Electromechanical fault prediction and diagnosis method and system based on big data

Abstract

The invention relates to an electromechanical fault prediction and diagnosis method and system based on big data, and the method comprises the steps: collecting the operation state data of electromechanical equipment in real time, and synchronously obtaining historical associated data; performing dynamic feature extraction on the operation state data and the historical associated data, constructing a sliding mean value feature matrix, and calculating dynamic weights of feature parameters; generating a fusion weight coefficient according to the dynamic weight and a preset fault threshold interval; extracting a distribution density curve of a historical fault occurrence probability, and calculating a dynamic threshold value; performing weighted reconstruction on the sliding mean feature matrix based on the fusion weight coefficient, and outputting a fault type and an occurrence probability through a pre-trained lightweight residual neural network model; and when the fault occurrence probability exceeds a dynamic threshold value adjusted based on a historical fault occurrence probability distribution density curve, generating an electromechanical fault diagnosis result so as to realize the purposes of real-time monitoring of the operation state of the electromechanical equipment and accurate fault prediction and diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of electromechanical equipment, and particularly relates to a method and system for predicting and diagnosing electromechanical faults based on big data. Background Art

[0002] With the intelligent upgrade of industrial equipment, the complexity and continuous operation requirements of electromechanical systems have increased significantly. Traditional fault diagnosis methods are facing severe challenges. Existing technologies mostly rely on fixed-threshold early warning and static weight models, which are difficult to adapt to dynamic changes such as equipment aging and environmental disturbances, resulting in high false alarm and missed alarm rates. At the same time, the utilization rate of historical data is insufficient, and there is a lack of quantitative correlation between fault characteristics and maintenance strategies. Maintenance decisions rely heavily on manual experience, with a lag in response and high costs. Although some existing technical solutions introduce machine learning algorithms, the model complexity does not match the edge computing resources, and real-time performance is difficult to guarantee. How to build an efficient and adaptive health management method for electromechanical equipment has become a key problem in the current industrial intelligent transformation. Summary of the Invention

[0003] The main purpose of the present invention is to provide a method and system for predicting and diagnosing electromechanical faults based on big data, which realizes the real-time monitoring of the operating state of electromechanical equipment, accurate fault prediction and diagnosis through multi-source data fusion analysis and intelligent algorithm modeling.

[0004] To achieve the above purpose, the present invention provides a method for predicting and diagnosing electromechanical faults based on big data, including the following steps: Collect the operating state data of electromechanical equipment in real time and synchronously obtain historical associated data; Extract dynamic features from the operating state data and historical associated data, construct a sliding mean feature matrix, and calculate the dynamic weights of each feature parameter; Generate a fusion weight coefficient according to the dynamic weight and a preset fault threshold interval; Extract the distribution density curve of the historical fault occurrence probability and calculate the dynamic threshold; Perform weighted reconstruction on the sliding mean feature matrix based on the fusion weight coefficient, and output the fault type and occurrence probability through a pre-trained lightweight residual neural network model; When the fault occurrence probability exceeds the dynamic threshold adjusted based on the historical fault occurrence probability distribution density curve, generate an electromechanical fault diagnosis result.

[0005] Further, the step of collecting the operating state data of electromechanical equipment in real time and synchronously obtaining historical associated data includes: Collect the operating state data of electromechanical equipment, including vibration signal data, temperature data, current data, and voltage data, and perform noise reduction processing on the vibration signal data through wavelet transform; Synchronously obtain historical associated data in the historical fault database of the electromechanical equipment, including historical environmental temperature and humidity data, equipment maintenance records, and load fluctuation data.

[0006] Further, the step of synchronously obtaining historical associated data further includes: Prioritize extracting data that meets the following conditions: The difference between the cumulative operating duration of the equipment at the time of data collection and the current duration ≤ ΔT, where ΔT is 20% - 30% of the mean time between failures of the equipment; The environmental temperature and humidity at the time of data collection are within the ±10% fluctuation range of the current environmental parameters.

[0007] Further, the steps of performing dynamic feature extraction on the operating state data and historical associated data, constructing a sliding mean feature matrix, and calculating the dynamic weights of each feature parameter include: Set the window length of the time window to 5 - 10 sampling periods, and the sliding step size to 1 - 2 sampling periods; Perform sliding mean calculation on the operating state data and historical associated data based on the time window to generate a sliding mean feature matrix; Calculate the dynamic weights of each feature parameter through the entropy weight method.

