mechanism of diagnosis-guidance-maintenance. The specific implementation method is as follows: The maintenance decision sequence includes the testing order of faulty components, replacement priority, maintenance steps, tool requirements, and safety precautions, serving as the core basis for dynamic updates. The system analyzes the execution status of the maintenance decision sequence in real time, activating alarm information for corresponding components sequentially according to the testing order. Untested components display a pending status, components under testing display a highlighted alarm status, and components that have completed maintenance display a normal status, achieving visualized control of the maintenance process.

[0089] The alarm information for the fault location includes the final fault type, confidence score, and remaining effective lifespan. The text color matches the AR layer color, and the alarm information for the faulty component is displayed in bold red font, scrolling to remind maintenance personnel to prioritize its handling. Maintenance guidance text is dynamically generated based on the decision sequence, providing standardized operating procedures for different fault types. For example, the guidance text for a loose bearing is: Turn off the equipment power → Remove the protective cover → Tighten the bearing fixing bolts → Detect the vibration value → Reset the protective cover. The text is concise and clear, and the steps are clear and executable.

[0090] Voice prompts and text guidance are triggered simultaneously. The voice broadcast uses clear human voice prompts at a moderate speed, and key points are repeated. For example, priority is given to handling loose drive motor bearings. The combination of voice and text is adapted to different maintenance operation scenarios, improving the convenience of guidance.

[0091] As the maintenance operation progresses, the system receives real-time feedback on maintenance completion, automatically updates the status indicators in the AR overlay, removes alarm information for repaired components, and proceeds to the next maintenance step. This continues until all faults are resolved, at which point the AR layer automatically clears all fault indicators, displaying the equipment's normal operating status. This dynamic update mechanism requires no manual intervention throughout the entire process, automatically matching maintenance progress and continuously providing accurate diagnosis and guidance, ultimately forming a stable and efficient AR fault diagnosis and maintenance guidance mechanism.

[0092] Another embodiment of the present invention provides an AR display system for fault diagnosis of a high-frequency shaking machine, see [link to documentation]. Figure 3 The system may include: The acquisition module 301 is used to acquire the vibration waveform, current harmonics, temperature field distribution and phase difference time series data of the high-frequency rocking machine in real time through a multi-source sensor array, and fuse them to generate a spatiotemporally synchronized multi-dimensional state feature matrix. The decomposition module 302 is used to perform time-frequency domain decomposition on the multidimensional state feature matrix, extract fault-sensitive feature vectors, and perform similarity matching with the historical fault mode library using a dynamic time warping algorithm to output candidate fault types and their corresponding time offsets. The diagnostic module 303 is used to input the candidate fault type, the time offset and the multidimensional state feature matrix into a pre-trained adaptive fuzzy neural network, and output an early diagnostic result containing the final fault type, confidence level and remaining effective life through fuzzy rule reasoning and membership degree calculation. The reasoning module 304 is used to call the fault propagation directed graph model to perform causal chain reasoning based on the early diagnosis results, and generate a hierarchical diffusion path diagram from the fault source to the associated components and a key maintenance decision sequence. Display module 305 is used to map the hierarchical diffusion path map and maintenance decision sequence to a three-dimensional visualization model of the real shaking machine using augmented reality technology, generate an AR fault indication layer aligned with the spatial position of the real equipment, overlay the AR fault indication layer on the real monitoring screen, and dynamically update the alarm information and maintenance guidance of the fault location.

[0093] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0094] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0095] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0096] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A fault diagnosis AR display method for a high-frequency shaking machine, characterized in that, The method includes: The vibration waveform, current harmonics, temperature field distribution and phase difference time series data of the high-frequency rocking machine are collected in real time by a multi-source sensor array and fused to generate a spatiotemporally synchronized multidimensional state feature matrix. The multidimensional state feature matrix is ​​decomposed in the time-frequency domain to extract fault-sensitive feature vectors. The dynamic time warping algorithm is used to perform similarity matching with the historical fault mode library to output candidate fault types and their corresponding time offsets. The candidate fault type, the time offset, and the multidimensional state feature matrix are input into a pre-trained adaptive fuzzy neural network. Through fuzzy rule reasoning and membership calculation, an early diagnostic result containing the final fault type, confidence level, and remaining effective lifetime is output. Based on the early diagnosis results, the fault propagation directed graph model is invoked to perform causal chain reasoning, generating a hierarchical diffusion path diagram from the fault source to the associated components and a key maintenance decision sequence. Augmented reality technology is used to map the hierarchical diffusion path map and maintenance decision sequence to a three-dimensional visualization model of the actual shaking machine, generating an AR fault indication layer aligned with the spatial position of the actual equipment. This AR fault indication layer is then overlaid on the real-time monitoring screen, and alarm information and maintenance instructions for the faulty parts are dynamically updated.

