Speed reducer fault positioning method and system based on artificial intelligence

Abstract

The invention provides a speed reducer fault positioning method and system based on artificial intelligence, and relates to the technical field of intelligent fault detection and positioning of industrial equipment. The method comprises the following steps: synchronously acquiring vibration, temperature and acoustic emission signals by using a multi-mode sensing module, carrying out filtering, correction and improved wavelet packet decomposition processing through an edge calculation unit, extracting time domain, frequency domain and nonlinear characteristics, calculating a vibration degradation degree and a state deviation index to construct a characteristic space, and calculating a state deviation index; fault large-class identification and fault source accurate positioning are realized through deep metric learning and a graph attention network, and credibility evaluation and online incremental learning optimization are carried out in combination with a D-S evidence theory. The system comprises a multi-modal sensing module, an edge computing unit, a cloud analysis platform, a self-adaptive optimization module and a man-machine interaction terminal. The method overcomes the defects that in a traditional method, a signal source is single, feature extraction depends on experience, real-time performance is insufficient, positioning is inaccurate and the like, and fault detection accuracy and equipment operation safety are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fault detection and location of industrial equipment, and particularly to a method and system for locating faults of a speed reducer based on artificial intelligence. Background Art

[0002] The speed reducer is a core component in the fields of industrial automation, mechanical manufacturing, wind power generation, etc. Its main function is to reduce the motor speed and increase the output torque to meet the operating requirements of various mechanical systems. However, due to the complex long-term operating environment, key components such as gears and bearings inside the speed reducer are vulnerable to factors such as fatigue damage, poor lubrication, and overload impact, which can cause faults, resulting in abnormal equipment operation and even unplanned shutdown of the production system.

[0003] Traditional speed reducer fault diagnosis methods mainly rely on signal processing technologies such as spectrum analysis, envelope demodulation, and empirical mode decomposition, combined with artificially set thresholds or statistical indicators for anomaly detection. However, such methods have certain limitations. For example, they rely on a single signal source and it is difficult to effectively extract comprehensive features of different types of faults; the feature extraction process depends on expert experience and it is difficult to adapt to the changing fault patterns under complex working conditions; most studies only focus on static fault features and lack modeling of the fault evolution trend, resulting in weak early fault warning capabilities. In addition, existing methods mostly use statistical analysis or simple classification models and it is difficult to accurately identify the location and propagation path of the fault source. Especially in a multi-stage transmission system, the propagation of fault signals has complex spatio-temporal relationships. At the same time, due to the complex and changeable industrial environment, traditional methods are difficult to adapt to the change of fault patterns under different working conditions and lack an online learning and adaptive optimization mechanism, resulting in poor generalization ability of the diagnostic model.

[0004] In recent years, the development of artificial intelligence technology has provided new solutions for speed reducer fault diagnosis. The application of technologies such as deep learning, graph neural networks, and metric learning enables intelligent fault diagnosis systems to automatically extract multi-level features, build dynamic health state assessment models, and achieve high-precision fault location. However, the current application of artificial intelligence in speed reducer fault diagnosis still faces some challenges, such as the difficulty of multi-source heterogeneous data fusion, how to efficiently fuse multi-modal data such as vibration, temperature, and acoustic emission to improve the expression ability of fault features; the problem of small sample learning, industrial fault data is usually difficult to obtain and the data distribution is uneven, resulting in difficulty in training and promoting deep learning models; high real-time requirements, the diagnostic system in the industrial field needs to have real-time processing capabilities to meet the needs of online monitoring and early warning.

[0005] In view of the above problems, the present invention provides a method for fault location of a speed reducer based on artificial intelligence, which combines multi-source data acquisition, feature fusion, dynamic degradation assessment, deep metric learning, graph neural network analysis and online learning optimization to construct a high-precision and multi-level fault diagnosis and location system, thereby improving the accuracy, real-time performance and intelligent level of fault detection. Summary of the Invention

[0006] In order to solve the technical problems in the prior art, such as single signal source dependence, feature extraction relying on expert experience, lack of fault evolution trend modeling, difficulty in accurately identifying the location of the fault source, poor generalization ability of the diagnosis model, small sample learning problems, and high real-time requirements, the present invention provides a method and system for fault location of a speed reducer based on artificial intelligence.