[0008] Further, the steps of generating a fusion weight coefficient according to the dynamic weight and a preset fault threshold interval include: Based on the fusion data set of the operating state data and the historical associated data, calculate the overall variance of each feature parameter; Adaptively adjust the bandwidth parameter of the Gaussian kernel function according to the variance, and weight the time decay factor based on the modified Gaussian kernel function to generate the decay weight of historical data; Perform product fusion on the dynamic weight and the decay weight to obtain the final fusion weight coefficient.

[0009] Further, the training steps of the pre-trained lightweight residual neural network model include: Extract the sliding mean feature matrix within the T time window before the occurrence of a fault from the historical fault database of the electromechanical equipment, and label the fault type and occurrence probability label, where the occurrence probability label is generated according to the exponential decay rule based on the time interval between the fault occurrence moment and the data collection moment; Adopt the cross-entropy loss function and L2 regularization constraint to jointly optimize the model parameters.

[0010] Further, the steps of extracting the distribution density curve of the historical fault occurrence probability and calculating the dynamic threshold include: Extract the distribution density curve of the historical fault occurrence probability and calculate the first derivative of the probability density; The inflection point where the first derivative changes from negative to positive is used as the trigger boundary for dynamic threshold adjustment to determine the dynamic threshold.

[0011] Further, the steps of weighted reconstruction of the sliding mean feature matrix based on the fusion weight coefficient and outputting the fault type and occurrence probability through a pre-trained lightweight residual neural network model include: Weight the sliding mean feature matrix based on the fusion weight coefficient; Normalize the weighted reconstructed sliding mean feature matrix to generate a reconstructed input feature vector; Extract fault features layer by layer through the multi-level residual blocks of the lightweight residual neural network and output the fault type and corresponding occurrence probability.

[0012] Further, the steps of generating an electromechanical fault diagnosis result when the fault occurrence probability exceeds the dynamic threshold adjusted based on the historical fault occurrence probability distribution density curve include: Match the sliding mean feature matrix corresponding to when the fault occurrence probability exceeds the dynamic threshold with the similar fault cases in the historical associated data; Screen out valid cases based on the dual screening conditions of Euclidean distance and elimination rate; Match the optimal repair path in the historical repair case library based on the valid cases, associate the fault location coordinates and mark the response priority of the repair strategy in combination with the topological structure model of the electromechanical equipment, and generate the electromechanical fault diagnosis result of the visual decision tree map.

[0013] The present invention also provides a big data-based electromechanical fault prediction and diagnosis system, including: A data acquisition unit for real-time collecting the operation status data of the electromechanical equipment and synchronously obtaining the historical associated data; A feature extraction unit for dynamically extracting features from the operation status data and the historical associated data, constructing a sliding mean feature matrix, and calculating the dynamic weights of each feature parameter; A weight generation unit for generating a fusion weight coefficient according to the dynamic weight and a preset fault threshold interval; A threshold calculation unit for extracting the distribution density curve of the historical fault occurrence probability and calculating the dynamic threshold; A fault prediction unit for weighted reconstruction of the sliding mean feature matrix based on the fusion weight coefficient and outputting the fault type and occurrence probability through a pre-trained lightweight residual neural network model; A result generation unit for generating an electromechanical fault diagnosis result when the fault occurrence probability exceeds the dynamic threshold adjusted based on the historical fault occurrence probability distribution density curve.

[0014] The electromechanical fault prediction and diagnosis method and system based on big data provided by the present invention have the following beneficial effects: Through the collaborative design of dynamic feature fusion and lightweight model, the present invention constructs a full-process closed-loop system for electromechanical fault prediction and diagnosis. By introducing the entropy weight method and time decay factor to dynamically fuse the weights of multi-source data, the sensitivity of traditional static models to equipment state drift is overcome, and the early warning accuracy under complex working conditions is significantly improved; based on the inflection point of historical probability density, the threshold is adaptively adjusted to accurately quantify the fault risk level and effectively reduce the misjudgment rate; a lightweight residual network and edge computing adaptation architecture are also adopted to greatly reduce the computing load while ensuring the model accuracy, meeting the industrial real-time requirements; and, through historical case matching and topological association, a visual decision tree map is generated to directly map the prediction results to executable maintenance strategies, breaking the information island between traditional diagnosis and maintenance links, and forming an intelligent closed loop from monitoring to maintenance. The present invention provides a highly reliable and low-latency integrated solution for electromechanical equipment health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a schematic flowchart of the electromechanical fault prediction and diagnosis method based on big data in an embodiment of the present invention; Figure 2 is a structural block diagram of the electromechanical fault prediction and diagnosis system based on big data in an embodiment of the present invention.

