Belt conveyor fault prediction method and system based on ECTimeMixer
Through the deep neural network model based on ECTimeMixer, the vibration, sound and temperature data of the belt conveyor are processed, and the problems of transparent fault information and low recognition accuracy in traditional fault prediction methods are solved, so as to detect faults in advance and reduce maintenance costs. The results improve the reliability and production efficiency of the equipment.
Patent Information
- Application Number
- CN202411170897.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-11-22
AI Technical Summary
The traditional belt conveyor fault prediction method faces the problem of difficulty in transparency of fault information and low accuracy in fault identification, which leads to unstable operation of port mechanical equipment and increases maintenance costs and downtime.
Using the ECTimeMixer-based belt drive failure prediction method, data processing and prediction are performed by obtaining vibration, sound and temperature data of belt drive parts, and dimensionality reduction technology and deep neural network models (including MultiScaleSeasonMixing, MultiScaleTrendMixing, EaaPastDecomposableMixing and Future_multi_mixing modules) through data processing and prediction, combined with Time-varying filtering empirical modal decomposition and data noise reduction method of eigenmode functions can detect potential faults in advance.
It improves the accuracy of fault prediction and equipment reliability, reduces downtime and maintenance costs, extends the service life of the equipment, and enhances the stability and production efficiency of port machinery and equipment.
Smart Images

Figure CN119004334A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of belt conveyor fault prediction, and in particular relates to a belt conveyor fault prediction method and system based on ECTimeMixer. Background Art
[0002] With the continuous expansion of trade and the rapid development of science and technology, the mechanical equipment of modern ports is undergoing unprecedented changes. The large-scale, efficient, automated, digitalized, intelligentized equipment and the pursuit of environmental protection and "green" technology have jointly promoted the progress of the port industry. The rapid growth of port throughput and the size of the shipping fleet has also put forward higher requirements and challenges for port machinery and equipment. In this context, as a key equipment in the port loading process, the stable and reliable operation of the belt conveyor is essential to ensure the port throughput efficiency. However, due to the frequent occurrence of harsh working conditions such as disturbances and strong impacts, as well as the complex service environment, the belt conveyor faces severe failure risks. Once a failure occurs, it will not only reduce the port's loading efficiency, but may also cause significant economic losses. Under the general trend of port automation and efficiency, it has become particularly urgent to conduct in-depth research on the fault diagnosis technology of belt conveyors. The types of belt conveyor failures are diverse and are closely related to their physical structure and working environment. Failures usually follow a certain development trend, gradually evolving from initial minor abnormalities to complete failure of equipment or components, and eventually leading to the shutdown of the entire system.
[0003] Due to the increasing complexity of equipment systems, traditional fault prediction methods face problems such as difficulty in making fault information transparent and low fault identification accuracy. Therefore, research on fault prediction of port belt conveyors tends to adopt advanced monitoring technologies and intelligent algorithms to more accurately identify and predict potential faults. At present, sensor networks integrated with Internet of Things (IoT) technology can monitor the operating status of belt conveyors in real time and collect key performance parameters. The application of big data analysis and machine learning algorithms can process these massive data and identify early signs of faults. The combination of these technologies not only improves the accuracy of fault prediction, but also makes predictive maintenance possible, which means that intervention can be carried out before a fault occurs, significantly reducing maintenance costs, reducing downtime, and extending the service life of equipment.
[0004] In summary, the fault prediction technology of belt conveyors is the key to ensuring efficient port operations, and is also an indispensable part of the transformation of modern ports to intelligence and digitalization. With the continuous advancement of technology, the fault prediction system of port belt conveyors will be more intelligent in the future, able to automatically adjust maintenance plans, and respond to various operating conditions in real time, further improving the overall operational efficiency and safety of ports. Summary of the invention
[0005] In view of the problems faced by traditional fault prediction methods, such as difficulty in making fault information transparent and low fault identification accuracy, the present invention provides a belt conveyor fault prediction method and system based on ECTimeMixer, so as to effectively detect potential faults of the belt conveyor in advance, reduce downtime, reduce maintenance costs, and improve equipment reliability and production efficiency. Its application can detect potential faults in advance and provide guidance for enterprises to make correct maintenance decisions.
[0006] According to a first aspect of the present invention, a belt conveyor fault prediction method based on ECTimeMixer is provided, the method comprising:
[0007] Obtain vibration, sound and temperature data of belt conveyor parts;
[0008] The vibration, sound and temperature data of belt conveyor parts are integrated by using dimensionality reduction technology to obtain one-dimensional time series data;
[0009] Downsample the one-dimensional time series data to obtain multi-scale time series data;
[0010] Construct the ECTimeMixer network model, which includes the MultiScaleSeasonMixing module, the MultiScaleTrendMixing module, the EaaPastDecomposableMixing module, and the Future_multi_mixing module;
[0011] The EaaPastDecomposableMixing module decomposes the multi-scale time series data into seasonal components and trend components, and processes them using the Eaa attention mechanism. The processed seasonal components are input into the MultiScaleSeasonMixing module, and the processed trend components are input into the MultiScaleTrendMixing module to merge features from different time scales.
[0012] The MultiScaleSeasonMixing module gradually merges high-scale seasonal information into low-scale information to obtain a multi-scale time series seasonal representation list; the MultiScaleTrendMixing module gradually transforms low-resolution trend signals into higher-resolution trend representations to generate a multi-scale time series trend representation list;
[0013] The future_multi_mixing module mixes a multiscale list of time series seasonal representations with a multiscale list of time series trend representations to obtain a multiscale sequence; generates a forecast for the multiscale sequence and sums it along the last dimension to obtain the forecast result.