[0007] The technical solutions provided by the present invention are as follows:

[0008] First aspect:

[0009] A method for fault location of a speed reducer based on artificial intelligence provided by the present invention includes:

[0010] S1. Multi-source data acquisition: Synchronously collect three-dimensional vibration signals, temperature gradient signals and broadband acoustic signals of the speed reducer through a vibration sensor array, a temperature sensor and an acoustic emission sensor, wherein the vibration sensor array includes at least 8 acceleration sensors symmetrically distributed in space;

[0011] S2. Signal preprocessing and feature fusion: Perform band-pass filtering, baseline correction and time-frequency synchronization processing on the collected original signals, and use an improved wavelet packet decomposition algorithm to extract the time-domain statistical features, frequency-domain energy features and non-linear dynamic features of each channel signal to construct a multi-dimensional feature matrix;

[0012] S3. Calculation of dynamic degradation parameters: Calculate the vibration feature degradation degree VFD and the operating state deviation index OSDI based on the feature matrix, where VFD reflects the coupling degradation degree of gear meshing state and bearing raceway damage, and OSDI characterizes the deviation amplitude of the overall operating state of the system relative to the reference healthy state;

[0013] S4. Fault feature space mapping: Construct a feature embedding network based on deep metric learning, project the multi-dimensional feature matrix into a low-dimensional fault feature space, and establish the corresponding relationship between fault types and feature distributions in this space;

[0014] S5. Multi-level fault location: Adopt a cascaded classification architecture. First, identify the major fault categories through a convolutional neural network CNN, then use a graph attention network GAT to analyze the spatial propagation path of the fault source, and finally determine the exact location of the fault component in combination with the spatio-temporal evolution laws of VFD and OSDI;

[0015] S6, Credibility Assessment and Feedback Optimization: Calculate the credibility index of the diagnostic result based on the D-S evidence theory. When the credibility is lower than the preset threshold, start the online learning mechanism and update the fault feature database using the incremental random forest algorithm.

[0016] Second aspect:

[0017] A reducer fault location system based on artificial intelligence provided by the present invention includes:

[0018] A multi-modal sensing module, an edge computing unit, a cloud analysis platform, an adaptive optimization module, and a human-computer interaction terminal.

[0019] The multi-modal sensing module includes a vibration sensor array with spatially symmetric distribution, an embedded temperature sensor, and a broadband acoustic emission sensor.

[0020] The edge computing unit is equipped with an FPGA chip and an ARM processor, and is used to implement the steps S1-S3 of the method described in the first aspect.

[0021] The cloud analysis platform deploys a deep metric learning network and a multi-level classifier, and is used to execute the steps S4-S5 of the method described in the first aspect.

[0022] The adaptive optimization module includes an incremental learning engine and a credibility assessment unit, and is used to execute the step S6 of the method described in the first aspect.

[0023] The human-computer interaction terminal provides a three-dimensional fault location visualization interface and maintenance decision-making suggestions.

[0024] The beneficial effects brought by the technical solution provided by the present invention at least include:

[0025] (1) In the present invention, a vibration sensor array, a temperature sensor, and an acoustic emission sensor are used to synchronously collect multi-modal signals of the reducer, and a high-dimensional feature matrix is constructed through an improved wavelet packet decomposition algorithm, time-frequency synchronization processing, and feature fusion technology. By this means, the present invention effectively fuses mechanical vibration, temperature gradient, and broadband acoustic information, improves the separability of fault features, and compared with traditional diagnostic methods relying on a single signal source, the present invention can comprehensively characterize the feature distribution of different types of faults and avoid the problems of missed detection and misjudgment caused by the limitations of a single sensor.

[0026] (2) In the present invention, a feature embedding network based on deep metric learning is constructed to project a multi-dimensional feature matrix into a low-dimensional fault feature space. At the same time, a graph attention network (GAT) is used to analyze the fault propagation path. Combining the vibration feature degradation degree (VFD) and the operating state deviation index (OSDI), accurate fault component localization is finally achieved. This method breaks through the bottleneck that traditional methods are difficult to analyze the fault propagation path in a multi-stage transmission system, significantly improves the fault tracing ability. Compared with traditional spectrum analysis and empirical mode decomposition methods, the present invention can effectively reduce the fault misjudgment rate under complex working conditions and improve the fault localization accuracy of a multi-stage gear transmission system;

[0027] (3) In the present invention, the D-S evidence theory is introduced to calculate the diagnostic credibility index. When the credibility is lower than the set threshold, the incremental random forest algorithm is triggered to achieve online learning and update the fault feature database. At the same time, a double-precision floating-point operation module is configured in the edge computing unit to ensure the real-time performance of complex calculations. This method overcomes the small-sample learning problem of traditional deep learning methods in industrial field applications, improves the adaptability of the diagnostic model under different working conditions. Through online learning optimization, the present invention can dynamically update fault features, improve the fault recognition accuracy during long-term use, and reduce the risk of diagnostic failure caused by working condition changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0029] Figure 1 It is a schematic flowchart of a method for fault localization of a reducer based on artificial intelligence provided by an embodiment of the present invention;

[0030] Figure 2 It is a schematic diagram of a difficult sample mining strategy for a method for fault localization of a reducer based on artificial intelligence provided by an embodiment of the present invention;

[0031] Figure 3 It is a schematic structural diagram of a system for fault localization of a reducer based on artificial intelligence provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The following will describe the technical solutions in the present invention with reference to the drawings.