[0016] The realization, functional characteristics and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0018] Referring to Figure 1 , which is a schematic flowchart of a method for predicting and diagnosing electromechanical faults based on big data proposed by the present invention, includes the following steps: S1, Real-time collect the operation state data of the electromechanical equipment and synchronously obtain the historical associated data; S2, Perform dynamic feature extraction on the operation state data and historical associated data, construct a sliding mean feature matrix, and calculate the dynamic weights of each feature parameter; S3, Generate a fusion weight coefficient according to the dynamic weight and a preset fault threshold interval; S4, Extract the distribution density curve of the historical fault occurrence probability and calculate the dynamic threshold; S4. Reconstruct the sliding mean feature matrix with weighted coefficients based on the fusion weight coefficients, and output the fault type and occurrence probability through a pre-trained lightweight residual neural network model; S5. If the fault occurrence probability exceeds the dynamic threshold, generate an electromechanical fault diagnosis result.

[0019] In one embodiment, for step S1, The steps of collecting the operation status data of the electromechanical equipment in real time and synchronously obtaining the historical correlation data include: Collect the operation status data of the electromechanical equipment, including vibration signal data, temperature data, current data, and voltage data, and perform noise reduction processing on the vibration signal data through wavelet transform; Synchronously obtain the historical correlation data in the historical fault database of the electromechanical equipment, including historical environmental temperature and humidity data, equipment maintenance records, and load fluctuation data.

[0020] In the specific implementation process, the selection of the operation status data is based on the multi-dimensional fault feature capture requirements: vibration signals are used to detect mechanical wear (for example, the energy in the 5kHz frequency band increases by 30% 2 hours before a bearing fault), temperature is monitored for winding overheating (sensitivity ±0.5°C / min), and current / voltage is analyzed for electrical abnormalities (THD>5% or voltage sag ≥10%). For the high-frequency noise problem of vibration signals, traditional Fourier transform cannot separate noise from high-frequency fault features (frequency domain aliasing). This technical solution uses adaptive threshold wavelet noise reduction to dynamically adjust the threshold according to the energy distribution of vibration signals: , where is the first-layer wavelet detail coefficient, is the signal length, and λ is the threshold. Perform 5-layer decomposition through the Daubechies 6 wavelet basis, and combine the hard threshold and the soft threshold to improve the signal-to-noise ratio (SNR) and increase the kurtosis value of the fault features. The synchronous acquisition of historical correlation data focuses on environmental temperature and humidity, maintenance records, and load fluctuation parameters. When the environmental humidity>70%, the probability of insulation failure increases by 3 times. Maintenance records can optimize the decision-making priority (for example, the bearing failure rate decreases by 60% within 30 days after lubrication). Load fluctuation is significantly correlated with the current harmonic distortion rate (r = 0.82). To achieve accurate screening of historical data, define the dynamic time window ΔT as 20%-30% of the mean time between failures (MTBF) of the equipment. The calculation formula is: , for example, for an equipment with MTBF = 2000 hours, ΔT = 500 hours, and constrain the historical environmental temperature and humidity fluctuation range within ±10% to avoid seasonal interference. S1 ensures data quality and historical comparability through hardware configuration (such as a piezoelectric accelerometer with a sampling rate of 20kHz) and dynamic rules (ΔT + temperature and humidity filtering), providing reliable input for subsequent feature fusion and model inference.

[0021] Experimental data of vibration signal noise reduction:

[0022] In one embodiment, for step S2, The steps of performing dynamic feature extraction on the operation state data and historical associated data, constructing a sliding mean feature matrix, and calculating the dynamic weights of each feature parameter include: Set the window length of the time window to 5 - 10 sampling periods, and the sliding step size to 1 - 2 sampling periods; Based on the time window, perform sliding mean calculation on the operation state data and historical associated data to generate a sliding mean feature matrix; Calculate the dynamic weights of each feature parameter by the entropy weight method.