[0014] Furthermore, the method also includes: filtering and denoising the acquired vibration, sound and temperature data of the belt conveyor parts by a data denoising method based on time-varying filter empirical mode decomposition TVFEMD and intrinsic mode function IMF energy entropy increment, as follows:
[0015] The multi-scale information of vibration data x(t) is separated by time-varying filtered empirical mode decomposition (TVFEMD) to determine its instantaneous amplitude A(t) and instantaneous phase. Right now:
[0016]
[0017] In the formula, is the Hilbert transform of x(t);
[0018] Define the analytical signal z(t) as:
[0019]
[0020] In the formula, e is a natural constant and j is an imaginary number;
[0021] The analytic signal z(t) is defined as the combination of two components:
[0022]
[0023] Calculate the local cutoff frequency
[0024]
[0025] Where a1(t) and a2(t) are the instantaneous amplitudes of the two-component signals, and are the instantaneous phases respectively;
[0026] Use a time-varying filter to filter the uncalculated signal to obtain the local mean, calculate the signal h(t), and record the final approximation result as m(t), that is:
[0027]
[0028] In the formula, is the local cutoff frequency;
[0029] The stopping conditions are:
[0030]
[0031] Among them, B Loughlin (t) is the Loughlin instantaneous bandwidth, is the weighted mean instantaneous frequency;
[0032] Continuously update x(t). If the signal satisfies θ(t)≤ξ, then x(t) is considered to be an IMF component at this time; otherwise, repeat the above steps until the decomposition is completed; where ξ is the bandwidth threshold;
[0033] The method of energy entropy increment of IMF components is used to judge false IMF components; that is, the signal energy of the t-th IMF component and the sum of the decomposed signal energy are calculated, that is:
[0034] E i(t) =i(t) 2
[0035]
[0036] In the formula, E i(t) is the energy value, i(t) is the tth IMF component, and n is the total number of IMF components;
[0037] Normalize each IMF component and obtain the energy spectrum:
[0038]
[0039] Q(i)=log[P(i)]
[0040] Where P(i) is the energy proportion of the IMF component, and Q(i) is the entropy value of the IMF component;
[0041] Calculate the energy entropy increment ΔQ(i) for each IMF component:
[0042] ΔQ(i)=-P(i)log[P(i)]
[0043] The larger the ΔQ(i) value is, the higher the authenticity of the IMF component is, and vice versa. To prevent the effective signal from being excessively eliminated, the discrimination threshold is set to ΔQ(i) / μ≥0.85, where μ=0.2.
[0044] Furthermore, the dimensionality reduction technology is used to integrate the vibration, sound and temperature data of the belt conveyor parts, including:
[0045] Combine the vibration, sound and temperature data of belt conveyor parts into three-dimensional data and standardize them;
[0046] Calculate the covariance matrix of three-dimensional data and solve for eigenvalues and eigenvectors;
[0047] The eigenvector with the largest eigenvalue is selected as the first principal component, and the three-dimensional data is projected onto the first principal component to obtain the one-dimensional time series data after dimensionality reduction and fusion.
[0048] Furthermore, the MultiScaleSeasonMixing module comprehensively considers the characteristics of time series data at different scales in a hierarchical manner. By fusing information layer by layer from high scale to low scale, the high-scale pattern information is gradually merged into the low scale, so that the network can integrate seasonal characteristics at different levels;
[0049] The MultiScaleSeasonMixing module defines and initializes two variables, out_high and out_low, to store the high- and low-scale time series data currently being processed. Out_season_list is then used to store the time series data after each scale has been processed. During the loop, the high-scale data out_high is downsampled through the downsampling layer, and then the result is added to the low-scale data out_low, and out_high is updated to the current out_low for use in the next loop step.
[0050] MultiScaleSeasonMixing has multiple levels, each of which downsamples the data, that is, reduces the number of data points to capture longer-period seasonal patterns. Each downsampling layer includes two linear variables and an activation function GELU. At each level, downsampling is performed and the result is merged with the data in the next level, mixing higher-resolution seasonal signals with lower-resolution seasonal signals, so that each level contains information on different time scales;
[0051] The MultiScaleSeasonMixing module starts by getting the highest resolution data of the sequence, and then processes it layer by layer. At each level, the data of the current resolution is mixed with the data of the previous level, and seasonal information at different scales is continuously added. Finally, a nested hierarchical list containing all possible seasonal cycles of the data is obtained, which is out_season_list.
[0052] Furthermore, the MultiScaleTrendMixing module gradually converts the captured low-resolution trend signals into higher-resolution trend representations by upsampling and mixing, integrates the information at all resolutions, and generates a multi-scale time series trend representation list;
[0053] The input trend_list contains time series trend representations at different downsampling levels, from the highest resolution trend to the lowest resolution trend. Use copy() to copy this list, and reverse() to make trend_list_reverse start from the lowest resolution and end at the highest resolution. Define out_low and out_high as the current lowest level trend signal and the next higher level trend signal, respectively. Initialize out_trend_list to contain only the lowest level trend signal.
[0054] MultiScaleTrendMixing has multiple levels, and each level will upsample the data, starting from the lowest resolution trend signal at the bottom level of the model, and gradually upsampling so that the original trend signal with the highest resolution can be reflected in the end. Upsampling means inserting new data points between the original data points of the time series or image to increase the resolution of the data. Each upsampling layer includes two linear variables and an activation function GELU.
[0055] Furthermore, the Eaa attention mechanism is used for processing, including:
[0056] Self-attention takes the seasonal component and the trend component x as input, where x∈R n×d , containing n tokens and d-dimensional embedding vectors; project the input x into query Q, key K, and value V, using three matrices WQ, WK, and WV. Each self-attention layer includes h heads, allowing the model to focus on different views of the input. Self-attention is described as:
[0057]
[0058] The final output is the processed time series data That is, the processed seasonal component and trend component.
