Speed reducer fault prediction system and method based on deep learning algorithm

Abstract

The invention discloses a speed reducer fault prediction system and method based on a deep learning algorithm, and relates to the technical field of mechanical equipment fault diagnosis and prediction, the fault prediction system comprises a data acquisition module, a data processing module, a feature extraction module, a fault prediction model module and an early warning module; the technical key points are as follows: based on a time sequence deep learning model, through an LSTM unit layer, a GRU unit layer and an attention mechanism layer, performing deep analysis on an input multi-dimensional feature vector, and when the model predicts a potential fault, immediately comparing a prediction result with an early warning threshold system; if the corresponding early warning condition is triggered, early warning information containing fault details and countermeasures is generated; the accurate prediction and intelligent early warning mechanism enables equipment maintenance to be changed from passive coping to active prevention, can arrange a maintenance plan in advance, avoids faults, greatly reduces equipment downtime and maintenance cost, and remarkably improves production efficiency and equipment reliability.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment fault diagnosis and prediction technology, specifically to a speed reducer fault prediction system and method based on deep learning algorithms. Background Technology

[0002] In modern smart manufacturing plants, large CNC machine tools, automated production lines, and other mechanical equipment operate 24 hours a day, undertaking core production tasks such as precision parts processing and product assembly. Once these devices suddenly malfunction, the entire production line will be forced to stop, which will not only delay the delivery of orders worth millions of dollars that day, but also scrap raw materials already put into production due to equipment failure, causing huge direct economic losses. At the same time, the production interruption caused by the failure will also disrupt subsequent logistics, product sales, and other links, with indirect losses that are difficult to estimate. In mining scenarios, crushers, conveyor belts, and other mechanical equipment operate under high load and harsh conditions for a long time. When a failure occurs, it will not only interrupt ore mining operations, but may even cause serious safety accidents such as collapses and mechanical injuries, threatening the lives of workers.

[0003] Traditional mechanical equipment fault diagnosis mainly relies on manual inspection and simple sensor monitoring. In a certain automobile manufacturing plant, inspection workers need to manually check hundreds of pieces of equipment in the workshop every two hours, judging the equipment status by listening to the operating sound and touching the surface temperature. This method is not only extremely inefficient, but also greatly affected by the worker's experience and subjective judgment, making it difficult to detect early potential faults. Even with simple sensor monitoring, only limited basic operating parameters of the equipment can be obtained, which cannot be used to conduct in-depth analysis of complex fault phenomena, resulting in poor fault diagnosis accuracy and frequent misdiagnosis and missed diagnosis.

[0004] With the development of industrial and artificial intelligence technologies, data-driven fault prediction methods have gradually emerged. However, in practical applications, existing methods still have many problems when processing operational data of complex mechanical equipment. In the field of aero-engine fault prediction, due to the nonlinear and non-stationary characteristics of vibration, temperature, and other data generated during engine operation, existing fault prediction models struggle to fully extract effective features, resulting in insufficient prediction accuracy and failing to meet the stringent requirements of aviation safety for early fault warnings. In wind power equipment, due to the complex and variable wind environment and significant differences in equipment operating conditions, existing warning threshold setting methods cannot be dynamically adjusted according to actual operating conditions, often leading to frequent false alarms or missed warnings of important faults, greatly affecting equipment maintenance efficiency and operational reliability. These problems make existing fault prediction technologies unable to meet the stringent requirements for high reliability and high safety of mechanical equipment in actual production, necessitating a more advanced and accurate fault prediction system and method. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides the following technical solution: A gearbox fault prediction system based on deep learning algorithms, comprising: The data acquisition module is located at key structural positions in mechanical equipment, such as bearing housings, gearboxes, and drive shafts. It consists of a distributed sensor array comprising piezoelectric vibration sensors, infrared temperature sensors, magnetoelectric speed sensors, and strain gauge load sensors. The piezoelectric vibration sensor, based on the piezoelectric effect, converts the acceleration signal generated by mechanical vibration into an electrical signal. The infrared temperature sensor obtains the surface temperature by detecting the infrared radiation energy of an object. The magnetoelectric speed sensor uses the principle of electromagnetic induction to convert rotational speed into a pulse signal. The strain gauge load sensor reflects the load magnitude by measuring the change in strain gauge resistance. Each sensor synchronously acquires vibration, temperature, rotational speed, and load status data during equipment operation at a constant sampling frequency. The data preprocessing module establishes a communication connection with the data acquisition module and includes a signal conditioning unit and a data purification unit. The signal conditioning unit first amplifies the original acquired signal using an operational amplifier and then performs filtering operations using low-pass and band-pass filters to remove high-frequency noise and interference signals, thus completing the preprocessing of the original acquired signal. In the data purification unit, for vibration signals, the discrete wavelet transform, a multi-resolution signal decomposition technique, is used to decompose the signal into multiple different frequency levels. By setting an appropriate number of decomposition levels and thresholds, noise is removed and effective signals are retained. For temperature data, a state-space model based on Kalman filtering is used, combined with the dynamic change law and noise characteristics of temperature data, to eliminate measurement noise and environmental interference. For speed and load data, the main trend components of the data are extracted through singular value decomposition, and then outliers are removed by combining median filters. Finally, the data is normalized to the [0,1] standard range using a normalization formula. The feature extraction module, built on a deep learning network architecture, is configured with a time-domain feature analysis unit and a frequency-domain feature analysis unit. The time-domain feature analysis unit extracts time-domain features by calculating statistical quantities such as the mean, variance, peak value, kurtosis, and skewness of the signal. The frequency-domain feature analysis unit uses Fast Fourier Transform to convert the time-domain signal to the frequency domain, obtains frequency-domain features such as power spectral density and frequency centroid, and generates time-frequency joint features through Short-Time Fourier Transform. Specifically, this module extracts local time-domain features of the data through multiple one-dimensional convolutional layers, extracts time-frequency distribution features through multiple two-dimensional convolutional layers, uses residual connection structures to solve the gradient vanishing problem in deep networks, enhances feature representation capabilities, and finally achieves feature dimensionality reduction and aggregation through average pooling layers or max pooling layers, ultimately constructing a multi-dimensional feature vector containing 50 feature values. The fault prediction model module employs a time-series deep learning model, including a training unit and a prediction unit. The input layer uses a feature vector mapping device to convert multi-dimensional feature vectors into a time-series format of a specific sequence length. The hidden layer consists of multiple long short-term memory (LSTM) layers or gated recurrent unit (ROU) layers. These units, through the collaborative work of forget gates, input gates, and output gates, capture the long-term dependencies and short-term variation features of the time-series data. The output layer is configured with a fully connected layer containing multiple neurons and a softmax classifier, outputting the probability distribution of fault types and the assessment results of fault severity. The attention mechanism layer generates attention weights by calculating the cosine similarity between the features at each time step and the context vector, highlighting key fault features. The training unit uses labeled data and employs a backpropagation algorithm combined with stochastic gradient descent or its improved version to adjust the model parameters. The prediction unit inputs real-time feature vectors into the trained model for fault prediction. The early warning module, which communicates with the fault prediction model module, includes a threshold comparison unit and an early warning generation unit. The threshold comparison unit sets up a three-level progressive early warning threshold system: the primary early warning corresponds to the early abnormal state of the equipment, and the trigger condition is that the predicted fault probability is between the first fault probability threshold P1 and the second fault probability threshold P2 within 10 consecutive sampling periods; the intermediate early warning corresponds to the potential fault state of the equipment, and the trigger condition is that the predicted fault probability is between the second fault probability threshold P2 and the third fault probability threshold P3 within 5 consecutive sampling periods; the advanced early warning corresponds to the emergency fault state of the equipment, and the trigger condition is that the predicted fault probability exceeds P3, or the fault development rate is determined to exceed the increase of 0.1 per sampling period by calculating the first derivative of the fault probability within consecutive sampling periods. The early warning generation unit generates corresponding level early warning information based on the threshold comparison results, including suggestions for countermeasures such as fault type, predicted occurrence time, enhanced monitoring, arranging shutdown for testing, and immediate shutdown for maintenance.