[0033] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as an "example" in the present invention should not be construed as being more preferred or more advantageous than other embodiments or design solutions. Rather, the use of the word "example" is intended to present concepts in a specific manner. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one of the two.

[0034] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when the difference is not emphasized, the meanings they express are the same. "(of)", "corresponding", and "corresponding" can sometimes be used interchangeably. It should be noted that when the difference is not emphasized, the meanings they express are the same.

[0035] In the embodiments of the present invention, sometimes a subscript such as W1 may be miswritten as a non-subscript form such as W1. When the difference is not emphasized, the meanings they express are the same.

[0036] To make the technical problems to be solved, technical solutions and advantages of the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments.

[0037] Refer to the attached Figure 1 , which shows a schematic flowchart of a method for locating faults in a speed reducer based on artificial intelligence provided by an embodiment of the present invention.

[0038] The embodiments of the present invention provide a method for locating faults in a speed reducer based on artificial intelligence. This method can be implemented by a device for locating faults in a speed reducer based on artificial intelligence. This device for locating faults in a speed reducer based on artificial intelligence can be a terminal or a server. The processing flow of a method for locating faults in a speed reducer based on artificial intelligence can include the following steps:

[0039] S1. Multi-source data acquisition: Synchronously collect three-dimensional vibration signals, temperature gradient signals, and broadband acoustic signals of the speed reducer through a vibration sensor array, a temperature sensor, and an acoustic emission sensor, where the vibration sensor array includes at least 8 acceleration sensors symmetrically distributed in space.

[0040] In this embodiment, the specific implementation method of step S1 is as follows:

[0041] Arrange 8 PCB 356A32 type acceleration sensors on the surface of the speed reducer housing, symmetrically distributed in space: arrange 2 axial sensors at each of the input / output shaft ends, and arrange 2 radial sensors on both sides of the gearbox. The synchronous acquisition parameters are set as:

[0042] Vibration signal sampling frequency: 25.6 kHz (meeting the 3-fold sampling requirement of the Nyquist theorem for the gear meshing frequency of 5 kHz);

[0043] Sampling interval of the temperature sensor (PT100): 1 second;

[0044] Frequency band range of the acoustic emission sensor (PAC R15α): 50 - 400 kHz.

[0045] S2. Signal preprocessing and feature fusion: Perform band-pass filtering, baseline correction, and time-frequency synchronization processing on the collected original signals. Use the improved wavelet packet decomposition algorithm to extract the time-domain statistical features, frequency-domain energy features, and non-linear dynamic features of the signals in each channel, and construct a multi-dimensional feature matrix.

[0046] S3. Calculation of dynamic degradation parameters: Calculate the Vibration Feature Degradation (VFD) and the Operation State Deviation Index (OSDI) based on the feature matrix. Among them, VFD reflects the coupling degradation degree of the gear meshing state and the bearing raceway damage, and OSDI characterizes the deviation amplitude of the overall operation state of the system relative to the reference healthy state.

[0047] S4. Fault feature space mapping: Construct a feature embedding network based on deep metric learning, project the multi-dimensional feature matrix into a low-dimensional fault feature space, and establish the corresponding relationship between the fault type and the feature distribution in this space.

[0048] S5. Multi-level fault location: Adopt a cascaded classification architecture. First, identify the major fault categories through a Convolutional Neural Network (CNN), then use a Graph Attention Network (GAT) to analyze the spatial propagation path of the fault source, and finally determine the exact location of the faulty component by combining the spatio-temporal evolution laws of VFD and OSDI.

[0049] S6. Confidence evaluation and feedback optimization: Calculate the confidence index of the diagnostic result based on the D-S evidence theory. When the confidence is lower than the preset threshold, start the online learning mechanism and update the fault feature database using the incremental random forest algorithm.

[0050] In a possible implementation, the improved wavelet packet decomposition algorithm in S2 specifically includes:

[0051] Introduce an adaptive threshold denoising mechanism on the basis of the traditional wavelet packet decomposition, where the threshold T of the decomposition coefficient at the j-th layer j The calculation formula refers to the general threshold theory:

[0052]

[0053] where σ j represents the median absolute deviation of the wavelet coefficients of the j-th layer, and N j represents the number of wavelet coefficients of the j-th layer. At the same time, a dynamic compensation factor is added where E r is the energy ratio of adjacent decomposition layers, and the final threshold is adjusted to T' j = T j × (1 + γ).

[0054] In this embodiment, in step S2, the vibration signal is band-pass filtered at 50 Hz - 10 kHz to eliminate power grid interference and high-frequency noise.