[0059] Furthermore, the Future_multi_mixing module is used for multimodal mixing processing, that is, combining the encoded time series data to generate fault predictions at future time points; first initialize a list called dec_out_list to store the decoded output, take out the first element in the input x_list list, ignore the remaining components, process only one time series component at a time, input it to the predict_layers prediction layer, and then use permute to convert the dimension to (0,2,1) and input it to the projection_laye projection layer. The purpose of the projection_laye layer is to map data of one dimension to another dimension, convert the output of the data into a certain shape for subsequent steps, adjust the result shape to match the output configuration, and add it to the dec_out_list list; the Future_multi_mixing module is expressed as:
[0060]
[0061] in, It represents the prediction of the m-th scale sequence for the future, and the final output is Predictorm() represents the predictor of the mth scale sequence, M represents the length of the multi-scale information, represents multi-scale past information, and F×C represents the prediction of the m-th scale sequence for the future.
[0062] Furthermore, the method further comprises:
[0063] During training, create two empty lists x_list and x_mark_list to store the processed data. Traverse all input data. If the time mark of the data is not empty, take out the corresponding time series x and mark x_mark one by one according to the index. Then perform the following steps for each time series:
[0064] Get its batch size B, time length T, and number of features N, and perform standard normalization through CrossNorm. In 2Dcnn, each instance has C feature maps of size H×W. Given a feature map A×RH×W, instance normalization first normalizes the feature map and then performs an affine transformation:
[0065]
[0066] Where μA and σA are the mean and standard deviation; γ and β represent learnable affine parameters;
[0067] CrossNorm exchanges μA and σA of channel A with μB and σB of channel B, that is, converts β and γ into μ and σ, as shown below:
[0068]
[0069] Among them, channel A and channel B refer to seasonal component and trend component features respectively; CrossNorm exchanges the mean and variance of the feature map channel-wise during training, making the model more robust to style changes; add the processed time series and time tags to the list created previously;
[0070] The processed data x_list is further preprocessed by the pre_enc function. The role of the pre_enc method in the model is to perform data preprocessing and prepare the embedding layer. Then, each set of preprocessed data and tags are embedded using enc_embedding to convert them into a form suitable for neural network processing. This process is performed for each input fragment, and the result is added to the enc_out_list list. Then a loop is entered. Each loop applies the EaaPastDecomposableMixing module to the current enc_out_list, which is responsible for extracting features from the time series. The number of EaaPastDecomposableMixing encoding blocks is determined by self .layer hyperparameters determine that each block processes enc_out_list, and the encoded output enc_out_list is passed to the decoder block Future_multi_mixing. Future_multi_mixing generates future predictions based on the historical information collected by the encoder. The decoded output dec_out_list is stacked using the stack function and summed along the last dimension to merge the information from each hidden state. The normalization layer CrossNorm is used to denormalize the prediction results to restore their original scale and distribution. The processed prediction data dec_out is returned, which is the future failure time and failure type of the belt conveyor predicted by the model based on the input historical data points.
[0071] According to a second aspect of the present invention, there is provided a belt conveyor fault prediction system based on ECTimeMixer for implementing any one of the belt conveyor fault prediction methods based on ECTimeMixer described above, the system comprising a multi-source information acquisition system and a fault prediction system;
[0072] The multi-source information acquisition system includes a temperature sensor, a sound sensor, a vibration sensor and a data acquisition unit; the temperature sensor, the sound sensor and the vibration sensor are respectively connected to the data acquisition unit; the temperature sensor is used to obtain temperature data of belt conveyor parts, the sound sensor is used to obtain sound data of belt conveyor parts, the vibration sensor is used to obtain vibration data of belt conveyor parts, and the data acquisition unit is used to input original signals of the temperature sensor, the sound sensor and the vibration sensor into the fault prediction system;
[0073] The fault prediction system includes a data denoising unit, a data fusion unit and a fault prediction unit; the data denoising unit is used to filter and reduce the noise of the original signal; the data fusion unit is used to fuse the vibration, sound and temperature data of the belt conveyor parts using dimensionality reduction technology to obtain one-dimensional time series data; the fault prediction unit uses the ECTimeMixer model to predict the time and type of fault.
[0074] Furthermore, the fault prediction system also includes a database unit for storing the original signal input by the data acquisition unit.
[0075] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:
[0076] The present invention collects data by applying a variety of sensors (such as temperature sensors, vibration sensors, and sound sensors), and trains a belt conveyor fault prediction model through a deep neural network. This enables the belt conveyor to detect potential problems before a failure occurs, so that maintenance and repair can be performed in advance to avoid equipment failure at critical moments. By performing fault prediction and preventive maintenance on the belt conveyor, the reliability and stability of the equipment can be improved, and the service life of the equipment can be extended. The problem of increased complexity of the equipment system and difficulty in making the equipment status transparent is solved. Therefore, fault prediction of the belt conveyor through the ECTimeMixer neural network can help detect potential faults in advance, reduce downtime, reduce maintenance costs, and improve equipment reliability and production efficiency.
[0077] The present invention performs noise reduction and other operations on multi-source data in the database by using time-varying filter empirical mode decomposition TVFEMD and intrinsic mode function IMF. Then, the same type of data of parts in the same position are fused and normalized through dimensionality reduction technology. Combined with the learning ability of ECTimeMixer, the features in the data are automatically extracted through EaaPastDecomposableMixing. The decomposition is applied to multi-scale sequences, and the decomposed seasonal components and trend components are mixed in the fine to coarse and coarse to fine directions respectively, and the micro seasonal and macro trend information are aggregated in turn. Future_multi_mixingg further integrates multiple predictors to utilize the complementary prediction capabilities of multi-scale observations, so as to successfully predict the failure of the belt conveyor. BRIEF DESCRIPTION OF THE DRAWINGS
[0078] Figure 1 It is a structural diagram of the belt conveyor fault prediction system based on ECTimeMixer;
[0079] Figure 2 Schematic diagram of the Eaa attention mechanism;
[0080] Figure 3 This is a schematic diagram of the ECTimeMixer network model framework. DETAILED DESCRIPTION
[0081] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with 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 intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0082] Embodiment 1
[0083] A belt conveyor fault prediction system based on ECTimeMixer consists of a fault prediction system and a multi-source information acquisition system. The data acquisition unit is connected to a vibration sensor, a sound sensor and a temperature sensor, which are used to obtain the vibration, sound and temperature data of various parts of the belt conveyor. These original signals are input into the fault prediction system through the data acquisition unit. The fault prediction system processes the data of the temperature sensor, the sound sensor and the vibration sensor, and predicts the belt conveyor fault.