[0006] Furthermore, the distributed sensor array of the data acquisition module is arranged according to the different functional components of the mechanical equipment, with each sensor corresponding to collect operating status data of a specific part; among them, vibration sensors are arranged on the bearing housing and gearbox surface, temperature sensors are arranged near heat-generating components such as bearings and motors, speed sensors are arranged on the transmission shaft, and load sensors are arranged on the power output end or key force-bearing components.

[0007] Furthermore, the signal conditioning unit of the data preprocessing module includes: The vibration data processing unit is equipped with a discrete wavelet transform device and a signal processing device with a 5-level decomposition structure. By setting appropriate wavelet basis functions and thresholds, it can achieve noise separation and feature enhancement of vibration signals. The temperature data processing unit is equipped with a moving average filter with a window size of 10 and a state-space model device based on Kalman filtering to smooth the temperature data and estimate the state, thereby eliminating measurement noise and environmental interference. The speed and load data processing unit is equipped with a singular value decomposition device with 10 singular value decomposition components and a median filter with a window size of 5, to realize the extraction of speed and load data trends and correction of outliers.

[0008] Furthermore, the deep learning network architecture of the feature extraction module includes: A one-dimensional convolutional layer group consisting of three convolutional layers with different kernel sizes is used to extract local temporal features of the signal with a stride of 1. A two-dimensional convolutional layer group consisting of two convolutional layers with specific kernel sizes is used to extract the time-frequency distribution features of the signal with a stride of 2. A residual connection structure group consisting of three residual blocks, each containing two convolutional layers, is used to solve the gradient vanishing problem in deep networks and enhance feature representation capabilities.

[0009] A pooling layer employing average pooling is used to achieve feature dimensionality reduction and aggregation.

[0010] Furthermore, the time-series deep learning model of the fault prediction model module includes: The input layer is configured with a feature vector mapping device to convert multidimensional feature vectors into a time series format with a sequence length of 100. The hidden layer consists of two layers of long short-term memory units and one layer of gated recurrent units, which are used to capture the long-term dependencies and short-term variation characteristics of time series data. The output layer configuration includes a fully connected layer with 64 neurons and a softmax classifier, outputting the probability distribution of fault types and the assessment results of fault severity. The attention mechanism layer generates attention weights by calculating the cosine similarity between the features at each time step and the context vector, thus highlighting key fault features.

[0011] Furthermore, the three-level progressive early warning threshold system configured in the threshold comparison unit of the early warning module includes: The primary warning corresponds to the early abnormal state of the equipment. The trigger condition is that the predicted failure probability is in the range of [P1,P2] within 10 consecutive sampling periods. Intermediate warning corresponds to the potential fault state of the equipment. The trigger condition is that the predicted fault probability is in the range of [P2,P3] within 5 consecutive sampling periods. Advanced warning corresponds to an emergency equipment failure state. The triggering condition is that the predicted failure probability exceeds P3, or the failure development rate exceeds the growth rate of 0.1 per sampling period. The number of sampling cycles, fault probability ranges, and growth rates corresponding to different warning thresholds are all preset and different for each level.

[0012] Furthermore, the early warning module also includes a threshold dynamic adjustment unit, which is used to acquire equipment operating condition parameters in real time, including equipment running time, ambient temperature, and load change rate; and adjust the thresholds at each level in the three-level progressive early warning threshold system according to the operating condition parameters and a preset operating condition-threshold mapping relationship table; the operating condition-threshold mapping relationship table is obtained by analyzing and statistically analyzing historical operating data and fault data of the equipment under different operating conditions.

[0013] Furthermore, the threshold dynamic adjustment unit is also used for: Statistics on the number of warning triggers and false alarms within a preset time period; If the number of warning triggers exceeds the first set number and the false alarm rate is lower than the first set false alarm rate, then the fault probability range corresponding to each warning threshold level is expanded according to the first adjustment coefficient. If the number of false alarms exceeds the second set number, the fault probability range corresponding to each level of warning threshold will be narrowed according to the second adjustment coefficient; The first adjustment coefficient, the second adjustment coefficient, the first set number of times, the second set number of times, and the first set false alarm rate are all preset.

[0014] Furthermore, the fault prediction system also includes: The model optimization module is configured with an online learning unit and a transfer learning unit. After acquiring new fault data, the online learning unit uses a mini-batch stochastic gradient descent algorithm to incrementally train the fault prediction model. The transfer learning unit uses fault data from other similar devices to fine-tune the last three layers of the pre-trained model, thereby improving the training efficiency and prediction accuracy of the model on the current device. The interpretive analysis module includes a feature importance assessment unit and a fault attribution analysis unit. The feature importance assessment unit evaluates the contribution of a feature to the model's prediction results by calculating the Shapley value of the feature. The fault attribution analysis unit combines important features and historical fault data, and uses methods such as decision tree analysis or association rule mining to analyze the causes of faults. The maintenance decision support module is configured with a maintenance strategy generation unit and a resource scheduling optimization unit. The maintenance strategy generation unit generates the optimal maintenance plan, which includes maintenance time, maintenance method, and required parts, based on the fault type, severity, and development trend, combined with historical maintenance data, and using case reasoning or rule reasoning methods. The resource scheduling optimization unit optimizes the allocation of maintenance personnel and the supply of parts by using genetic algorithms or particle swarm optimization algorithms based on the maintenance plan and existing resources.

[0015] Furthermore, the system is deployed using a distributed architecture, including: Edge computing nodes, deployed at the equipment site, integrate data acquisition and data preprocessing modules. They use ARM-based embedded processors to achieve local data acquisition and preliminary processing, and transmit the processed data to intermediate computing nodes through field communication networks such as ModbusTCP and PROFINET. Intermediate computing nodes are deployed in regional networks, configured with feature extraction modules and some fault prediction model modules. They utilize x86 architecture server clusters to complete feature extraction and edge inference, and transmit the inference results and key data to remote computing centers via fiber optic networks. The remote computing center, deployed in a remote data center, is equipped with a complete fault prediction model module, early warning module and other extended modules. It utilizes a large-scale cloud computing cluster to achieve global model training, advanced analysis and decision support, and feeds back the decision results to intermediate computing nodes and edge computing nodes.

[0016] A fault prediction method for mechanical equipment based on deep learning algorithms includes the following steps: By deploying a distributed sensor array in key parts of mechanical equipment, vibration, temperature, speed and load status data during equipment operation are collected synchronously at a constant sampling frequency. The collected raw data were denoised, outlier corrected and standardized; the vibration signal was decomposed and denoised using discrete wavelet transform; the temperature data was estimated using a state-space model based on Kalman filtering; the rotational speed and load data were trend extracted and outlier corrected by combining singular value decomposition and median filtering; and the data were normalized to the standard range of [0,1].