[0055] In the improved wavelet decomposition, the db10 wavelet basis is selected for 5-layer decomposition, and the decomposition bandwidth is 800 Hz. Then the calculation of the adaptive threshold is as follows:

[0056] The median absolute deviation of the decomposition coefficients of the 3rd layer σ3 = 0.12

[0057]

[0058] Energy ratio of adjacent layers

[0059] The adjusted threshold T'3 = 0.43 × (1 + 0.71) = 0.74

[0060] The signal-to-noise ratio of the reconstructed signal is increased to 42 dB, which is 28% higher than the traditional method.

[0061] In a possible implementation manner, the calculation method of the vibration feature deterioration degree VFD in S3 specifically includes:

[0062]

[0063] where α and β are feature weight coefficients determined by the entropy weight method, satisfying α + β = 1, and E b is the energy of the current bearing feature frequency band, E0 is the energy of the corresponding frequency band in the reference state, and S k is the kurtosis value of the gear meshing sideband, S0 is the kurtosis threshold in the reference state, and C m is the envelope spectrum correlation coefficient, and C0 is the correlation coefficient threshold in the healthy state. The calculation method of the weight coefficient is:

[0064] β = 1 - α

[0065] where H(E b ) represents the information entropy of the bearing feature frequency band energy, and H(S k) represents the sideband kurtosis information entropy, and the calculation method of the bearing characteristic band energy information entropy is H(X) = -∑p(x)logp(x).

[0066] In this embodiment, the calculation is performed on the bearing inner ring fault sample:

[0067] Characteristic band energy: E b = 3.2×10 -3 V 2 , reference E0 = 1.5×10 -3 V 2

[0068] Sideband kurtosis: S k = 8.7, reference S0 = 3.2

[0069] Envelope spectrum correlation coefficient: C m = 0.62, healthy threshold C0 = 0.92

[0070] Weight calculation:

[0071] H(E b ) = 2.1 bits

[0072] H(S k ) = 1.8 bits

[0073]

[0074] Among them, it is calculated through the frequency band energy distribution histogram.

[0075]

[0076] 2.23 > 1.5, and it is determined as abnormal when exceeding the threshold.

[0077] In a possible implementation manner, the calculation method of the operating state deviation index OSDI in S3 specifically includes:

[0078]

[0079] Among them, w i represents the dynamic weight of the i-th characteristic parameter, d(F i , F i0 ) represents the Mahalanobis distance between the current characteristic parameter F i and the reference value F i0 , R i represents the allowable fluctuation range of the characteristic parameter, and the dynamic weight w i is calculated using the improved fuzzy entropy method:

[0080]

[0081] Among them, E i represents the singular spectrum entropy value of the i-th characteristic parameter, and E total represents the sum of the singular spectrum entropies of all characteristic parameters.

[0082] In the OSDI dynamic weight optimization part of this embodiment, it is necessary to select 5 key characteristic parameters:

[0083] Parameter <![CDATA[Singular spectrum entropy E i > <![CDATA[Dynamic weight w i > F1 0.85 0.24 F2 1.12 0.21 F3 0.68 0.28 F4 1.05 0.22 F5 0.92 0.25

[0084] The total entropy E total = 4.62, and each weight value is calculated to ensure that high-frequency features (such as F3) obtain higher weights.

[0085] In a possible implementation manner, the deep metric learning network in S4 adopts a triplet loss function, specifically including:

[0086] L = max(d(a, p) - d(a, n) + margin, 0)

[0087] where a is the anchor sample, that is, the current detection sample, p is the positive sample, that is, the same-class fault sample, n is the negative sample, that is, the different-class fault sample, d(·) is the cosine similarity distance in the feature space, margin is the preset boundary threshold, and a hard sample mining strategy is adopted during network training, and the most difficult positive sample pair and the most difficult negative sample pair are selected for optimization in each batch.

[0088] In a possible implementation manner, the calculation method of the Mahalanobis distance specifically includes:

[0089]

[0090] where Σ represents the covariance matrix of the historical health data of the characteristic parameters, and the covariance matrix is updated using the sliding window method during calculation, and the window length is L = 200 sampling periods.

[0091] In a possible implementation manner, the calculation method of the envelope spectrum correlation coefficient C m specifically includes:

[0092]

[0093] where E k represents the amplitude of the measured envelope spectrum at the k-th frequency point, and H k represents the amplitude of the envelope spectrum of the health state benchmark at the k-th frequency point, and are the means of the corresponding amplitudes respectively, and K represents the total number of analysis frequency bands, and the value range is [50, 200] times the meshing frequency.

[0094] In a possible implementation manner, such as Figure 2As shown, the difficult sample mining strategy specifically includes:

[0095] S501. For each anchor sample a, select the positive sample that satisfies arg max p d(a, p) from the same-class samples;

[0096] S502. Select the negative sample that satisfies arg min n d(a, n) from the different-class samples;

[0097] S503. Set the dynamic boundary threshold margin = μ·σ d , where μ is the proportionality coefficient with a default value of 1.2, and σ d is the standard deviation of the distances of the current batch of samples.