[0084] In addition, the fault prediction system consists of a database, a data denoising unit, a data fusion unit and a fault prediction unit. These units are connected together in sequence. The database is used to store the original signal input by the data acquisition unit. The data denoising unit performs denoising on the data stored in the database. The data fusion unit fuses the denoised data. The fault prediction unit uses the ECTimeMixer model to train the fused data to predict the time and type of fault.
[0085] A belt conveyor fault prediction method based on ECTimeMixer comprises the following steps:
[0086] Step 1: Obtain the temperature, sound and vibration data of various parts of the belt conveyor through temperature sensors, sound sensors and vibration sensors;
[0087] Step 2: The data acquisition unit inputs the original signals of these sensors into the database;
[0088] Step 3: The database stores the original signal and historical fault data input by the data acquisition unit;
[0089] Step 4: The data denoising unit denoises the data stored in the database;
[0090] Step 5: The data fusion unit performs data fusion on the feature extracted data;
[0091] Step 6: The fault prediction unit uses the ECTimeMixer model to train the fused data and predict the expected fault time and fault type.
[0092] Among them, the specific implementation method of step 4 is: the data denoising unit uses a data denoising method based on time-varying filtering empirical mode decomposition (TVFEMD) and intrinsic mode function (IMF) energy entropy increment to filter and denoise the original signal, wherein the TVFEMD method is used to dissociate the multi-scale information of the signal, and the IMF energy entropy increment method is used to accurately extract the effective signal.
[0093] TVFEMD is used to separate the multi-scale information of bridge monitoring data and determine the instantaneous amplitude A(t) and instantaneous phase of the signal to be decomposed x(t). Right now:
[0094]
[0095] in: is the Hilbert transform of x(t).
[0096] Define the analytical signal z(t) as:
[0097]
[0098] In the formula, e is a natural constant and j is an imaginary number;
[0099] The analytic signal z(t) is defined as the combination of two components:
[0100]
[0101] Calculate the local cutoff frequency
[0102]
[0103] Where a1(t) and a2(t) are the instantaneous amplitudes of the two-component signals, and are the instantaneous phases respectively;
[0104] Use a time-varying filter to filter the uncalculated signal to obtain the local mean, calculate the signal h(t), and record the final approximation result as m(t), that is:
[0105]
[0106] In the formula, is the local cutoff frequency;
[0107] The stopping conditions are:
[0108]
[0109] Among them, B Loughlin (t) is the Loughlin instantaneous bandwidth, is the weighted mean instantaneous frequency;
[0110] Continuously update x(t). If the signal satisfies θ(t)≤ξ, then x(t) is considered to be an IMF component at this time; otherwise, repeat the above steps until the decomposition is completed; where ξ is the bandwidth threshold.
[0111] Next, the energy entropy increment method of the IMF component is used to determine the false IMF component, as follows:
[0112] Calculate the signal energy of the t-th IMF component and the sum of the decomposed signal energy, that is:
[0113] E i(t) =i(t) 2
[0114]
[0115] Each IMF component is normalized and the energy spectrum is obtained respectively:
[0116]
[0117] Q(i)=log[P(i)]
[0118] Where: i(t) is the tth IMF component; n is the total number of IMF components.
[0119] Calculate the energy entropy increment for each IMF component:
[0120] ΔQ(i)=-P(i)log[P(i)]
[0121] The larger the ΔQ(i) value, the higher the authenticity of the IMF component, and vice versa. To prevent the effective signal from being excessively eliminated, the discrimination threshold is set to ΔQ(i) / μ≥0.85, where μ=0.2.
[0122] In step 5, PCA is used to fuse the vibration, sound and temperature data of the belt conveyor parts as follows:
[0123] First, standardize the data: combine the vibration, sound and temperature data of the belt conveyor parts into three-dimensional data, standardize the three-dimensional data (such as subtracting the mean and dividing by the standard deviation) to make the numerical range of each dimension the same, so as to avoid a certain dimension from having too much influence on the result. Then calculate the covariance matrix: calculate the covariance matrix for the standardized data. Then solve the eigenvalues and eigenvectors: calculate the eigenvalues and eigenvectors of the covariance matrix. The eigenvalues reflect the variance of the data in the direction of the corresponding eigenvector. Then select the principal component: usually select the eigenvector with the largest eigenvalue as the first principal component. The direction corresponding to this eigenvector is the direction with the largest variance of the data in three-dimensional space. Finally, project the data to the principal component: project the original three-dimensional data to the direction of this principal component to obtain one-dimensional data. This step is achieved by multiplying the original data with the eigenvector of the first principal component.
[0124] Embodiment 2
[0125] A belt conveyor fault prediction system based on ECTimeMixer, such as Figure 1As shown, it includes a fault prediction system and a multi-source information acquisition system, wherein the multi-source information acquisition system includes a temperature sensor, a sound sensor, a vibration sensor and a data acquisition unit, the temperature sensor, the sound sensor and the vibration sensor are respectively connected to the data acquisition unit, the temperature sensor is used to obtain the temperature of the parts of various parts of the belt conveyor, the sound sensor is used to obtain the sound of the parts of various parts of the belt conveyor, the vibration sensor is used to obtain the vibration data of the parts of various parts of the belt conveyor, the data acquisition unit is used to input the original signals of the temperature sensor, the sound sensor and the vibration sensor into the fault prediction system, and the fault prediction system is used to determine the type of belt conveyor fault according to the data of the temperature sensor, the sound sensor and the vibration sensor.