[0017] Based on a deep learning network architecture, feature extraction is performed on the preprocessed data; the time-domain statistical features of the signal are calculated, the frequency-domain features are obtained using the Fast Fourier Transform, and the time-frequency joint features are generated through the Short Time Fourier Transform; local time-domain features are extracted using three one-dimensional convolutional layers with different kernel sizes, and time-frequency distribution features are extracted using two two-dimensional convolutional layers with specific kernel sizes; the feature representation capability is enhanced through a residual connection structure; and feature dimensionality reduction and aggregation are achieved through an average pooling layer to construct a multi-dimensional feature vector. The time-series deep learning model is trained using multi-dimensional feature vectors. The feature vectors are converted into a time-series format with a sequence length of 100 and input into the model. Time-series features are captured through two layers of Long Short-Term Memory (LSTM) units (128 units each) and one layer of GRU units (64 units each). Attention mechanisms are used to highlight key fault features. The fault type probability distribution and fault severity assessment results are output through fully connected layers and a softmax classifier. Backpropagation and Adam optimization algorithms are used, with cross-entropy loss function as the optimization objective, to iteratively update the model parameters. Dropout is used to prevent overfitting and improve the model's generalization ability. The data collected in real time and preprocessed and feature extracted are input into the trained time series deep learning model. Through time series analysis of the model's hidden layer and classification calculation of the output layer, the model predicts the type, degree and development trend of equipment failure. The prediction results are compared with a preset three-level progressive threshold system. When the predicted fault probability is in the range of [P1, P2] for 10 consecutive sampling periods, a first-level warning is triggered. When the predicted fault probability is in the range of [P2, P3] for 5 consecutive sampling periods, a second-level warning is triggered. When the predicted fault probability exceeds P3 or the fault development rate exceeds the increase of 0.1 per sampling period, a third-level warning is triggered. Based on the threshold comparison results, corresponding levels of early warning information are generated. Level 1 early warning information includes minor anomaly alerts and suggestions for enhanced monitoring; Level 2 early warning information includes potential fault types and suggestions for scheduling shutdown for testing; Level 3 early warning information includes emergency fault warnings and suggestions for immediate shutdown for maintenance. Based on the early warning information, the maintenance strategy generation unit combines the fault type, severity, and development trend with historical maintenance data to determine the maintenance time, maintenance method, and required parts using a case-based reasoning method. The resource scheduling optimization unit uses a genetic algorithm to rationally allocate maintenance personnel and parts based on the maintenance plan and existing resources. Based on the newly collected fault data, the parameters of the time-series deep learning model were adjusted using the mini-batch stochastic gradient descent algorithm through online learning. Transfer learning technology was used to fine-tune the parameters of the last three layers of the pre-trained model with the help of fault data from other similar devices, thereby improving the model's prediction accuracy and generalization ability.

[0018] This invention provides a speed reducer fault prediction system and method based on deep learning algorithms, which has the following beneficial effects: 1. By employing techniques such as discrete wavelet transform and Kalman filtering, the original data is denoised, corrected, and normalized, effectively improving data quality. The pre-processed vibration data removes environmental noise interference, clearly presenting the true operating state of the equipment. Based on this, a deep learning network architecture is used to extract data features from multiple dimensions, including the time domain, frequency domain, and time-frequency domain. Through operations such as convolutional layers, residual connection structures, and pooling layers, a multi-dimensional feature vector containing 50 feature values ​​is constructed. These deeply processed and extracted features can accurately depict subtle changes and potential fault modes in equipment operation, providing high-quality input for the fault prediction model. This enables the model to more sensitively capture fault characteristics, significantly enhancing its ability to identify and predict faults. The model is thus imbued with reconnaissance capabilities, allowing it to perceive potential hazards in equipment operation.

[0019] 2. Based on a temporal deep learning model, the system performs in-depth analysis of the input multidimensional feature vectors through LSTM, GRU, and attention mechanisms, enabling accurate prediction of equipment fault types, severity, and development trends. For example, when the model predicts a potential fault, it immediately compares the prediction result with a three-level progressive early warning threshold system. If the corresponding early warning conditions are triggered, it quickly generates early warning information containing fault details and countermeasures. This accurate prediction and intelligent early warning mechanism transforms equipment maintenance from a passive response to proactive prevention, allowing for advance planning of maintenance schedules, avoiding fault occurrences, significantly reducing equipment downtime and maintenance costs, and substantially improving production efficiency and equipment reliability.

[0020] 3. Based on case-based reasoning, and combining historical maintenance data with current fault information, the system quickly formulates the optimal maintenance plan, including maintenance time, methods, and required parts. Simultaneously, based on the maintenance plan and existing resources, a genetic algorithm is used to rationally allocate maintenance personnel and parts resources. This enables the generation of detailed maintenance plans and the allocation of maintenance teams and parts in a short time, ensuring equipment is repaired in the shortest possible time and effectively reducing the impact of equipment downtime on production. This collaborative work model achieves seamless integration of fault warning and maintenance decision-making, greatly improving the efficiency and scientific nature of equipment maintenance and reducing operation and maintenance costs.

[0021] 4. By utilizing online learning and transfer learning technologies, the fault prediction model parameters are continuously adjusted based on newly collected fault data to improve the model's prediction accuracy and generalization ability. Using acquired equipment operating condition parameters, the warning threshold is adjusted according to the operating condition-threshold mapping table, and further optimized based on the number of warning triggers and false alarms. Timely adjustment of the warning threshold ensures the accuracy of warnings. Simultaneously, new fault data is used to optimize the model, enabling the system to continuously adapt to fault prediction needs under different operating conditions. This continuous optimization and adaptive adjustment mechanism ensures that the system maintains efficient and accurate fault prediction and warning capabilities throughout long-term operation, continuously improving the overall performance and reliability of the system. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1: Please refer to Figure 1 This embodiment provides a gearbox fault prediction system based on deep learning algorithms. The specific contents of the fault prediction system are as follows: With the accumulation of industrial big data and the development of artificial intelligence technology, data-driven fault prediction methods have gradually emerged. Among them, machine learning algorithms such as support vector machines (SVM) and random forests (RF) have been widely used in mechanical equipment fault prediction. These methods establish fault prediction models by learning from historical data, but they have problems such as difficulty in feature extraction and insufficient model generalization ability when dealing with high-dimensional and nonlinear data.

[0025] Deep learning, as an important branch of artificial intelligence, has achieved remarkable results in areas such as image recognition and natural language processing, and is increasingly being applied to mechanical equipment fault prediction. Deep learning can automatically extract features from large amounts of data without requiring manual design of feature extraction algorithms, giving it a unique advantage in handling complex data. For example, Convolutional Neural Networks (CNNs) can effectively extract local features from data, while Recurrent Neural Networks (RNNs) and their variants, Long Short-Term Memory Networks (LSTM) and Gated Recurrent Units (GRUs), can process time-series data and capture long-term dependencies. However, existing deep learning applications for mechanical equipment fault prediction still suffer from problems such as low model training efficiency and imperfect early warning mechanisms. In practical applications, the lack of unified standards for model parameter settings makes it difficult to reproduce different research results. Early warning thresholds are often fixed, failing to adapt to dynamic changes in equipment operating conditions, thus affecting the accuracy and reliability of fault prediction. Specific improvement schemes are as follows: (a) Data acquisition and transmission: 1. Distributed sensor array deployment: By deploying distributed sensor arrays consisting of piezoelectric vibration sensors, infrared temperature sensors, magnetoelectric speed sensors, and strain gauge load sensors at key structural locations in mechanical equipment, such as bearing housings, gearboxes, and drive shafts, multi-dimensional data acquisition is achieved. Taking a reducer used in steel rolling production as an example, piezoelectric vibration sensors are installed on the bearing housings of the rolling mill, and bearing vibration data is collected at a constant sampling frequency of 1000Hz; infrared temperature sensors are installed on the surface of the gearbox to detect the surface temperature of the gearbox; magnetoelectric speed sensors are installed at the drive shaft to convert the drive shaft speed into a pulse signal using the principle of electromagnetic induction; and strain gauge load sensors are used to measure the load conditions during the operation of the rolling mill, thereby obtaining multi-dimensional information about the equipment operation.

[0026] 2. Data transmission architecture: The system adopts a distributed architecture to realize data transmission. Edge computing nodes are deployed on-site, using ARM-based embedded processors, such as Raspberry Pi 4B, to complete local data acquisition and preliminary processing, and transmit the data to the intermediate computing nodes through industrial standard communication protocols. In actual use, Raspberry Pi 4B collects sensor data in real time, performs preliminary noise reduction on the vibration signals, and transmits the data to the intermediate computing nodes in the workshop through the Modbus TCP protocol.