[0023] In the specific implementation process, the time window length is set to 5 - 10 sampling periods (for example, when the sampling interval is 1 minute, the window covers 5 - 10 minutes of data), and the sliding step size is 1 - 2 periods. Its design basis is the time scale of the evolution of electromechanical fault characteristics. Experiments show that the abnormal signal at the initial stage of bearing wear lasts about 8 minutes (corresponding to a window length of 8 periods), while the current fluctuation caused by sudden load change reaches the peak within 2 - 3 minutes (corresponding to a sliding step size of 2 periods). The calculation formula of the sliding mean feature matrix is: , where L is the window length, is the value of the th feature at the th sampling point. The rows of the matrix represent the time window, and the columns represent the feature parameters. The instantaneous noise can be suppressed through this matrix. The entropy weight method is used to calculate the dynamic weights, and the information entropy of the feature parameters is quantified to reflect the data uncertainty, and the information entropy is evaluated and the weights are assigned. Information entropy calculation formula: , where is the information entropy of the th feature parameter, and its value range is [0, 1], which is used to quantify the data chaos degree of this feature (the larger the entropy value, the more dispersed the information and the lower the contribution to fault prediction); N is the total number of time windows (for example, if the total data duration is 60 minutes and the window length is 5 minutes, then N = 12); is the normalized value of the th feature parameter within the th time window, satisfying , and the calculation formula , where is the original feature value; is the normalization coefficient to ensure the comparability of entropy values under different window numbers. The importance of features is dynamically judged through the entropy value For example, the entropy value of the vibration signal is low ( ), it is sensitive to mechanical failures and is given a high weight; when the entropy value of the vibration signal is high ( ), the environmental interference is large and a low weight is given, and then the weights are allocated. The weight allocation formula is , where is the dynamic weight of the th characteristic parameter, and the value range is (0, 1), and the sum of all weights is 1 ( ); m is the total number of characteristic parameters (for example, there are 4 characteristic parameters of vibration, temperature, current, and voltage in total, then m = 4); is the information entropy of the kth characteristic; is the sum of the information entropy inversion values of all characteristics, which is used to normalize the weights. By

[0024] converting the entropy value into an information utility value, the lower the entropy, the higher the utility, and by normalizing to ensure the fairness of weight allocation and avoid weight deviation caused by differences in the number of characteristics.

[0025] The above results show that when the window length is 8 cycles and the step size is 2 cycles, the optimal balance is achieved between delay and accuracy. The dynamic weights of the entropy weight method further solve the problem of feature redundancy (for example, when the correlation between current harmonics and vibration signals > 0.7, the current weight is automatically reduced by 30%), thereby improving the generalization ability of the model. Step S2 realizes dynamic feature extraction and weight allocation through the time window mechanism and the entropy weight method, and solves the problem that traditional static features cannot capture the gradual change of device state.

[0026] In one embodiment, for step S3, The step of generating a fusion weight coefficient according to the dynamic weight and a preset fault threshold interval includes: Calculating the overall variance of each characteristic parameter based on the fusion data set of the operating state data and the historical correlation data; Adapting and adjusting the bandwidth parameter of the Gaussian kernel function according to the variance, and weighting the time decay factor based on the modified Gaussian kernel function to generate the decay weight of the historical data; Multiplying and fusing the dynamic weight and the decay weight to obtain the final fusion weight coefficient.

[0027] In the specific implementation process, based on the fusion data set of the operating state data and the historical correlation data, calculate the overall variance of each characteristic parameter, and the calculation formula is , where is the overall variance of the characteristic parameter , which reflects the overall volatility of the real-time and historical fusion data set; is the A sampling point feature set; is the characteristic mean, and N is the total number of samples (covering the real-time data window and the historical data period). Variance calculation can quantify the volatility of the equipment state. The larger the variance, the more severe the fluctuation of the equipment state (such as sudden load change or mechanical loosening). The system will automatically adjust the subsequent processing parameters. Based on the calculated variance, the bandwidth parameter of the Gaussian kernel function is adaptively adjusted. , to avoid interference from outdated data. The bandwidth parameter is used to control the time decay rate of historical data ,in, is the bandwidth of the Gaussian kernel function, which is positively correlated with the square root of the variance; is the decay weight of historical data at timestamp t, is the current time; An exponential function used to quantify the time decay effect. It calculates the decay weight of historical data at timestamp t. When device state fluctuations are large (high variance), increasing the bandwidth parameter σ makes the decay curve smoother, preserving more of the contribution of recent historical data. Conversely, decreasing the bandwidth accelerates the decay of the weight of more distant historical data.