[0126] The installation position of the sensor will affect the data collection intensity and noise level, so the installation position needs to be selected according to the characteristics of the belt conveyor. The belt conveyor is generally installed in the following parts: Belt: Sensors can be installed at different parts of the belt, such as the feed end, discharge end or middle position of the belt. Bearing: Installing sensors at the bearings of the belt conveyor can determine whether the bearings are faulty or worn. Drive device: Installing sensors on the drive device of the belt conveyor can detect whether there are overheating, abnormal vibration or noise problems. Tensioning device: Installing sensors on the tensioning device of the belt conveyor can determine whether there is insufficient tension or abnormal conditions. Support structure: Installing sensors on the supporting structure of the belt conveyor can detect whether there are problems such as structural looseness, deformation or wear. The installation position of the sensor of the belt conveyor should be determined according to the design and operation of the belt conveyor. When installing the sensor, factors such as the contact method between the sensor and the monitored object, the sensitivity and reliability of the sensor should be considered to ensure that the relevant data can be accurately and reliably monitored.
[0127] At the same time, different types or quantities of sensors can be added to the data acquisition system according to the application scenario and belt conveyor requirements. The collected sound, vibration and temperature data are first filtered and denoised by a data denoising method based on time-varying filter empirical mode decomposition TVFEMD and intrinsic mode function IMF energy entropy increment, and then the data is fused and normalized using dimensionality reduction technology. Then, the data corresponding to the equipment that requires multi-source data for fault prediction is used as the input of the fault prediction module.
[0128] The fault prediction system consists of a database, a data denoising unit, a data fusion unit and a fault prediction unit. These units are connected in sequence to realize a complete set of processes. The database is used to store the original signal input by the data acquisition unit, and the data denoising unit is responsible for processing the outliers of the data stored in the database. The fault prediction unit uses machine learning algorithms to train the data and optimize the parameters to achieve accurate fault identification of the belt conveyor. Through this system, we can effectively collect, process and predict the fault information of the belt conveyor.
[0129] A belt conveyor fault prediction method based on ECTimeMixer comprises the following steps:
[0130] Step S1: Use different types of sensors to collect temperature, sound, and vibration data of the belt conveyor's operating environment. The same type of data is often used to predict faults in the same parts of different parts of the belt conveyor. Specifically, belts: vibration and sound sensors can be installed to detect whether there are abnormal conditions. Bearings: Temperature and vibration sensors can be installed to determine whether the bearings are faulty or worn. Drive device: Install temperature and vibration sensors to detect whether there are overheating or abnormal vibration problems. Tensioning device: Install temperature, vibration, and sound sensors to determine whether there is insufficient tension or abnormal conditions. Support structure: Install a vibration sensor to monitor the vibration of the structure to detect whether the structure is loose, deformed, or worn.
[0131] Step S2: The collected sound, vibration and temperature data are first filtered and denoised by a data denoising method based on time-varying filtered empirical mode decomposition (TVFEMD) and intrinsic mode function (IMF) energy entropy increment to remove random noise in the data. Then, the data is fused and normalized using dimensionality reduction technology.
[0132] Step S3: Build the MultiScaleSeasonMixing module of the ECTimeMixer network model, such as Figure 3 As shown in the figure, the MultiScaleSeasonMixing class comprehensively considers the characteristics of time series data at different scales in a hierarchical manner. It helps capture the long-term dependencies and multi-scale seasonal patterns in time series data by fusing information from high scale to low scale. By gradually merging high-scale pattern information to low scale, the network can integrate seasonal characteristics at different levels.
[0133] MultiScaleSeasonMixing defines two variables out_high and out_low and initializes them to save the high- and low-scale time series data currently being processed, and then uses out_season_list to store the time series data after each scale is processed. During the loop, the high-scale data out_high is downsampled through the downsampling layer, and then the result is added to the low-scale data out_low, and out_high is updated to the current out_low for use in the next loop step.
[0134] MultiScaleSeasonMixing has multiple levels, each of which downsamples the data, that is, reduces the number of data points to capture longer-period seasonal patterns. Each downsampling layer includes two linear variables and an activation function GELU. At each level, the model performs downsampling and merges the results with the data in the next level. This mixes seasonal signals with higher resolution (short cycle) and seasonal signals with lower resolution (long cycle), so that each level contains information on different time scales.
[0135] The MultiScaleSeasonMixing model starts with the highest resolution data of the sequence, and then processes it layer by layer, mixing the current resolution data with the data of the previous layer at each level, continuously adding seasonal information at different scales. In the end, we get a nested hierarchical list containing all possible seasonal cycles of the data, which is out_season_list.
[0136] Step S4: Build the MultiScaleTrendMixing module of the ECTimeMixer network model. MultiScaleTrendMixing gradually converts the captured low-resolution trend signal into a higher-resolution trend representation by upsampling and mixing, integrates information at all resolutions, and generates a multi-scale time series trend representation list.
[0137] The input trend_list contains time series trend representations at different downsampling levels, from the highest resolution trend (shortest short term) to the lowest resolution trend (longest term). Use copy() to copy this list and reverse() so that trend_list_reverse starts from the lowest resolution and ends at the highest resolution. Define out_low and out_high as the current lowest level trend signal and the next higher level trend signal, respectively. Initialize out_trend_list to contain only the lowest level trend signal.
[0138] MultiScaleTrendMixing has multiple levels, and each level will upsample the data. These layers start from the lowest resolution trend signal at the bottom layer of the model and gradually upsample so that the original trend signal with the highest resolution can be reflected in the end. Upsampling is to insert new data points between the original data points of the time series or image to increase the resolution of the data. Each upsampling layer includes two linear variables and an activation function GELU.