[0027] The intermediate computing nodes utilize x86 architecture server clusters to perform feature extraction and partial fault prediction model calculations, and then transmit critical data to the remote computing center.

[0028] Remote computing centers rely on large-scale cloud computing clusters, such as those built on OpenStack, to achieve global model training, advanced analysis and decision support, and feed the decision results back to intermediate computing nodes and edge computing nodes.

[0029] (ii) Data preprocessing: Data preprocessing mainly consists of two parts: firstly, signal conditioning, which involves denoising, filtering, and other data processing; and secondly, purifying the pre-processed data according to a preset data processing method and normalizing the data to the standard range [0, 1]. Specifically: Signal conditioning unit: The original acquired signal is amplified by an operational amplifier to increase the signal amplitude. At the same time, low-pass filters and band-pass filters are used to filter high-frequency noise and interference in different types of signals. For example, when processing the vibration signal of the rolling mill in the steel rolling production line, due to the large amount of electromagnetic interference on site, the high-frequency electromagnetic interference is effectively filtered out by setting a low-pass filter with the corresponding cutoff frequency, making the vibration signal clearer.

[0030] Data cleansing unit: For vibration signals, a 5-level decomposition is performed using Discrete Wavelet Transform (DWT). Based on the equipment's operating characteristics and noise intensity, suitable wavelet basis functions such as db4 are selected. A soft thresholding method is used to set the threshold, achieving noise removal and effective signal preservation. Assuming the original vibration signal sequence is x(n), after j-level DWT decomposition, the approximate coefficient a is obtained. j (n) and detail coefficients d j (n), where j = 1, 2, ..., 5. The soft thresholding formula is: ; In the formula, Let d be the value of the detail coefficient of the j-th layer at position n after soft thresholding. j (n) represents the value of the original detail coefficients of the j-th layer at position n, where j is the decomposition layer number and n is the position in the signal sequence, sgn(d j (n)) is d j The sign function of (n), λ j is the threshold value for the j-th layer.

[0031] In vibration signal processing, detail coefficients d at different scales are obtained through DWT decomposition. j (n), the above formula is used to perform soft thresholding on the detail coefficients after discrete wavelet transform decomposition, with the aim of removing noise from the vibration signal. For each detail coefficient d j (n), first determine its absolute value and threshold λ. j The size relationship.

[0032] If |d j (n)|≥λ j Then through sgn(d j (n))×(|d j (n)|-λ j ) is processed, sgn(d j (n) is the sign function, used to preserve the sign of the original coefficients and subtract the threshold λ. j This achieves the shrinkage of the coefficients, thereby suppressing noise; If |d j (n)|<λ j If the coefficient is zero, then the coefficient is set to 0, and it is assumed that the coefficient is mainly caused by noise.

[0033] The threshold λ in the discrete wavelet transform soft thresholding process is calculated dynamically. j In actual use, the noise characteristics and signal strength of the reducer vibration signal vary under different operating conditions. The threshold is determined based on the statistical characteristics of the signal itself, and the specific formula is as follows: ; In the formula, λ j The threshold for the j-th layer, The basic threshold component is calculated based on the Gaussian distribution characteristics of the noise, where, N represents the variance of the detail coefficients at the j-th layer, reflecting the degree of signal fluctuation. j Let d be the length of the detail coefficients at the j-th layer. j (n) represents the value of the original detail coefficient of the j-th layer at position n; It is an adjustment factor based on the local characteristics of the signal. By calculating the relationship between the square and variance of the detail coefficients, the threshold is further optimized to better adapt to signal changes.

[0034] Temperature data is analyzed using a Kalman-filter-based state-space model. Based on the dynamic temperature change patterns and noise characteristics, state and observation equations are established. By continuously updating the state estimate and covariance matrix, measurement noise and environmental interference are eliminated. Let the state equation for temperature (i.e., the equation describing the temperature state of the rolling mill bearings in the steel rolling production line as a function of time) be: x k =Φ k,k-1 ×x k-1 +Γ k,k-1 ×w k-1 ; In the formula, x k Φ is the state vector at time k, containing state variables such as temperature and its first derivative, used to comprehensively describe the system's state at that time; k,k-1This is the state transition matrix, whose elements are parameterized based on the equipment's heat conduction model and historical operating data. It reflects the system state transition relationship from time k-1 to time k, taking into account the influence of the equipment's thermal inertia and ambient temperature changes. x k-1 Γ is the state vector at time k-1; k,k-1 The process noise driving matrix is ​​defined as follows: process noise w k-1 Introducing the system, w k-1 For the process noise at time k-1, w k-1 Follows a Gaussian distribution N(0,Q) k-1 ), which represents the uncertain interference that the system experiences during operation.

[0035] The relationship between the internal temperature state of the rolling mill bearing and the sensor measurements was established, and the observation equation is as follows: z k =H k ×x k +vk; In the formula, z k H is the observation vector at time k, i.e., the temperature value measured by the sensor; k The observation matrix is ​​used to represent the system's state vector x. k Mapped to the observation space, this reflects how the sensor observes the system state; v k To observe the noise, we assume it follows a Gaussian distribution N(0,R). k ), which represents the error present in the sensor measurement process.

[0036] The Kalman filter update process includes two steps: prediction and update. The prediction formula is: ; In the formula, To estimate the predicted state at time k based on information from time k-1, Φ k,k-1 Let k be the state transition matrix from time k-1 to time k. For the optimal state estimate at time k-1, P k|k-1 P is the covariance of the predicted state estimate at time k based on information from time k-1. k-1|k-1 Let Γ be the covariance of the optimal state estimate at time k-1. k,k-1 Q is the process noise driving matrix. k-1 Let be the covariance matrix of the process noise at time k-1.

[0037] In the prediction step of the Kalman filter update process, the first formula is used to estimate the optimal state at time k-1. Through the state transition matrix Φ k,k-1 Obtain the predicted state estimate at time k The second formula is used to calculate the uncertainty (covariance) P of the predicted state estimate. k|k-1 In the formula, For propagation based on the covariance and state transition of the previous time step, The influence of process noise on covariance is taken into account; in rolling mill temperature monitoring, these two formulas can be used to preliminarily estimate the current temperature state and the reliability of the estimate.

[0038] The updated formula is: ; In the formula, K k P is the Kalman gain at time k. k|k-1 H is the covariance of the predicted state estimate at time k based on information at time k-1. k Let R be the observation matrix at time k. k Let be the covariance matrix of the observation noise at time k. This is the optimal state estimate based on all information at time k. To estimate the predicted state at time k based on information from time k-1, z k Let P be the observation vector at time k. k|k Let I be the optimal state estimation covariance based on all information at time k, where I is the identity matrix.

[0039] In the Kalman filter update step, the first formula is used to calculate the Kalman gain K. k It determines how to reasonably integrate the predicted state estimate and new observation data; the second formula uses the Kalman gain K... k Estimation of predicted state By making corrections and combining them with the observed data zk, a more accurate state estimate can be obtained. The third formula is used to update the covariance Pk|k to reflect the uncertainty of the new estimate. In the processing of rolling mill temperature data, these formulas effectively eliminate random noise and environmental interference in the measurement, and improve the stability and accuracy of temperature data.

[0040] The speed and load data are first processed by singular value decomposition (SVD) to extract the main trend components, and then combined with a median filter to remove outliers; let the speed or load data matrix be X∈R. m×n Perform SVD decomposition: X = UΣV T This formula is used for data matrices X∈R m×n Perform singular value decomposition; the decomposition yields an orthogonal matrix U, a diagonal matrix Σ, and the transpose V of the orthogonal matrix V. TX is the original speed or load data matrix; by retaining the components corresponding to larger singular values, the main trends and features of speed and load data can be extracted, and noise caused by equipment vibration, electromagnetic interference, etc. can be removed; for example, when rolling steel of different specifications in a rolling mill, the changes in equipment operating status lead to complex data, and SVD can effectively simplify the data structure and highlight key information.