[0028] Experimental data:

[0029] The dynamic weight obtained in step S2 With decay weight Perform product fusion to obtain the fusion weight coefficient: , integrating the contribution of spatiotemporal dimension features.

[0030] Experimental comparison:

[0031] Step S3 solves the problem of insufficient correlation between real-time data and historical data in traditional methods by integrating dynamic weights with historical data attenuation weights.

[0032] In one embodiment, the training steps of the pre-trained lightweight residual neural network model include: Extracting a sliding mean feature matrix within a time window T before the fault occurs from a historical fault database of the electromechanical equipment, and labeling the fault type and occurrence probability label, wherein the occurrence probability label is generated according to an exponential decay rule based on the time interval between the fault occurrence time and the data collection time; The cross entropy loss function and L2 regularization constraint are used to jointly optimize the model parameters.

[0033] Specifically, the sliding mean feature matrix of the T time window (usually 5-10 sampling cycles) before the fault occurs is extracted from the historical fault database. The matrix contains key parameters such as vibration and temperature. The training label is generated using the exponential decay rule, and the formula is , where Δt represents the time difference between the data acquisition moment and the fault occurrence moment, λ is the attenuation coefficient (usually taken as 0.05 - 0.1), which gives higher weights to the data near the fault moment and effectively guides the model to focus on key features. The model training adopts a joint optimization strategy, combining a weighted cross-entropy loss function and L2 regularization constraints. A weight coefficient inversely proportional to the fault category frequency is introduced into the cross-entropy loss function. For example, a weight of 9.0 is assigned to the bearing fault samples that only account for 10%, significantly improving the data imbalance problem. The L2 regularization term prevents overfitting by constraining the weight magnitude (λ is usually taken as 1e-4~1e-5). To adapt to the edge computing environment, the trained model also undergoes channel pruning and 8-bit quantization processing, removing the feature channels with low contribution degrees and compressing the number of parameters to 30% - 40% of the original.

[0034] In one embodiment, for step S4, The steps of extracting the distribution density curve of the historical fault occurrence probability and calculating the dynamic threshold include: Extracting the distribution density curve of the historical fault occurrence probability and calculating the first derivative of the probability density; Taking the inflection point where the first derivative changes from negative to positive as the trigger boundary for dynamic threshold adjustment to determine the dynamic threshold.

[0035] Specifically, to extract the distribution density curve of the historical fault occurrence probability, the kernel density estimation algorithm is used to perform non-parametric modeling on the historical probability data. The formula is , where, is the historical fault occurrence probability value, is the bandwidth parameter (usually taken as 0.05 - 0.1), and K is the Gaussian function. The constructed probability density curve reflects the distribution characteristics of the fault probability. For example, the historical data of a certain type of water pump shows that the fault probability presents a bimodal distribution in the interval [0.65, 0.82], corresponding to two main fault modes of bearing wear and winding overheating. Based on the density curve, the first derivative is further calculated, and the inflection point is located through the change of the derivative sign. When the first derivative changes from negative to positive, it corresponds to the critical point where the probability density curve changes from decreasing to increasing, and this point represents the mutation boundary of the fault probability distribution. For example, in a case of a fan bearing fault, the density curve has an inflection point at , indicating that when the real-time prediction probability exceeds this value, the fault risk shows an exponential upward trend. In one embodiment, for step S5, The steps of reweighting and reconstructing the sliding mean feature matrix based on the fusion weight coefficient and outputting the fault type and occurrence probability through a pre-trained lightweight residual neural network model include: Reweighting the sliding mean feature matrix based on the fusion weight coefficient; Normalize the sliding mean feature matrix after weighted reconstruction to generate the reconstructed input feature vector; Extract fault features layer by layer through the multi-level residual blocks of the lightweight residual neural network, and output the fault type and the corresponding occurrence probability.