[0139] Step S5: Build the EaaPastDecomposableMixing module of the ECTimeMixer network model. Each input sequence is first decomposed by the decompsition module and decomposed into seasonal and trend components according to the settings. These components help the model understand the long-term trends and periodic patterns in the time series. For the seasonal and trend components, the seasonality is input to the MultiScaleSeasonMixing module, and the trend is input to the MultiScaleTrendMixing module. The features from different time scales are merged. This step is to allow the model to capture more fine-grained time series information at different scales. The decomposed components can be further processed by the EfficientAdditiveAttnetion (abbreviated as Eaa) attention mechanism. The Eaa attention mechanism allows the model to better capture the long-distance interdependence in time series data. The processed seasonal and trend components are then merged with the original data. The PDM block can be formalized as:
[0140]
[0141] The season part 1 and the trend part t l is a multi-scale time series X l Decomposed.
[0142] The Eaa attention mechanism replaces the traditional quadratic matrix multiplication operation with linear element-wise multiplication, effectively reducing computational complexity. The key-value interaction is simplified to a linear layer without explicit key-value interaction, further reducing the amount of computation, such as Figure 2 Specifically, self-attention takes x as input, where x∈R n×d , containing n tokens and d-dimensional embedding vectors. The input x is projected into query (Q), key (K), and value (V) using three matrices WQ, WK, and WV. Each self-attention layer includes h heads, which allows the model to focus on different views of the input. Self-attention can be described as:
[0143]
[0144] The decomposed seasonal components and trend components are input into EAA respectively, and the final output is the processed time series data, which contains the long-term trend and seasonal information of the original data.
[0145] Step S6: Build the Future_multi_mixing module of the ECTimeMixer network model. Its purpose is to perform multimodal mixing processing, that is, to combine the encoded time series data to generate fault predictions at future time points. First, initialize a list called dec_out_list to store the decoded output, take out the first element in the input x_list list, ignore the remaining components, and only process one time series component at a time. Input it to the predict_layers prediction layer to get the result. Use permute to convert the dimension to (0,2,1) and input it to the projection_laye projection layer. The purpose of the projection_laye layer is to map data of one dimension to another dimension and convert the output of the data into a certain shape for subsequent steps. Adjust the shape of the result to match the output configuration and add it to the dec_out_list list. The Future_multi_mixing block is represented as:
[0146]
[0147] in Represents the prediction of the m-th scale sequence for the future, and the final output is ∈R F×C Predictorm(·) represents the predictor of the m-th scale sequence,
[0148] Step S7: The data is divided into different training sets and test sets in a ratio of 7:3. The training set is used to train the network model, and the test set is used to evaluate the accuracy of the prediction model. The training set is used as a training sample and the training sample is input into the ECTimeMixer network. The specific process is as follows:
[0149] Create two empty lists x_list and x_mark_list to store the processed data, traverse all input data, and if the time mark of the data is not empty (x_mark_encisnotNone), take out the corresponding time series x and mark x_mark one by one according to the index. Next, perform the following steps for each time series:
[0150] Get its (referring to the time series x and the mark x_mark) batch size (B), time length (T), and number of features / variables (N), and perform standard normalization through CrossNorm. In 2Dcnn, each instance has C feature maps of size H×W. Given a feature map A×RH×W, instance normalization first normalizes the feature map and then performs an affine transformation:
[0151]
[0152] μA and σA are the mean and standard deviation; γ and β represent learnable affine parameters. In order to expand the training distribution, CrossNorm swaps μA and σA of channel A with μB and σB of channel B, that is, converts β and γ into μ and σ, as shown below:
[0153]
[0154] In this invention, channel A and channel B represent the two feature channels of season and trend respectively. The purpose of CrossNorm is to exchange the mean and variance of the feature map channel-wise during training, making the model more robust to style changes. If channel_independent is set, transform the data to match the expected input size to feed the model, and finally add the processed time series and time tags to the list created previously.
[0155] The processed data x_list is further preprocessed by the pre_enc function. The role of the pre_enc method in the model is to perform data preprocessing and prepare the embedding layer. Then, each set of preprocessed data and tags are embedded using enc_embedding to convert them into a form suitable for neural network processing. This process is performed for each input fragment, and the result is added to the enc_out_list list. Then a loop is entered. Each loop applies the EaaPastDecomposableMixing block to the current enc_out_list, which is responsible for extracting features from the time series. The number of EaaPastDecomposableMixing encoding blocks is determined by self.lay The er hyperparameter determines that each block processes enc_out_list, and the encoded output enc_out_list is passed to the decoder block Future_multi_mixingg. Future_multi_mixingg generates future predictions based on the historical information collected by the encoder. The decoded output dec_out_list is stacked using the stack function and summed along the last dimension to merge the information from each hidden state. The first normalization layer CrossNorm is used to "de-normalize" the prediction results to restore their original scale and distribution. The processed prediction data dec_out is returned, which is the future failure time and failure type of the belt conveyor predicted by the model based on the input historical data points.
[0156] Step S8: In order to predict faults of parts such as the same motor in different parts, we used the ECTimeMixer network for training. Since parts in different parts have differences in vibration signals, sound and temperature signals, we need to perform separate data processing and prediction for each part. We input the test set of parts in various parts of the belt conveyor into the model to verify the accuracy of the model. In each training, we evaluate the prediction accuracy of the algorithm by comparing the output value of the model with the actual value, and obtain a more accurate fault prediction model through iterative training of the data. Once the trained prediction model receives the input of real-time monitoring data, it can give the current prediction result.