2. The method according to claim 1, characterized in that, The process involves real-time acquisition of vibration waveforms, current harmonics, temperature field distribution, and phase difference time-series data of a high-frequency rocking machine using a multi-source sensor array, and fusing these data to generate a spatiotemporally synchronized multidimensional state feature matrix, including: Accelerometers, current transformers, thermocouple arrays and Hall sensors are deployed at key measuring points of the shaking machine to synchronously collect vibration waveforms, current harmonics, temperature field distribution and phase difference time series data, and generate multi-source heterogeneous time series datasets. Timestamp alignment is performed on multi-source heterogeneous time-series datasets, and an interpolation algorithm is used to unify the sampling frequency of each sensor, eliminate data acquisition delay deviation, and generate time-synchronized multi-channel data streams. Based on time-synchronized multi-channel data streams, peak values ​​and frequency components are extracted from vibration waveforms, distortion rate is calculated for current harmonics, spatial grid interpolation is performed on the temperature field, relative angles are calculated for phase differences, and feature sequences of each mode are generated. The modal feature sequences are spliced ​​together along the time axis to construct a spatiotemporally synchronized high-dimensional feature matrix. Principal component analysis is then used for dimensionality reduction and fusion to finally generate a multidimensional state feature matrix.

3. The method according to claim 2, characterized in that, The process involves performing time-frequency domain decomposition on the multidimensional state feature matrix to extract fault-sensitive feature vectors, and then using a dynamic time warping algorithm to perform similarity matching with a historical fault pattern library to output candidate fault types and their corresponding time offsets, including: Discrete wavelet packet decomposition is performed on the multidimensional state feature matrix to decompose the signal into multiple frequency band components. The energy proportion and singular entropy of each frequency band are calculated to generate a time-frequency domain feature set. Based on the time-frequency domain feature set, the mutual information criterion is used to select the feature parameters with the highest correlation to the fault, eliminate redundant features, and generate a fault-sensitive feature vector. The fault-sensitive feature vector is used as a query sequence and dynamically time-normalized and matched with the template sequences in the historical fault mode library one by one. The normalized path distance and the optimal alignment path are calculated to generate a similarity score sequence. Based on the similarity score sequence, select the top-scoring candidate fault types, extract the time offset corresponding to each candidate fault from the optimal alignment path, and finally output the candidate fault types and their time offsets.

4. The method according to claim 3, characterized in that, The process involves inputting the candidate fault type, the time offset, and the multidimensional state feature matrix into a pre-trained adaptive fuzzy neural network. Through fuzzy rule reasoning and membership calculation, the network outputs an early diagnostic result containing the final fault type, confidence level, and remaining effective lifetime, including: Candidate fault types are encoded into fuzzy label vectors, time offsets are normalized into time delay features, and concatenated with multidimensional state feature matrices to generate fused diagnostic feature vectors. The fused diagnostic feature vector is input into the preamble of the adaptive fuzzy neural network, and the activation intensity of each input variable to the fuzzy rule is calculated by the Gaussian membership function to generate the rule activation intensity vector. Based on the rule-based activation intensity vector, fuzzy inference is performed in the consequent network, and the centroid method is used to defuzzify the data, output the confidence score of each fault type, and generate the fault probability distribution. By combining the failure probability distribution and the current operating time, the Weibull proportional hazards model is used to predict the remaining effective lifetime, and the final output includes the final failure type, confidence level and remaining effective lifetime as an early diagnostic result.