[0036] In the specific implementation process, weight the sliding mean feature matrix based on the fusion weight coefficient. The weighting formula is , where ⊙ represents element-wise multiplication of matrices, is the fusion weight coefficient matrix generated in step S3, is the sliding mean feature matrix. By weighting, key features can be dynamically strengthened. After weighted reconstruction, the amplitude range of the current signal (0 - 10A) is much larger than that of the temperature (20 - 100 °C). When not normalized, the model is easily dominated by the current. Therefore, it is necessary to normalize the feature matrix to eliminate the dimensional difference. The calculation formula is , where μ and σ are the matrix mean and standard deviation respectively. Input the normalized feature vector into the pre-trained lightweight residual neural network, and extract fault features layer by layer through the multi-level residual blocks. Each residual block contains two convolutional layers with a kernel size of 3×1 and skip connections. After channel pruning, the number of parameters is reduced by 62% (for example, from 64 channels to 24 channels), and then compressed to 1 / 4 of the original volume through 8-bit fixed-point quantization. The deep residual block is dedicated to capturing the periodic features of faults, such as the energy distribution of the impact signal of bearing wear in the time-frequency domain, and combines with the temperature-current correlation features extracted by the middle layer block to form a comprehensive fault criterion. Compared with the traditional ResNet, the lightweight design can improve the accuracy while reducing the number of parameters. S5 strengthens the key period features through the fusion weight, normalizes to eliminate the dimensional deviation, and the lightweight network takes into account both accuracy and efficiency.

[0037] In one embodiment, for step S6, When the fault occurrence probability exceeds the dynamic threshold adjusted based on the historical fault occurrence probability distribution density curve, the steps to generate the electromechanical fault diagnosis result include: Match the sliding mean feature matrix corresponding to when the fault occurrence probability exceeds the dynamic threshold with the similar fault cases in the historical associated data; Based on the dual screening conditions of Euclidean distance and elimination rate, screen out the effective cases; Based on the effective cases, match the optimal repair path in the historical repair case library, combine the fault location coordinates associated with the topological structure model of the electromechanical equipment, mark the response priority of the repair strategy, and generate the electromechanical fault diagnosis result of the visual decision tree map.

[0038] In the specific implementation process, when the real-time failure probability exceeds the dynamic threshold adjusted based on the historical distribution density curve, the current sliding mean feature matrix is extracted (such as a window length of 8 cycles, including parameters such as the vibration signal mean of 0.48 m / s² and the temperature of 68°C), and similarity matching is performed with the historical case library. The matching process uses dual screening conditions: Euclidean distance screening, calculating the Euclidean distance between the current matrix and historical cases , retain the cases of; only select the cases where the actual elimination rate of the historical maintenance strategy is ≥90% to ensure the effectiveness of the recommendation (for example, in a certain pump case, the failure recurrence time of the high elimination rate strategy exceeds 6 months). The effective cases passed through the screening will be sorted by priority to match the optimal maintenance path, and the sorting rule is: the smallest Euclidean distance > the highest elimination rate > the shortest maintenance time. For example, a certain bearing failure matches 3 paths, and path A (distance 0.12, elimination rate 95%, time-consuming 2 hours) is recommended first. At the same time, based on the equipment topology structure model (such as CAD or digital twin), the fault location coordinates are associated, the vibration anomaly is mapped to the bearing at the motor drive end, and the fault point is highlighted in red on the visualization interface, the maintenance-related components are marked in green (such as the end cover, bearing housing), and the associated subsystems are prompted in yellow (such as the cooling fan). The generated decision tree spectrum contains a multi-level structure: the fault type and probability on the root node (such as "bearing wear, probability 92%"); the historical cases arranged by priority on the first-level branch, marking the similarity and elimination rate; the detailed maintenance steps on the second-level branch (dismantle the end cover → measure the clearance → replace the grease); the spare parts list (such as bearing model), tool requirements (such as torque wrench), and safety specifications (such as power-off locking process) on the leaf node. Each maintenance result is fed back to the case library through the dynamic update mechanism to automatically correct the elimination rate data. For example, after a certain maintenance fails, the system lowers the priority of this path and triggers incremental training of the model. Step S6 realizes the closed-loop management from fault warning to maintenance strategy through the dynamic threshold trigger mechanism and data-driven decision tree generation