[0041] Considering the periodic and trend-like changes in equipment operation, a weighted singular value reconstruction method is adopted, retaining the first s singular values ​​to reconstruct the data matrix. : In the formula, Let be the reconstructed data matrix, s be the number of singular values ​​retained, and ω be the number of singular values ​​retained. i σ is the weighting coefficient corresponding to the i-th singular value. i For the i-th singular value, u i Let be the i-th column vector of matrix U. Let V be the transpose of the vector in the i-th row of matrix V.

[0042] The weighting coefficient ω is determined based on the mill's operating conditions and historical data. i This highlights outlier components that are more relevant to the actual operating status of the equipment. For example, when rolling steel of different specifications in a rolling mill, the weighting coefficients are adjusted to further enhance the expression of useful information in the rotation speed data. Finally, an adaptive window-size median filter is used to remove outliers. The window size is dynamically adjusted according to statistical quantities such as the standard deviation of the data. The data is then normalized to the range of [0,1] to eliminate the dimensional differences between different data features.

[0043] (III) Feature Extraction: Basic feature calculation: The time-domain feature analysis unit of the feature extraction module obtains the basic characteristics of the signal in the time dimension by calculating statistical measures such as signal mean, variance, peak value, kurtosis, and skewness. Due to the non-stationarity and trend of the data, a weighted method is used when calculating these statistical measures: Where N is the length of the data sequence, x n The preprocessed data sequence has undergone steps such as discrete wavelet transform and soft thresholding of vibration signals, and Kalman filtering of temperature data; w n The weighting coefficients are determined based on data acquisition time and equipment operating conditions. For example, during the mill start-up phase, the vibration signal fluctuates greatly, corresponding to the weighting coefficients for data points. nLarger values ​​will be assigned to the data. The calculation of variance, peak value, kurtosis, and skewness also relies on the preprocessed data and considers weighting coefficients to more accurately reflect the signal characteristics. Taking the mean calculation as an example, when calculating the weighted mean of the mill vibration signal in the steel rolling production line, a corresponding weighting coefficient will be assigned to each data point according to the data acquisition time and equipment operating conditions. During the mill start-up phase, due to the large changes in the vibration signal, the data in this phase will be given a larger weight to reflect its importance. Then, each data point is multiplied by its weighting coefficient and summed, and then divided by the sum of the weighting coefficients to obtain the weighted mean. The calculation of variance, peak value, kurtosis, and skewness also considers weighting coefficients in a similar way.

[0044] The frequency domain feature analysis unit uses Fast Fourier Transform (FFT) to convert the time-domain signal to the frequency domain and extract frequency domain features such as power spectral density and frequency centroid. When calculating the power spectral density P(f) of the rolling mill vibration signal, considering the multi-condition characteristics of equipment operation, a multi-window spectrum estimation method is adopted, assuming the use of L different window functions w. l Given (t), l=1,2,…,L, the formula for calculating the power spectral density is: ; Where P(f) is the power spectral density of the rolling mill vibration signal at frequency f, L is the number of window functions used, x(t) is the preprocessed vibration signal, and w l (t) is the l-th window function. For the window function w l Energy of (t); frequency centroid f cg The calculation considers the influence weight α(f) of different frequency components on equipment failure, where f is the frequency variable, and the formula is: ; In the formula, f cg为 The frequency centroid reflects the concentration trend of signal frequency components and is an important indicator for measuring signal frequency distribution. In equipment fault diagnosis, changes in the frequency centroid can be used as one of the bases for judging the equipment status. f is a frequency variable that traverses all frequency components of the signal. P(f) is the power spectral density calculated earlier, which reflects the energy magnitude of different frequency components. α(f) is the influence weight of different frequency components on equipment faults, which is determined according to the structural characteristics and fault mechanism of the equipment. For example, some frequency components are closely related to the faults of key components of the equipment, and their corresponding α(f) will be set to be larger.

[0045] This method generates joint time-frequency features using Short-Time Fourier Transform (STFT), and the calculation results of these basic features will serve as input for subsequent feature extraction by deep learning networks. The method processes the signal using multiple different window functions, calculates the power spectral density of the signal after processing by each window function, and then takes the average value as the final power spectral density result, thus more accurately reflecting the frequency component distribution of the vibration signal under different operating conditions. The frequency centroid calculation considers the influence weight of different frequency components on equipment failure, multiplying each frequency value by the corresponding power spectral density and frequency weight coefficient, summing the results, and then dividing by the sum of the products of the power spectral density and the frequency weight coefficient. Furthermore, joint time-frequency features are generated using STFT to more comprehensively analyze signal characteristics.

[0046] Feature extraction using deep learning networks: Based on a deep learning network architecture, multiple one-dimensional convolutional layers are used to extract local temporal features of the data, and multiple two-dimensional convolutional layers are used to extract time-frequency distribution features; the input of the one-dimensional convolutional layer is... Where N is the time step, C in The number of input channels is and the convolution kernel is . K is the kernel size, C out The number of output channels, with an offset of... The output of a one-dimensional convolutional layer The calculation formula is: ; In the formula, Y i,c To represent the value of the output feature map of a one-dimensional convolutional layer at the i-th time step and the c-th channel, it is the result of the convolution calculation and includes the extracted local temporal features; i is the time step index, ranging from 0 to N-K+1, where N is the time step size of the input data and K is the convolution kernel size; c is the output channel index, ranging from 0 to C. out -1, C out is the number of output channels; k is the position index within the convolution kernel, used to traverse each element of the convolution kernel; c in The input channel index ranges from 0 to C. in -1, C in Input the number of channels; For the input feature map at time step i+k, the c-th time... in The values ​​of each channel are the input data for the convolution calculation, which come from the signal features after the basic feature calculation. For the convolution kernel at position k and position c in The weight parameters of the c-th input channel and the c-th output channel are key parameters for model learning, determining the feature extraction capability of the convolutional layer; b c This is the bias parameter for the c-th output channel, used to adjust the output of the convolutional layer.

[0047] During the calculation process, considering the dynamic changes in equipment operation data, a dynamic convolution kernel update mechanism is introduced; during each training iteration, the gradient of the convolution kernel W is calculated based on the loss function L output by the model. The update amount ΔW of the convolution kernel is calculated using the gradient descent algorithm as follows: , where η is the learning rate; the updated convolutional kernel W′ is: W′=W+ΔW.

[0048] This process relies on feedback from model training, and the input to model training is data that has undergone data collection, preprocessing, and basic feature calculation, forming a tight data correlation chain.

[0049] Multiple two-dimensional convolutional layers extract time-frequency distribution features, with the time-frequency map generated by STFT as input; let the input feature map of the two-dimensional convolutional layer be... H and W are the height and width of the feature map, respectively, and the convolution kernel is... K H K W The kernel size is defined by its height and width, with an offset of 1 / 2. The two-dimensional convolutional layer outputs a feature map. The calculation formula is: ; In the formula, F out (i,j,c) represents the value of the output feature map of the 2D convolutional layer in the i-th row, j-th column, and c-th channel, containing the extracted time-frequency distribution features; i is the row index of the feature map, ranging from 0 to HK. H +1, where H is the height of the input feature map, and K... H is the kernel height; j is the column index of the feature map, ranging from 0 to WK. W +1, W is the width of the input feature map, K W is the kernel width; c is the output channel index, ranging from 0 to C. out -1, C out The output channel number is h; the kernel index is h in the height direction; the kernel index is w in the width direction; and c is c. in The input channel index ranges from 0 to C. in -1, C in F is the number of input channels; in (i+h,j+w,c in ) represents the input feature map at row i+h, column j+w, and column c. in The values ​​of each channel are the input data for the two-dimensional convolution calculation, derived from the time-frequency plot generated by STFT; W conv (h,w,c inc) represents the two-dimensional convolution kernel at row h, column w, and column c. in The weight parameters for the c-th input channel and the c-th output channel; b conv (c) represents the bias parameter of the c-th output channel.