5. The method according to claim 4, characterized in that, Based on the early diagnostic results, a directed graph model for fault propagation is invoked to perform causal chain reasoning, generating a hierarchical diffusion path diagram from the fault source to related components and a sequence of key maintenance decisions, including: The final fault type and confidence level in the early diagnosis results are analyzed, and the corresponding fault propagation directed graph model is loaded from the fault knowledge base. The nodes of the model represent components and the edges represent causal relationships, resulting in a weighted directed graph instance. Starting with the component corresponding to the fault type, perform a breadth-first search in the weighted directed graph instance, calculate the causal influence strength and propagation depth of each reachable node, and generate a set of fault propagation paths. Based on the failure probability and maintenance cost of each node in the fault propagation path set, a dynamic programming algorithm is used to generate the optimal maintenance decision sequence, including the detection order and replacement priority, and to generate the critical maintenance decision sequence. The node coordinates and connections in the fault propagation path set are visualized to generate a hierarchical propagation path diagram from the fault source to the associated components, and finally outputs the hierarchical propagation path diagram and the key maintenance decision sequence.

6. The method according to claim 5, characterized in that, The process involves using augmented reality (AR) technology to map the hierarchical diffusion path diagram and maintenance decision sequence onto a 3D visualization model of the actual shaking machine, generating an AR fault indication layer aligned with the spatial position of the actual equipment. This AR fault indication layer is then overlaid on the real-time monitoring screen, and alarm information and maintenance instructions for the faulty location are dynamically updated. This includes: Establish a three-dimensional digital model of each component of the shaking machine, and use a calibration algorithm to spatially register the model coordinate system with the world coordinate system of the real equipment to generate a spatially aligned three-dimensional visualization model; Map the nodes and edges in the hierarchical diffusion path diagram to the spatial locations of the corresponding components in the 3D visualization model to generate an AR fault indication layer with color coding and arrow indicators. Real-time monitoring images are captured by augmented reality display devices, and the device attitude is tracked in real time using visual inertial odometry. The AR fault indication layer is rendered on the screen in a semi-transparent overlay manner to ensure that the layer changes synchronously with the viewing angle and generate an AR overlay image. The alarm information of the fault location and the maintenance guidance text in the AR overlay screen are dynamically updated according to the maintenance decision sequence. The maintenance operation is guided by voice and text prompts, and finally a dynamically updated AR fault diagnosis and maintenance guidance mechanism is formed.

7. A fault diagnosis AR display system for a high-frequency shaking machine, characterized in that, The system includes: The acquisition module is used to acquire the vibration waveform, current harmonics, temperature field distribution and phase difference time series data of the high-frequency rocking machine in real time through a multi-source sensor array, and fuse them to generate a spatiotemporally synchronized multi-dimensional state feature matrix. The decomposition module is used to perform time-frequency domain decomposition on the multidimensional state feature matrix, extract fault-sensitive feature vectors, and perform similarity matching with the historical fault mode library using a dynamic time warping algorithm to output candidate fault types and their corresponding time offsets. The diagnostic module is used to input the candidate fault type, the time offset, and the multidimensional state feature matrix into a pre-trained adaptive fuzzy neural network, and output an early diagnostic result containing the final fault type, confidence level, and remaining effective lifetime through fuzzy rule reasoning and membership degree calculation. The reasoning module is used to perform causal chain reasoning by calling the fault propagation directed graph model based on the early diagnosis results, and to generate a hierarchical diffusion path diagram from the fault source to the associated components and a key maintenance decision sequence. The display module is used to map the hierarchical diffusion path diagram and maintenance decision sequence to a three-dimensional visualization model of the actual shaking machine using augmented reality technology, generate an AR fault indication layer aligned with the spatial position of the actual equipment, and overlay the AR fault indication layer on the real-time monitoring screen, and dynamically update the alarm information and maintenance instructions of the fault location.

8. The system according to claim 7, characterized in that, The acquisition module is specifically used for: Accelerometers, current transformers, thermocouple arrays and Hall sensors are deployed at key measuring points of the shaking machine to synchronously collect vibration waveforms, current harmonics, temperature field distribution and phase difference time series data, and generate multi-source heterogeneous time series datasets. Timestamp alignment is performed on multi-source heterogeneous time-series datasets, and an interpolation algorithm is used to unify the sampling frequency of each sensor, eliminate data acquisition delay deviation, and generate time-synchronized multi-channel data streams. Based on time-synchronized multi-channel data streams, peak values ​​and frequency components are extracted from vibration waveforms, distortion rate is calculated for current harmonics, spatial grid interpolation is performed on the temperature field, relative angles are calculated for phase differences, and feature sequences of each mode are generated. The modal feature sequences are spliced ​​together along the time axis to construct a spatiotemporally synchronized high-dimensional feature matrix. Principal component analysis is then used for dimensionality reduction and fusion to finally generate a multidimensional state feature matrix.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-6 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-6.

Images