[0098] In this embodiment, the deep metric learning network is configured as follows:

[0099] Network structure:

[0100] Input layer (256 dimensions) → Fully connected layer (128 dimensions, ReLU) → Embedding layer (32 dimensions)

[0101] Triplet loss parameters: margin = 0.5, initial learning rate 1e-4.

[0102] Difficult sample mining:

[0103] Select the anchor sample a (bearing outer ring fault) from 1000 training samples;

[0104] Positive sample p: The sample with the lowest cosine similarity to a in the same class (d(a, p) = 0.82);

[0105] Negative sample: The sample with the highest similarity to a in different classes (gear tooth breakage) (d(a, n) = 0.65);

[0106] Dynamically adjust margin = 1.2×0.18 (current batch σ d = 0.15) → margin = 0.22.

[0107] Training results: The clustering diameter of the same-class samples is reduced to 0.35 (initial value 0.78), and the inter-class distance is expanded to 1.2 (initial value 0.5).

[0108] In the multi-level fault location part of this embodiment, the working process of the cascade classifier is specifically as follows:

[0109] CNN rough classification:

[0110] Input a 32×32 time-frequency diagram (generated by STFT of vibration signals)

[0111] The network structure is 2 convolutional layers (3×3 kernels) + MaxPooling → fully connected layer (Softmax output)

[0112] Output failure category probability (bearing failure: 87%, gear failure: 9%, others: 4%)

[0113] GAT spatial propagation analysis:

[0114] Construct a sensor node graph: 8 sensors are nodes, and edge weights are calculated by signal coherence

[0115] Attention coefficient calculation:

[0116]

[0117] Where d=32 is the feature dimension and a is the learnable parameter vector.

[0118] Positioning results: The fault energy propagates along nodes 3→5→7. Combined with the mutation points of the OSDI time series, it is determined that the fault source is the second-stage gear shaft (closest to sensor 3).

[0119] In a possible implementation, the credibility index calculation in S6 adopts DS evidence theory, which specifically includes:

[0120]

[0121] Among them, Bel(A i ) indicates fault type A i The credibility function value, Pl(A i ) indicates fault type A i The likelihood function value of m represents the total number of candidate fault types. When the credibility is lower than 0.85, the online learning mechanism is triggered.

[0122] In the online learning mechanism verification of this embodiment, when the credibility is detected to be 0.79 (lower than the threshold value 0.85):

[0123] 200 new samples were added to the training set, and the random forest was updated, with the number of trees increased from 100 to 120 and the maximum depth limited to 8 layers to trigger incremental learning. After the update, the recall rate of similar faults increased from 82% to 91%, and the false alarm rate decreased from 15% to 9%.

[0124] It should be noted that the sensor installation must ensure that the surface roughness Ra ≤ 3.2μm to avoid signal distortion.

[0125] It should be noted that during online learning, incremental data needs to be cleaned using the 3σ criterion to remove abnormal samples.

[0126] It should be noted that the edge computing unit FPGA needs to be configured with a double-precision floating-point operation module to ensure real-time performance.

[0127] The beneficial effects brought by the technical solution provided by the embodiment of the present invention at least include:

[0128] (1) In the present invention, a vibration sensor array, a temperature sensor, and an acoustic emission sensor are used to synchronously collect multi-modal signals of the reducer, and a high-dimensional feature matrix is constructed through an improved wavelet packet decomposition algorithm, time-frequency synchronization processing, and feature fusion technology. By this means, the present invention effectively fuses mechanical vibration, temperature gradient, and broadband acoustic information, improves the separability of fault features, and compared with the traditional diagnosis method relying on a single signal source, the present invention can comprehensively depict the feature distribution of different types of faults, avoiding the problems of missed detection and misjudgment caused by the limitations of a single sensor;

[0129] (2) In the present invention, a feature embedding network based on deep metric learning is constructed to project the multi-dimensional feature matrix into a low-dimensional fault feature space. At the same time, a graph attention network GAT is used to analyze the fault propagation path, combined with the vibration feature degradation degree VFD and the operating state deviation index OSDI, and finally accurate fault component localization is realized. This method breaks through the bottleneck that traditional methods are difficult to analyze the fault propagation path in a multi-stage transmission system, significantly improves the fault traceability ability, and compared with the traditional spectrum analysis and empirical mode decomposition methods, the present invention can effectively reduce the fault misjudgment rate under complex working conditions and improve the fault localization accuracy of a multi-stage gear transmission system;

[0130] (3) In the present invention, the D-S evidence theory is introduced to calculate the diagnostic credibility index. When the credibility is lower than the set threshold, the incremental random forest algorithm is triggered to realize online learning and update the fault feature database. At the same time, a double-precision floating-point operation module is configured in the edge computing unit to ensure the real-time performance of complex calculations. This method overcomes the small sample learning problem in the industrial field application of traditional deep learning methods, improves the adaptability of the diagnostic model under different working conditions, and through online learning optimization, the present invention can dynamically update fault features, improve the fault recognition accuracy in long-term use, and reduce the risk of diagnostic failure caused by working condition changes.