[0050] To address local feature variations in time-frequency images, an adaptive convolution stride strategy is employed. First, the gradient G of the time-frequency image at each location is calculated, and then the absolute values ​​of the gradients at all locations are summed. G(i,j) is the gradient of the time-frequency image at the position of the i-th row and j-th column, divided by the total number of elements in the feature map H×W×C. in After multiplying by the adjustment factor γ and adding it to the initial stride s0, the dynamically adjusted convolution stride s is obtained as follows: The residual connection structure is used to solve the gradient vanishing problem in deep networks and enhance the feature representation ability. The average pooling layer is used to achieve feature dimensionality reduction and aggregation, and finally a multi-dimensional feature vector containing 50 feature values ​​is constructed. These feature vectors will be used as input data for the fault prediction model, and their calculation depends on the data processing and calculation results of the previous steps. The formulas are closely related.

[0051] Considering the dynamic changes in equipment operation data, a dynamic convolutional kernel update mechanism is introduced. Specifically, during each training iteration, the update amount of the convolutional kernel is calculated using the gradient descent algorithm based on the real-time data of equipment operation and the feedback information from model training. This update amount is then multiplied by the learning rate and added to the original convolutional kernel to obtain the updated convolutional kernel. Taking the motor current data processing of a steel rolling production line as an example, by dynamically updating the convolutional kernel, the model can better adapt to changes in current under different loads and extract more effective local time-domain features.

[0052] In the computation of two-dimensional convolutional layers, an adaptive convolution stride strategy is adopted to address local feature changes in time-frequency images. First, the gradient of the time-frequency image at each location is calculated. The absolute values ​​of the gradients at all locations are summed, divided by the total number of elements in the feature map, multiplied by an adjustment coefficient, and then added to the initial stride to obtain the dynamically adjusted convolution stride. This strategy can dynamically adjust the convolution stride according to changes in the features of the time-frequency image, resulting in more accurate feature extraction. For example, when analyzing the time-frequency image of a rolling mill gearbox vibration signal, when local wear occurs in the gear, the time-frequency characteristics of the vibration signal will change. The adaptive convolution stride strategy can dynamically adjust the stride according to this change, resulting in more accurate extraction of abnormal features.

[0053] The vanishing gradient problem in deep networks is solved by using residual connection structures (composed of multiple residual blocks containing two convolutional layers), which enhances the feature representation capability. Feature dimensionality reduction and aggregation are achieved through average pooling layers, and finally a multi-dimensional feature vector containing 50 feature values ​​is constructed. In actual training, the feature extraction effect is optimized by adjusting parameters such as the number of convolutional layers, the size of the convolutional kernel, the stride, and the number of residual blocks.

[0054] (iv) Fault prediction model construction: Model architecture design: A hybrid architecture integrating LSTM, attention mechanisms, and CNNs is employed. LSTM serves as the primary framework for processing multi-dimensional time-series data, including data on equipment vibration, temperature, and rotational speed. The core computational unit of LSTM processes the time-series data through a series of formulas. ; In the formula, f t The output of the forget gate at time t, with a value between 0 and 1, is used to control the cell state C at the previous time step. t-1 The degree to which information is retained; i t The output of the input gate at time t, with a value between 0 and 1, determines the current input x. t How much information will be added to the cell state? The candidate cell state is determined based on the current input x. t and the hidden state h from the previous moment t-1 The calculated temporary cell state; C t The cell state updated at time t combines retained information from the previous time step with the new information from the current input; t The output of the output gate at time t, with a value between 0 and 1, controls the cell state C. t How much information will be output to the hidden state h? t h t The hidden state at time t contains feature information about the device's operational data at the current time, and is passed as the output of the LSTM to the next time step or subsequent network layers; x t The input at time t is a multi-dimensional feature vector output by the feature extraction module, containing multi-dimensional information such as equipment vibration, temperature, and rotational speed; h t-1 C represents the hidden state at the previous time step (t-1); t-1 The cell state at the previous time step (t-1); W f W i W C W o These are the weight matrices for the forget gate, input gate, cell state update, and output gate, respectively; b f b ib C b o These are the bias vectors corresponding to the forget gate, input gate, cell state update, and output gate, respectively; σ is the Sigmoid activation function, which maps the input to between 0 and 1 and is used to control the degree of opening and closing of the gates; tanh is the hyperbolic tangent activation function, which maps the input to between -1 and 1 and is used to generate candidate cell states and hidden states.

[0055] At each time step, LSTM uses these formulas to selectively forget past information, update cell states, and output hidden states, thereby capturing the changing trends of bearing vibration data over long time series and identifying weak signals of faults such as early wear.

[0056] An attention mechanism module is embedded between LSTM layers. The output ht of the LSTM layer is mapped to a new dimension through a fully connected layer, and the correlation score between the hidden states at different time steps is calculated. , where e ij Hide state h at different time steps i and h j The correlation score between the features at two time steps indicates a higher correlation. i h j W represents the hidden state output of the LSTM layer at time steps i and j. a It is a learnable weight matrix; for example, when analyzing vibration data of a rolling mill gearbox, this formula can be used to calculate the correlation of vibration characteristics at different time points; then, the softmax function is used to convert the scores into attention weights. In the formula, α ij The attention weights reflect the importance of the hidden state at time step j to time step i, and their values ​​range from 0 to 1. The sum of the attention weights for all time steps is 1; T is the total number of time steps; c i The feature vector, after weighted summation using an attention mechanism, highlights information important for fault prediction, allowing the model to focus on key time steps of abnormal gearbox vibration changes, such as the moment of sudden vibration change when gear wear intensifies; finally, weighted summation is performed... Highlight information that is important for fault prediction.

[0057] Connecting multiple CNN layers further extracts local features from the processed features; the convolution operation formula for a two-dimensional convolutional layer is... In the formula, U and V are the number of rows and columns of the convolution kernel, respectively, and F... m+1 (i,j) represents the value of the output feature map of the (m+1)th convolutional layer in the i-th row and j-th column; F m (i+u,j+v) represents the value of the input feature map of layer m in row i+u and column j+v; W m(u,v) represents the weight of the m-th layer convolutional kernel in the u-th row and v-th column; b m As the bias of the m-th layer, when processing the time-frequency feature map of the rolling mill motor current signal, the convolution kernel slides on the time-frequency map. The output feature map is calculated using this formula to extract features such as local frequency changes in the current signal. The output layer is a fully connected layer, with 5 neurons set according to 5 possible fault types in the rolling mill, such as bearing wear, gear breakage, and motor overheating, to output the fault prediction results. , The output of the fault prediction model is a vector, where each element corresponds to the predicted probability of a fault type; W fc This is the weight matrix of the fully connected layer, used to weight the input feature vector h. out Mapped to fault type space; h out It is a fully connected layer.

[0058] Model training and optimization: The model is trained using the Adam optimizer, whose update formula is as follows: ; In the formula, m t β1 is the first-moment estimate at time t, recording the average information of historical gradients; β1 is the decay factor of the first-moment estimate, usually set to a value close to 1 (e.g., 0.9), used to control the degree of retention of historical gradient information; the closer the value is to 1, the more historical gradient information is retained; m t-1 The first moment estimate at time t-1; g t The gradient of the model parameters calculated at time t is obtained by differentiating the loss function with respect to the model parameters, indicating the direction and magnitude of the model parameter updates; v t β2 is the second-order moment estimate at time t, reflecting the change in the magnitude of the gradient; β2 is the decay factor of the second-order moment estimate, typically set to 0.999, used to control the decay rate of historical information in the second-order moment estimate; v t-1 For the second-order moment estimate at time t-1, β1 and β2 are usually set to 0.9 and 0.999, respectively, to control the decay rate of the moment estimate.