[0157] In summary, the present invention uses the ECTimeMixer network to perform fault prediction on the sound, vibration and temperature data of different parts of the belt conveyor. EaaPastDecomposableMixing can successfully aggregate detailed seasonal information starting from the most refined sequence and go deep into the macro trend components and knowledge from coarser scales. In the prediction stage, Future_multi_mixingg integrates multiple prediction factors to utilize the complementary prediction capabilities of multi-scale observations. The present invention performs fusion analysis based on multi-source signals and then uses the ECTimeMixer network for prediction, which enables the belt conveyor to detect potential problems before the failure occurs, so as to perform maintenance and repair in advance and avoid equipment failure at critical moments. By performing fault prediction and preventive maintenance on the belt conveyor, the reliability and stability of the equipment can be improved, and the accuracy of fault prediction can be improved. Its application can detect potential faults in advance and provide guidance for enterprises to make correct maintenance decisions, reduce downtime, and help improve equipment utilization and reduce maintenance time costs.
[0158] It should be noted that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.
[0159] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, and two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.
[0160] It will be easily understood by those skilled in the art that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A belt conveyor fault prediction method based on ECTimeMixer, characterized in that: The method includes: Obtain vibration, sound and temperature data of belt conveyor parts; The vibration, sound and temperature data of belt conveyor parts are integrated by using dimensionality reduction technology to obtain one-dimensional time series data; Downsample the one-dimensional time series data to obtain multi-scale time series data; Construct the ECTimeMixer network model, which includes the MultiScaleSeasonMixing module, the MultiScaleTrendMixing module, the EaaPastDecomposableMixing module, and the Future_multi_mixing module; The EaaPastDecomposableMixing module decomposes the multi-scale time series data into seasonal components and trend components, and processes them using the Eaa attention mechanism. The processed seasonal components are input into the MultiScaleSeasonMixing module, and the processed trend components are input into the MultiScaleTrendMixing module to merge features from different time scales. The MultiScaleSeasonMixing module gradually merges high-scale seasonal information into low-scale information to obtain a multi-scale time series seasonal representation list; the MultiScaleTrendMixing module gradually transforms low-resolution trend signals into higher-resolution trend representations to generate a multi-scale time series trend representation list; The future_multi_mixing module mixes a multiscale list of time series seasonal representations with a multiscale list of time series trend representations to obtain a multiscale sequence; generates a forecast for the multiscale sequence and sums it along the last dimension to obtain the forecast result.
2. The belt conveyor fault prediction method based on ECTimeMixer according to claim 1, characterized in that: The method also includes: filtering and denoising the acquired vibration, sound and temperature data of the belt conveyor parts by a data denoising method based on time-varying filter empirical mode decomposition TVFEMD and intrinsic mode function IMF energy entropy increment, as follows: The multi-scale information of vibration data x(t) is separated by time-varying filtered empirical mode decomposition (TVFEMD) to determine its instantaneous amplitude A(t) and instantaneous phase. Right now: In the formula, is the Hilbert transform of x(t); Define the analytical signal z(t) as: In the formula, e is a natural constant and j is an imaginary number; The analytic signal z(t) is defined as the combination of two components: Calculate the local cutoff frequency Where a1(t) and a2(t) are the instantaneous amplitudes of the two-component signals, and are the instantaneous phases respectively; Use a time-varying filter to filter the uncalculated signal to obtain the local mean, calculate the signal h(t), and record the final approximation result as m(t), that is: In the formula, is the local cutoff frequency; The stopping conditions are: Among them, B L oughlin(t) is the Loughlin instantaneous bandwidth, is the weighted mean instantaneous frequency; Continuously update x(t). If the signal satisfies θ(t)≤ξ, then x(t) is considered to be an IMF component at this time; otherwise, repeat the above steps until the decomposition is completed; where ξ is the bandwidth threshold; The method of energy entropy increment of IMF components is used to judge false IMF components; that is, the signal energy of the t-th IMF component and the sum of the decomposed signal energy are calculated, that is: E i(t) =i(t) 2 In the formula, E i(t) is the energy value, i(t) is the tth IMF component, and n is the total number of IMF components; Normalize each IMF component and obtain the energy spectrum: Q(i)=log[P(i)] Where P(i) is the energy proportion of the IMF component, and Q(i) is the entropy value of the IMF component; Calculate the energy entropy increment ΔQ(i) for each IMF component: ΔQ(i)=-P(i)log[P(i)] The larger the ΔQ(i) value is, the higher the authenticity of the IMF component is, and vice versa. To prevent the effective signal from being excessively eliminated, the discrimination threshold is set to ΔQ(i) / μ≥0.85, where μ=0.
2.
3. The belt conveyor fault prediction method based on ECTimeMixer according to claim 1, characterized in that: Utilize dimensionality reduction technology to fuse vibration, sound and temperature data of conveyor belt parts, including: Combine the vibration, sound and temperature data of belt conveyor parts into three-dimensional data and standardize them; Calculate the covariance matrix of three-dimensional data and solve for eigenvalues and eigenvectors; The eigenvector with the largest eigenvalue is selected as the first principal component, and the three-dimensional data is projected onto the first principal component to obtain the one-dimensional time series data after dimensionality reduction and fusion.
4. The belt conveyor fault prediction method based on ECTimeMixer according to claim 1, characterized in that: The MultiScaleSeasonMixing module comprehensively considers the characteristics of time series data at different scales in a hierarchical manner. By fusing information layer by layer from high scale to low scale, the high-scale pattern information is gradually merged into the low-scale, so that the network can integrate seasonal characteristics at different levels. The MultiScaleSeasonMixing module defines and initializes two variables, out_high and out_low, to store the high- and low-scale time series data currently being processed. Out_season_list is then used to store the time series data after each scale has been processed. During the loop, the high-scale data out_high is downsampled through the downsampling layer, and then the result is added to the low-scale data out_low, and out_high is updated to the current out_low for use in the next loop step. MultiScaleSeasonMixing has multiple levels, each of which downsamples the data, that is, reduces the number of data points to capture longer-period seasonal patterns. Each downsampling layer includes two linear variables and an activation function GELU. At each level, downsampling is performed and the result is merged with the data in the next level, mixing higher-resolution seasonal signals with lower-resolution seasonal signals, so that each level contains information on different time scales; The MultiScaleSeasonMixing module starts by getting the highest resolution data of the sequence, and then processes it layer by layer. At each level, the data of the current resolution is mixed with the data of the previous level, and seasonal information at different scales is continuously added. Finally, a nested hierarchical list containing all possible seasonal cycles of the data is obtained, which is out_season_list.