[0131] Refer to the attached Figure 3 , which shows the structural schematic diagram of a reducer fault localization system based on artificial intelligence provided by the embodiment of the present invention.

[0132] The present invention also provides a reducer fault localization system based on artificial intelligence, which is applied to the above-mentioned reducer fault localization method based on artificial intelligence, and includes:

[0133] A multi-modal sensing module, an edge computing unit, a cloud analysis platform, an adaptive optimization module, and a human-computer interaction terminal,

[0134] The multi-modal sensing module includes a vibration sensor array with spatially symmetric distribution, an embedded temperature sensor, and a broadband acoustic emission sensor;

[0135] The edge computing unit is equipped with an FPGA chip and an ARM processor, and is used to implement the steps S1 - S3 as in the method embodiment;

[0136] The cloud analysis platform deploys a deep metric learning network and a multi-level classifier, and is used to execute the steps S4 - S5 as in the method embodiment;

[0137] The adaptive optimization module includes an incremental learning engine and a credibility evaluation unit, and is used to execute the step S6 as in the method embodiment;

[0138] The human-computer interaction terminal provides a three-dimensional fault location visualization interface and maintenance decision-making suggestions.

[0139] In this embodiment, the multi-modal sensing module consists of a vibration sensor array, an embedded temperature sensor, and a broadband acoustic emission sensor. Among them, the vibration sensor uses a PCB 356A32 type acceleration sensor, which is arranged on the surface of the reducer housing in a spatially symmetric manner, including axial sensors at the input / output shaft ends and radial sensors on both sides of the gearbox. The temperature sensor selects PT100 to collect the temperature gradient signals of key parts, while the acoustic emission sensor (PAC R15α) is used to capture high-frequency structural impact signals. All sensor signals are uniformly collected by a high-precision synchronous acquisition card to ensure time-frequency alignment.

[0140] After data acquisition is completed, the signals are preprocessed by the edge computing unit, which integrates an FPGA chip and an ARM processor. It can perform operations such as band-pass filtering, baseline correction, and improved wavelet packet decomposition at a sampling rate of 25.6 kHz, and further calculate the vibration feature degradation degree (VFD) and the operating state deviation index (OSDI) to form a multi-dimensional feature matrix. The FPGA is responsible for real-time calculation of high-dimensional features, and the ARM is used to perform low-complexity statistical analysis.

[0141] The processed feature data is transmitted to the cloud analysis platform through the network. This platform deploys a deep metric learning network and a multi-level classifier. On this platform, the multi-dimensional feature matrix is projected into a low-dimensional fault feature space through a deep feature embedding network optimized by triplet loss, and the fault type identification and precise positioning are completed through a cascaded CNN and GAT model. The CNN first performs a rough classification on the time-frequency diagram of the vibration signal to identify the major fault categories, such as bearing faults, gear faults, or other faults. Subsequently, based on the spatial propagation path of the sensor signals, the GAT combines the spatio-temporal evolution laws of VFD and OSDI to analyze the fault propagation trend and accurately locate the faulty components.

[0142] In addition, the system of the present invention includes an adaptive optimization module, which includes an incremental learning engine and a credibility evaluation unit for optimizing the fault diagnosis performance in real time. When the cloud analysis platform calculates the fault diagnosis result, the D-S evidence theory is used to evaluate the credibility of the diagnosis result. If the credibility is lower than 0.85, the incremental learning mechanism is triggered. At this time, the incremental learning engine dynamically updates the random forest model, expands the training samples, continuously improves the fault feature library, and enhances the generalization ability of the model under different working conditions.

[0143] Finally, the human-machine interaction terminal of the system displays the fault location result in a three-dimensional visualization interface and provides maintenance decision-making suggestions. This terminal can run on the operator workstation, tablet computer or smart phone at the industrial site. Users can view the location of the fault source in real time, predict the remaining service life, and obtain targeted maintenance suggestions, such as replacing specific gears or bearings.

[0144] Through this embodiment, the present invention realizes the deep combination of multi-modal sensing, edge computing, cloud analysis and adaptive optimization, effectively improves the accuracy, real-time performance and adaptability of the reducer fault diagnosis, and provides strong support for the intelligent operation and maintenance of industrial equipment.