[0039] Refer to Figure 2 , for the structural block diagram of the electromechanical fault prediction and diagnosis system based on big data in an embodiment of the present invention, including: The data acquisition unit is used to collect the operation status data of the electromechanical equipment in real time and synchronously obtain the historical associated data; The feature extraction unit is used to perform dynamic feature extraction on the operation status data and historical associated data, construct a sliding mean feature matrix, and calculate the dynamic weights of each feature parameter; The weight generation unit is used to generate a fusion weight coefficient according to the dynamic weight and the preset fault threshold interval; The threshold calculation unit is used to extract the distribution density curve of the historical fault occurrence probability and calculate the dynamic threshold;[[ID=ID=18]] A fault prediction unit, configured to perform weighted reconstruction on the sliding mean feature matrix based on the fusion weight coefficient, and output the fault type and occurrence probability through a pre-trained lightweight residual neural network model; A result generation unit, configured to generate an electromechanical fault diagnosis result when the fault occurrence probability exceeds a dynamic threshold adjusted based on the historical fault occurrence probability distribution density curve.

[0040] For the specific implementation of each unit in the above device example, please refer to that described in the above method embodiment, and details are not repeated herein.

[0041] In summary, the present invention collects the operation state data of electromechanical equipment in real time and synchronously obtains historical associated data; performs dynamic feature extraction on the operation state data and historical associated data, constructs a sliding mean feature matrix, and calculates the dynamic weights of each feature parameter; generates a fusion weight coefficient according to the dynamic weights and a preset fault threshold interval; extracts the distribution density curve of the historical fault occurrence probability, and calculates the dynamic threshold; performs weighted reconstruction on the sliding mean feature matrix based on the fusion weight coefficient, and outputs the fault type and occurrence probability through a pre-trained lightweight residual neural network model; when the fault occurrence probability exceeds the dynamic threshold adjusted based on the historical fault occurrence probability distribution density curve, generates an electromechanical fault diagnosis result, so as to achieve the purpose of real-time monitoring, accurate fault prediction and diagnosis of the operation state of electromechanical equipment.

[0042] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above method embodiments. Among them, any reference to the memory, storage, database or other media provided by the present invention and used in the embodiments can include non-volatile and / or volatile memories. The non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. The volatile memory can include random access memory (RAM) or an external cache. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0043] It should be noted that in this document, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article or method comprising a series of elements not only includes those elements but also other elements not expressly listed, or further includes elements inherent to such process, apparatus, article or method. Without further limitation, an element defined by the phrase "comprising an..." does not exclude the presence of additional identical elements in the process, apparatus, article or method comprising such element.

[0044] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be similarly included in the patent protection scope of the present invention.

Claims

1. A method for predicting and diagnosing mechanical and electrical faults based on big data, characterized in that, Including the following steps: Collect the operation status data of the electromechanical equipment in real time and synchronously obtain the historical associated data; Perform dynamic feature extraction on the operation status data and historical associated data, construct a sliding mean feature matrix, and calculate the dynamic weights of each feature parameter; Generate a fusion weight coefficient according to the dynamic weight and a preset fault threshold interval; Extract the distribution density curve of the historical fault occurrence probability and calculate the dynamic threshold; Perform weighted reconstruction on the sliding mean feature matrix based on the fusion weight coefficient, and output the fault type and occurrence probability through a pre-trained lightweight residual neural network model; When the fault occurrence probability exceeds the dynamic threshold adjusted based on the historical fault occurrence probability distribution density curve, generate an electromechanical fault diagnosis result.

2. The method for predicting and diagnosing mechanical and electrical faults based on big data according to claim 1, wherein, The step of collecting the operation status data of the electromechanical equipment in real time and synchronously obtaining the historical associated data includes: Collect the operation status data of the electromechanical equipment, including vibration signal data, temperature data, current data, and voltage data, and perform noise reduction processing on the vibration signal data through wavelet transform; Synchronously obtain the historical associated data in the historical fault database of the electromechanical equipment, including historical ambient temperature and humidity data, equipment maintenance records, and load fluctuation data.

3. The method for predicting and diagnosing mechanical and electrical faults based on big data according to claim 1, wherein The step of synchronously obtaining the historical associated data further includes: Prioritize extracting data that meets the following conditions: The difference between the cumulative operation duration of the equipment at the time of data collection and the current duration ≤ ΔT, where ΔT is 20%-30% of the average fault interval time of the equipment; The ambient temperature and humidity at the time of data collection are within the ±10% fluctuation range of the current ambient parameters.