[0059] and It is a corrected moment estimate, avoiding bias in the early stages of training. In the early stages of training, due to m t and v t The initial value is 0, and their estimates will be biased towards 0; these two formulas are derived by applying m... t and v t Corrections are made to eliminate biases from the initial training phase, ensuring that the first and second moment estimates more accurately reflect the true gradient information; as the number of training iterations t increases, and As the value gradually approaches zero, the correction effect gradually weakens, ensuring the stability of the estimated value in the later stages of training. This is the corrected estimate of the first moment at time t; To estimate the corrected second moment at time t, the learning rate η is initially set to 0.001, and ϵ is set to 10. -8 To prevent the denominator from being zero, these formulas are used to continuously update the model parameters θ. t This allows the model to gradually adapt to the characteristics of rolling mill data.

[0060] The loss function chosen is the cross-entropy loss function. Assuming there are 100 samples in a training batch (N=100), and the rolling mill has 5 fault types (C=5), y nc This indicates whether sample n belongs to category c (0 or 1). It is the probability that the model predicts that sample n belongs to category c; by calculating this loss function, the difference between the model's prediction and the true label is measured, driving model optimization.

[0061] The dataset was divided into training, validation, and test sets in a 7:1:2 ratio. During training, the learning rate was multiplied by 0.9 and decayed every 10 rounds, with an L2 regularization term introduced. (λ=0.0001) To prevent overfitting, L is the original loss function, λ is the regularization strength hyperparameter, used to control the influence of the L2 regularization term on the loss function; the larger the value of λ, the stronger the constraint of regularization, and the model parameters will be compressed closer to 0. N is the number of training samples, that is, the total number of data samples participating in the model training. W is the set of all trainable parameters (weights) in the model, and w is a single weight parameter in the set W. Training stops when the validation set loss value no longer decreases for 5 consecutive rounds. At this time, the model achieves good performance in predicting the failure of rolling mill equipment.

[0062] (v) Early warning mechanism design; A three-tiered progressive early warning threshold system; Taking a certain characteristic of the vibration signal of the rolling mill in a steel rolling production line (such as the root mean square value of vibration acceleration) as an example, statistical analysis of the historical data of normal equipment operation yielded a mean value μ = 2.5 m / s². 2 The standard deviation σ = 0.3 m / s 2 The first-level warning threshold is set as μ + 1.5σ = 2.5 + 1.5 × 0.3 = 2.95 m / s 2 When the vibration characteristic value exceeds this threshold, such as when the root mean square value of vibration acceleration reaches 3.0 m / s² in a certain monitoring, 2 The system issued a Level 1 warning, prompting maintenance personnel to pay attention to the rolling mill status.

[0063] The secondary warning threshold is set as μ + 2.5σ = 2.5 + 2.5 × 0.3 = 3.25 m / s2 If the vibration characteristic value exceeds this threshold, it indicates that the anomaly is intensifying, and the system issues a level-two warning; the level-three warning threshold is set as μ + 3.5σ = 2.5 + 3.5 × 0.3 = 3.55 m / s². 2 When the threshold is reached, a Level 3 warning will be issued immediately and a shutdown for maintenance will be recommended.

[0064] The isolated forest algorithm was used to assist in adjusting the threshold; the isolated forest model was trained on a large amount of historical operating data of the rolling mill equipment, and the anomaly score of each data point was calculated. When analyzing rolling mill bearing temperature data, if the isolation path length h(x) of the temperature data at a certain moment is short, the calculated anomaly score S(x) is close to 1, indicating that the data point is very likely to be an anomaly. This can be used to optimize the warning thresholds at all levels and improve the accuracy of the warning.

[0065] Early warning information push and processing When an anomaly occurs in the rolling mill and triggers an early warning, the system pushes information via SMS, email, and pop-up windows in the enterprise's internal management system. For example, if a Level 1 early warning is triggered, the SMS message will read: "Rolling mill bearing, minor anomaly warning, current root mean square value of vibration acceleration is 3.0 m / s²." 2 Exceeding the first-level threshold of 2.95 m / s 2 Time: [Specific Time]”; After receiving the message, maintenance personnel can view the historical vibration data curves, real-time monitoring values, and related characteristic analysis charts of the rolling mill through the system interface; For Level 1 warnings, a regular inspection is arranged within 3 days. If insufficient bearing lubricating grease is found during the inspection, the grease is replenished in time; For Level 2 warnings, targeted testing is carried out, such as using an ultrasonic flaw detector to check the gearbox gears. If local wear is found, the machine is shut down for maintenance; For Level 3 warnings, the machine is shut down immediately for maintenance, forming a closed-loop management system and continuously optimizing the warning effect.

[0066] Example 2: Based on Example 1, this example also provides a method for predicting mechanical equipment faults based on deep learning algorithms, including the following specific steps: By deploying a distributed sensor array in key parts of mechanical equipment, vibration, temperature, speed and load status data during equipment operation are collected synchronously at a constant sampling frequency. The collected raw data were denoised, outlier corrected and standardized; the vibration signal was decomposed and denoised using discrete wavelet transform; the temperature data was estimated using a state-space model based on Kalman filtering; the rotational speed and load data were trend extracted and outlier corrected by combining singular value decomposition and median filtering; and the data were normalized to the standard range of [0,1].

[0067] Based on a deep learning network architecture, feature extraction is performed on the preprocessed data; the time-domain statistical features of the signal are calculated, the frequency-domain features are obtained using the Fast Fourier Transform, and the time-frequency joint features are generated through the Short Time Fourier Transform; local time-domain features are extracted using three one-dimensional convolutional layers with different kernel sizes, and time-frequency distribution features are extracted using two two-dimensional convolutional layers with specific kernel sizes; the feature representation capability is enhanced through a residual connection structure; and feature dimensionality reduction and aggregation are achieved through an average pooling layer to construct a multi-dimensional feature vector. The time-series deep learning model is trained using multi-dimensional feature vectors. The feature vectors are converted into a time-series format with a sequence length of 100 and input into the model. Time-series features are captured through two layers of Long Short-Term Memory (LSTM) units (128 units each) and one layer of GRU units (64 units each). Attention mechanisms are used to highlight key fault features. The fault type probability distribution and fault severity assessment results are output through fully connected layers and a softmax classifier. Backpropagation and Adam optimization algorithms are used, with cross-entropy loss function as the optimization objective, to iteratively update the model parameters. Dropout is used to prevent overfitting and improve the model's generalization ability. The data collected in real time and preprocessed and feature extracted are input into the trained time series deep learning model. Through time series analysis of the model's hidden layer and classification calculation of the output layer, the model predicts the type, degree and development trend of equipment failure. The prediction results are compared with a preset three-level progressive threshold system. When the predicted fault probability is in the range of [P1, P2] for 10 consecutive sampling periods, a first-level warning is triggered. When the predicted fault probability is in the range of [P2, P3] for 5 consecutive sampling periods, a second-level warning is triggered. When the predicted fault probability exceeds P3 or the fault development rate exceeds the increase of 0.1 per sampling period, a third-level warning is triggered. Based on the threshold comparison results, corresponding levels of early warning information are generated. Level 1 early warning information includes minor anomaly alerts and suggestions for enhanced monitoring; Level 2 early warning information includes potential fault types and suggestions for scheduling shutdown for testing; Level 3 early warning information includes emergency fault warnings and suggestions for immediate shutdown for maintenance. Based on the early warning information, the maintenance strategy generation unit combines the fault type, severity, and development trend with historical maintenance data to determine the maintenance time, maintenance method, and required parts using a case-based reasoning method. The resource scheduling optimization unit uses a genetic algorithm to rationally allocate maintenance personnel and parts based on the maintenance plan and existing resources. Based on the newly collected fault data, the parameters of the time-series deep learning model were adjusted using the mini-batch stochastic gradient descent algorithm through online learning. Transfer learning technology was used to fine-tune the parameters of the last three layers of the pre-trained model with the help of fault data from other similar devices, thereby improving the model's prediction accuracy and generalization ability.