[0145] The beneficial effects brought by the technical solution provided by the embodiment of the present invention at least include:

[0146] (1) In the present invention, a vibration sensor array, a temperature sensor and an acoustic emission sensor are used to synchronously collect multi-modal signals of the reducer, and a high-dimensional feature matrix is constructed through an improved wavelet packet decomposition algorithm, time-frequency synchronization processing and feature fusion technology. By this means, the present invention effectively fuses mechanical vibration, temperature gradient and broadband acoustic information, improves the separability of fault features, and compared with the traditional diagnosis method relying on a single signal source, the present invention can comprehensively describe the feature distribution of different types of faults and avoid the problems of missed detection and misjudgment caused by the limitations of a single sensor.

[0147] (2) In the present invention, a feature embedding network based on deep metric learning is constructed to project the multi-dimensional feature matrix into a low-dimensional fault feature space. At the same time, a graph attention network (GAT) is used to analyze the fault propagation path, combined with the vibration feature degradation degree (VFD) and the operating state deviation index (OSDI). Finally, accurate fault component positioning is realized. This method breaks through the bottleneck that traditional methods are difficult to analyze the fault propagation path in a multi-stage transmission system, significantly improves the fault tracing ability, and compared with the traditional spectrum analysis and empirical mode decomposition methods, the present invention can effectively reduce the fault misjudgment rate under complex working conditions and improve the fault location accuracy of the multi-stage gear transmission system.

[0148] (3) In the present invention, the D-S evidence theory is introduced to calculate the diagnostic credibility index. When the credibility is lower than the set threshold, the incremental random forest algorithm is triggered to realize online learning and update the fault feature database. At the same time, a double-precision floating-point operation module is configured in the edge computing unit to ensure the real-time performance of complex calculations. This method overcomes the small-sample learning problem of traditional deep learning methods in industrial field applications, improves the adaptability of the diagnostic model under different working conditions. Through online learning optimization, the present invention can dynamically update fault features, improve the fault recognition accuracy in long-term use, and reduce the risk of diagnostic failure caused by working condition changes.

[0149] The above content is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

[0150] The following points need to be explained:

[0151] (1) The attached drawings of the embodiments of the present invention only relate to the structures involved in the embodiments of the present invention, and other structures can refer to the general design.

[0152] (2) For clarity, in the attached drawings used to describe the embodiments of the present invention, the thickness of layers or regions is enlarged or reduced, that is, these drawings are not drawn according to the actual scale. It can be understood that when an element such as a layer, film, region or substrate is referred to as being "on" or "under" another element, the element can be "directly" on or under the other element or there can be intermediate elements.

[0153] (3) Without conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other to obtain new embodiments.

[0154] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. The protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A speed reducer fault location method based on artificial intelligence, characterized in that: include: S1. Multi-source data acquisition: synchronously collect the three-dimensional vibration signal, temperature gradient signal and broadband acoustic signal of the reducer through the vibration sensor array, temperature sensor and acoustic emission sensor, wherein the vibration sensor array includes at least 8 acceleration sensors distributed symmetrically in space; S2, signal preprocessing and feature fusion: bandpass filtering, baseline correction and time-frequency synchronization processing are performed on the collected original signals, and the time domain statistical characteristics, frequency domain energy characteristics and nonlinear dynamic characteristics of each channel signal are extracted using the improved wavelet packet decomposition algorithm to construct a multi-dimensional feature matrix; S3. Dynamic degradation parameter calculation: Calculate the vibration characteristic degradation degree VFD and the operating state deviation index OSDI based on the characteristic matrix, where VFD reflects the coupling degradation degree of the gear meshing state and the bearing raceway damage, and OSDI represents the deviation amplitude of the overall operating state of the system relative to the benchmark health state; S4, fault feature space mapping: construct a feature embedding network based on deep metric learning, project the multi-dimensional feature matrix into the low-dimensional fault feature space, and establish the corresponding relationship between fault type and feature distribution in this space; S5. Multi-level fault location: Using a cascade classification architecture, we first use the convolutional neural network (CNN) to identify the major fault categories, then use the graph attention network (GAT) to analyze the spatial propagation path of the fault source, and finally determine the precise location of the faulty component by combining the spatiotemporal evolution of VFD and OSDI. S6. Credibility evaluation and feedback optimization: The credibility index of the diagnosis result is calculated based on the DS evidence theory. When the credibility is lower than the preset threshold, the online learning mechanism is started and the fault feature database is updated using the incremental random forest algorithm.

2. The method for locating reducer faults based on artificial intelligence according to claim 1 is characterized in that: The improved wavelet packet decomposition algorithm in S2 specifically includes: Based on the traditional wavelet packet decomposition, an adaptive threshold denoising mechanism is introduced, where the threshold T of the j-th layer decomposition coefficient is j The calculation formula refers to the general threshold theory: Among them, σ j represents the median absolute deviation of the j-th layer wavelet coefficients, N j Indicates the number of wavelet coefficients in the jth layer, and increases the dynamic compensation factor Where E r is the energy ratio of adjacent decomposition layers, and the final threshold is adjusted to T' j =T j ×(1+γ).