4. The method for predicting and diagnosing mechanical and electrical faults based on big data according to claim 1, wherein The step of performing dynamic feature extraction on the operation status data and historical associated data, constructing a sliding mean feature matrix, and calculating the dynamic weights of each feature parameter includes: Set the window length of the time window to 5-10 sampling periods and the sliding step to 1-2 sampling periods; Perform sliding mean calculation on the operation status data and historical associated data based on the time window to generate a sliding mean feature matrix; Calculate the dynamic weights of each feature parameter through the entropy weight method.

5. The method for predicting and diagnosing mechanical and electrical faults based on big data according to claim 1, characterized in that The step of generating a fusion weight coefficient according to the dynamic weight and a preset fault threshold interval includes: Based on the fusion data set of the operation status data and the historical associated data, calculate the overall variance of each feature parameter; Adaptively adjust the bandwidth parameter of the Gaussian kernel function according to the variance, and weight the time decay factor based on the modified Gaussian kernel function to generate the decay weight of the historical data; Perform product fusion on the dynamic weight and the decay weight to obtain the final fusion weight coefficient.

6. The method for predicting and diagnosing mechanical and electrical faults based on big data according to claim 1, wherein The training steps of the pre-trained lightweight residual neural network model include: Extract the sliding mean feature matrix within the T time window before the fault occurrence from the historical fault database of the electromechanical equipment, and label the fault type and occurrence probability label, where the occurrence probability label is generated according to the exponential decay rule based on the time interval between the fault occurrence moment and the data collection moment; Jointly optimize the model parameters using the cross-entropy loss function and L2 regularization constraint.

7. The method for predicting and diagnosing mechanical and electrical faults based on big data according to claim 1, characterized in that, The step of extracting the distribution density curve of the historical fault occurrence probability and calculating the dynamic threshold includes: Extract the distribution density curve of the historical fault occurrence probability and calculate the first derivative of the probability density; Take the inflection point where the first derivative changes from negative to positive as the trigger boundary for dynamic threshold adjustment to determine the dynamic threshold.

8. The method for predicting and diagnosing mechanical and electrical faults based on big data according to claim 1, wherein, The step of weighted reconstruction of the sliding mean feature matrix based on the fusion weight coefficient and outputting the fault type and occurrence probability through a pre-trained lightweight residual neural network model includes: Weight the sliding mean feature matrix based on the fusion weight coefficient; Perform normalization processing on the weighted and reconstructed sliding mean feature matrix to generate a reconstructed input feature vector; Extract fault features layer by layer through the multi-level residual blocks of the lightweight residual neural network and output the fault type and the corresponding occurrence probability.

9. The method for predicting and diagnosing mechanical and electrical faults based on big data according to claim 1, wherein The step of generating an electromechanical fault diagnosis result when the fault occurrence probability exceeds the dynamic threshold adjusted based on the historical fault occurrence probability distribution density curve includes: Match the similarity between the sliding mean feature matrix corresponding to the fault occurrence probability exceeding the dynamic threshold and the same type of fault cases in the historical associated data; Screen out valid cases based on the double screening conditions of Euclidean distance and elimination rate; Match the optimal maintenance path in the historical maintenance case library based on the valid cases, combine the topological structure model of the electromechanical equipment to associate the fault location coordinates, mark the response priority of the maintenance strategy, and generate a visual decision tree map of the electromechanical fault diagnosis result.

10. An electromechanical fault prediction and diagnosis system based on big data, characterized in that, Include: A data acquisition unit for real-time collecting the operation status data of the electromechanical equipment and synchronously obtaining the historical associated data; A feature extraction unit for dynamically extracting features from the operation status data and the historical associated data, constructing a sliding mean feature matrix, and calculating the dynamic weights of each feature parameter; A weight generation unit for generating a fusion weight coefficient according to the dynamic weight and a preset fault threshold interval; A threshold calculation unit for extracting the distribution density curve of the historical fault occurrence probability and calculating the dynamic threshold; A fault prediction unit for performing weighted reconstruction of the sliding mean feature matrix based on the fusion weight coefficient and outputting the fault type and occurrence probability through a pre-trained lightweight residual neural network model; A result generation unit for generating an electromechanical fault diagnosis result when the fault occurrence probability exceeds the dynamic threshold adjusted based on the historical fault occurrence probability distribution density curve.

Images