[0068] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0069] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0070] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A speed reducer failure prediction system based on a deep learning algorithm, characterized by, The fault prediction system comprises: A data acquisition module is arranged on the speed reducer and is used to collect running state data of the speed reducer, wherein the running state data includes vibration signals, temperature data, rotation speed data and torque data during running; A data processing module has a signal conditioning unit and a data purification unit inside, the signal conditioning unit performs preliminary processing on the collected running state data through signal amplification and filtering operation; the data purification unit purifies the preliminarily processed data according to a preset data processing method and normalizes the data to a standard range; A feature extraction module extracts time domain features and frequency domain features from the processed data based on a deep learning network architecture, extracts time-frequency joint features through a data transformation method, and constructs a multi-dimensional feature vector after feature enhancement and dimension reduction; A fault prediction model module adopts a time series deep learning model to analyze the extracted multi-dimensional feature vector, predict the fault type probability, degree evaluation result and development trend of the speed reducer; An early warning module is used to compare the predicted fault result with a threshold value, divide the fault level according to the comparison result, and generate corresponding early warning information.

2. The deep learning algorithm based gearbox failure prediction system of claim 1, wherein, The data purification unit purifies the preliminarily processed data through a preset data processing method, and the specific content is as follows: A wavelet transform filter is used to perform multi-resolution analysis on the vibration signals, decompose different frequency components to remove noise; A Kalman filter is used to estimate the state of the temperature data and eliminate environmental interference factors; A singular spectrum analysis processor and a median filter are used to extract trends and correct abnormal values of the rotation speed and torque data.

3. The deep learning algorithm based reduction gearbox failure prediction system of claim 1, wherein, Time domain features are extracted by calculating the statistics of the signals, the time domain signals are converted to the frequency domain by using fast Fourier transform, the frequency domain features including power spectral density and frequency centroid are obtained, and time-frequency joint features are generated by using short-time Fourier transform, wherein the deep learning network architecture of the feature extraction module comprises: A one-dimensional convolution layer, which includes three convolution layers with convolution kernel sizes of 3, 5 and 7 respectively, and the step length is 1, which is used to extract local time domain features of the vibration signals; A two-dimensional convolution layer, which includes two convolution layers with convolution kernel sizes of 5x5, and the step length is 2, which is used to extract time-frequency distribution features of the vibration signals; A residual connection layer, which is composed of three residual blocks, each residual block includes two convolution layers, which is used to solve the gradient vanishing problem in deep network and enhance the feature expression ability; A global pooling layer, which uses average pooling method, is used to reduce dimension and aggregate the extracted features.

4. The deep learning algorithm based reduction gearbox failure prediction system of claim 1, wherein, The time series deep learning model of the fault prediction model module comprises: An input layer, which converts the multi-dimensional feature vector into a sequence format for time series deep learning network processing through a feature vector mapping unit according to time sequence; A hidden layer, which includes a long short-term memory unit layer and a gated recurrent unit layer, is used to capture long-term dependence and short-term change patterns in time series data; An output layer, which is configured with a fully connected layer and a softmax classifier, is used to output fault type probability distribution and fault degree evaluation result; The attention mechanism layer generates attention weights by calculating the similarity between the features of each time step and the context vector, giving different weights to the features of different time steps and highlighting key fault features.

5. The deep learning algorithm based reduction gearbox failure prediction system of claim 1, wherein, The fault level configured by the early warning module adopts a three-level progressive early warning threshold system, the specific content is as follows: Primary early warning, corresponding to the early abnormal state of the speed reducer, the triggering condition is that the predicted failure probability exceeds the first failure probability threshold P1 and is lower than the second failure probability threshold P2 in the continuous N1 sampling period; Intermediate early warning, corresponding to the potential failure state of the speed reducer, the triggering condition is that the predicted failure probability exceeds the second failure probability threshold P2 and is lower than the third failure probability threshold P3 in the continuous N2 sampling period; High-level early warning, corresponding to the emergency failure state of the speed reducer, the triggering condition is that the predicted failure probability exceeds the third failure probability threshold P3 or the failure development rate exceeds V; Wherein, N1 < N2, P1 < P2 < P3, V is the preset failure development rate threshold.

6. The deep learning algorithm based reduction gearbox failure prediction system of claim 5, wherein, The threshold value set in the early warning module is adjusted by real-time acquisition of equipment operating condition parameters, including equipment operating time, environmental temperature, load change rate; and according to the operating condition parameter and the preset condition-threshold mapping relationship table, the threshold value of each level in the three-level progressive early warning threshold system is adjusted. Wherein, the condition-threshold mapping relationship table is obtained by analyzing and counting the historical operation data and failure data of the speed reducer under different working conditions.

7. The deep learning algorithm based gearbox failure prediction system of claim 6, wherein, During the threshold adjustment process, the number of early warning triggers and the number of false alarms in the preset time period are counted; If the number of early warning triggers exceeds the first preset number and the false alarm rate is lower than the first preset false alarm rate, then the failure probability interval corresponding to each level of early warning threshold is adjusted upward according to the first adjustment coefficient; If the number of false alarms exceeds the second preset number, then the failure probability interval corresponding to each level of early warning threshold is adjusted downward according to the second adjustment coefficient; Otherwise, no response is made.

8. The deep learning algorithm based gearbox failure prediction system according to any one of claims 1-7, wherein, The system also includes: The model optimization module uses the small batch stochastic gradient descent algorithm to incrementally train the failure prediction model after new failure data is collected; at the same time, the model parameters are adjusted using the failure data of other similar equipment; The explanatory analysis module evaluates the importance of features by calculating the contribution of each feature to the model prediction result; and analyzes the causes of failure according to important features and historical failure data; The maintenance decision support module generates the optimal maintenance plan including maintenance time, maintenance method and required spare parts according to the failure type, failure degree and failure development trend, combined with maintenance history data; and then adjusts the scheduling arrangement of resources according to the maintenance plan and the existing resource situation.

9. The deep learning algorithm based gearbox failure prediction system of claim 8, wherein, The failure prediction system adopts a distributed architecture for deployment, including: Edge computing nodes are deployed on site to collect and preliminarily process data, and transmit the processed data to the fog computing nodes; Fog computing nodes are deployed on the local area network in the factory to extract features and perform edge inference, and transmit the inference results and key data to the cloud computing center; The cloud computing center is deployed in a remote data center to perform global model training, advanced analysis and decision support, and feedback the decision results to the fog computing nodes and edge computing nodes.

10. A method for failure prediction of a speed reducer based on a deep learning algorithm, characterized by, The fault prediction method comprises the following steps: With a fixed sampling period, vibration signals, temperature data, speed data and load data during the operation of the speed reducer are synchronously collected; The collected raw data is denoised and abnormal value corrected, and the data is normalized to a standard range; Based on a deep learning network architecture, the preprocessed data is feature extracted to construct a multi-dimensional feature vector; The multi-dimensional feature vector is used to train a time series deep learning model; The extracted feature vector is input into the trained time series deep learning model, and the model predicts the fault type, fault degree and fault development trend of the equipment according to the input feature vector, time series analysis of the hidden layer and classification calculation of the output layer; The prediction result is compared with a preset multi-level progressive threshold system, and corresponding level warning information is generated according to the threshold comparison result; By combining the fault type, degree and development trend, and the maintenance history data, the maintenance time, maintenance method and required spare parts are determined; and then according to the maintenance scheme and the existing resource condition, the maintenance personnel and the spare parts are arranged; The time series deep learning model is incrementally trained and the parameters are updated according to the newly collected fault data.

Images