3. The method for locating reducer faults based on artificial intelligence according to claim 1 is characterized in that: The calculation method of the vibration characteristic degradation degree VFD in S3 specifically includes: Among them, α and β are the feature weight coefficients determined by the entropy weight method, satisfying α+β=1, E b is the characteristic frequency band energy of the current bearing, E0 is the frequency band energy corresponding to the reference state, S k is the kurtosis value of the gear meshing sideband, S0 is the reference state kurtosis threshold, C m is the envelope spectrum correlation coefficient, C0 is the health status correlation coefficient threshold, and the weight coefficient is calculated as follows: Among them, H(E b ) represents the energy information entropy of the bearing characteristic frequency band, H(S k ) represents the sideband kurtosis information entropy, and the calculation method of the bearing characteristic band energy information entropy is H(X)=-∑p(x)logp(x).

4. The method for locating reducer faults based on artificial intelligence according to claim 1 is characterized in that: The calculation method of the operating state deviation index OSDI in S3 specifically includes: Among them, w i represents the dynamic weight of the i-th feature parameter, d(F i ,F i0 ) represents the current feature parameter F i With the reference value F i0 The Mahalanobis distance, R i Represents the allowable fluctuation range of the characteristic parameters, where the dynamic weight w i The calculation of is based on the improved fuzzy entropy method: Among them, E i represents the singular spectral entropy value of the i-th characteristic parameter, E total represents the sum of the singular spectral entropies of all characteristic parameters.

5. The method for locating reducer faults based on artificial intelligence according to claim 1 is characterized in that: The deep metric learning network in S4 adopts a triplet loss function, specifically including: Among them, a is the anchor sample, that is, the current detection sample, p is the positive sample, that is, the same type of fault sample, n is the negative sample, that is, the heterogeneous fault sample, d(·) is the cosine similarity distance in the feature space, margin is the preset boundary threshold, and the difficult sample mining strategy is adopted during network training. The most difficult positive sample pair and the most difficult negative sample pair are selected for optimization in each batch.

6. The method for locating reducer faults based on artificial intelligence according to claim 4 is characterized in that: The calculation method of the Mahalanobis distance specifically includes: Among them, Σ represents the historical health data covariance matrix of the characteristic parameters. The sliding window method is used to update the covariance matrix during calculation, and the window length is L = 200 sampling periods.

7. The method for locating reducer faults based on artificial intelligence according to claim 3 is characterized in that: The envelope spectrum correlation coefficient C m The specific calculation methods include: Among them, E k represents the amplitude of the measured envelope spectrum at the kth frequency point, H k represents the amplitude of the healthy state reference envelope spectrum at the kth frequency point, and are the means of the corresponding amplitudes respectively, K represents the total number of analysis frequency bands, and its value range is [50,200] times the meshing frequency.

8. The method for locating reducer faults based on artificial intelligence according to claim 5 is characterized in that: The difficult sample mining strategy specifically includes: S501. For each anchor sample a, select samples of the same type that satisfy argmax p Positive samples of d(a,p); S502, select samples that satisfy argmin from heterogeneous samples n Negative samples of d(a,n); S503, set dynamic boundary threshold margin=μ·σ d , where μ is the proportionality coefficient, σ d is the standard deviation of the sample distances of the current batch.

9. The method for locating reducer faults based on artificial intelligence according to claim 1 is characterized in that: The credibility index calculation in S6 adopts DS evidence theory, which specifically includes: Among them, Bel(A i ) indicates fault type A i The credibility function value, Pl(A i ) indicates fault type A i The likelihood function value of m represents the total number of candidate fault types. When the credibility is lower than 0.85, the online learning mechanism is triggered.

10. A speed reducer fault location system based on artificial intelligence, comprising a multimodal sensing module, an edge computing unit, a cloud analysis platform, an adaptive optimization module and a human-computer interaction terminal, characterized in that: The multimodal sensing module comprises a spatially symmetrically distributed vibration sensor array, an embedded temperature sensor and a broadband acoustic emission sensor; The edge computing unit is equipped with an FPGA chip and an ARM processor, and is used to implement steps S1-S3 of the method according to any one of claims 1-9; The cloud analysis platform deploys a deep metric learning network and a multi-level classifier for performing steps S4-S5 of the method as claimed in claim 1; The adaptive optimization module includes an incremental learning engine and a credibility evaluation unit, which is used to execute step S6 of the method according to claim 1; The human-computer interaction terminal provides a three-dimensional fault location visualization interface and maintenance decision suggestions.

Images