5. The belt conveyor fault prediction method based on ECTimeMixer according to claim 1, characterized in that: The MultiScaleTrendMixing module gradually converts the captured low-resolution trend signal into a higher-resolution trend representation by upsampling and mixing, integrates the information at all resolutions, and generates a multi-scale time series trend representation list; The input trend_list contains time series trend representations at different downsampling levels, from the highest resolution trend to the lowest resolution trend. Use copy() to copy this list, and reverse() to make trend_list_reverse start from the lowest resolution and end at the highest resolution. Define out_low and out_high as the current lowest level trend signal and the next higher level trend signal, respectively. Initialize out_trend_list to contain only the lowest level trend signal. MultiScaleTrendMixing has multiple levels, and each level will upsample the data, starting from the lowest resolution trend signal at the bottom level of the model, and gradually upsampling so that the original trend signal with the highest resolution can be reflected in the end. Upsampling means inserting new data points between the original data points of the time series or image to increase the resolution of the data. Each upsampling layer includes two linear variables and an activation function GELU.
6. The belt conveyor fault prediction method based on ECTimeMixer according to claim 1, characterized in that: Use the Eaa attention mechanism for processing, including: Self-attention takes the seasonal component and the trend component x as input, where x∈R n×d , containing n tokens and d-dimensional embedding vectors; project the input x into query Q, key K, and value V, using three matrices WQ, WK, and WV. Each self-attention layer includes h heads, allowing the model to focus on different views of the input. Self-attention is described as: The final output is the processed time series data That is, the processed seasonal component and trend component.
7. The belt conveyor fault prediction method based on ECTimeMixer according to claim 1, characterized in that: The Future_multi_mixing module is used for multimodal mixing processing, that is, combining the encoded time series data to generate fault predictions at future time points; first initialize a list called dec_out_list to store the decoded output, take out the first element in the input x_list list, ignore the remaining components, process only one time series component at a time, input it to the predict_layers prediction layer, and then use permute to convert the dimension to (0,2,1) and input it to the projection_laye projection layer. The purpose of the projection_laye layer is to map data of one dimension to another dimension and convert the output of the data into a certain shape for subsequent steps. The result is reshaped to match the output configuration and added to the dec_out_list list; the Future_multi_mixing module is represented as: in, It represents the prediction of the m-th scale sequence for the future, and the final output is Predictorm() represents the predictor of the mth scale sequence, M represents the length of the multi-scale information, represents multi-scale past information, and F×C represents the prediction of the m-th scale sequence for the future.
8. The belt conveyor fault prediction method based on ECTimeMixer according to claim 1, characterized in that: The method further includes: During training, create two empty lists x_list and x_mark_list to store the processed data. Traverse all input data. If the time mark of the data is not empty, take out the corresponding time series x and mark x_mark one by one according to the index. Then perform the following steps for each time series: Get its batch size B, time length T, and number of features N, and perform standard normalization through CrossNorm. In 2Dcnn, each instance has C feature maps of size H×W. Given a feature map A×RH×W, instance normalization first normalizes the feature map and then performs an affine transformation: Where μA and σA are the mean and standard deviation; γ and β represent learnable affine parameters; CrossNorm exchanges μA and σA of channel A with μB and σB of channel B, that is, converts β and γ into μ and σ, as shown below: Among them, channel A and channel B refer to seasonal component and trend component features respectively; CrossNorm exchanges the mean and variance of the feature map channel-wise during training, making the model more robust to style changes; add the processed time series and time tags to the list created previously; The processed data x_list is further preprocessed by the pre_enc function. The role of the pre_enc method in the model is to perform data preprocessing and prepare the embedding layer. Then, each set of preprocessed data and tags are embedded using enc_embedding to convert them into a form suitable for neural network processing. This process is performed for each input fragment, and the result is added to the enc_out_list list. Then a loop is entered. Each loop applies the EaaPastDecomposableMixing module to the current enc_out_list, which is responsible for extracting features from the time series. The number of EaaPastDecomposableMixing encoding blocks is determined by self .layer hyperparameters determine that each block processes enc_out_list, and the encoded output enc_out_list is passed to the decoder block Future_multi_mixing. Future_multi_mixing generates future predictions based on the historical information collected by the encoder. The decoded output dec_out_list is stacked using the stack function and summed along the last dimension to merge the information from each hidden state. The normalization layer CrossNorm is used to denormalize the prediction results to restore their original scale and distribution. The processed prediction data dec_out is returned, which is the future failure time and failure type of the belt conveyor predicted by the model based on the input historical data points.
9. A belt conveyor fault prediction system based on ECTimeMixer for implementing the belt conveyor fault prediction method based on ECTimeMixer as claimed in any one of claims 1 to 8, characterized in that: The system includes a multi-source information acquisition system and a fault prediction system; The multi-source information acquisition system includes a temperature sensor, a sound sensor, a vibration sensor and a data acquisition unit; the temperature sensor, the sound sensor and the vibration sensor are respectively connected to the data acquisition unit; the temperature sensor is used to obtain temperature data of belt conveyor parts, the sound sensor is used to obtain sound data of belt conveyor parts, the vibration sensor is used to obtain vibration data of belt conveyor parts, and the data acquisition unit is used to input original signals of the temperature sensor, the sound sensor and the vibration sensor into the fault prediction system; The fault prediction system includes a data denoising unit, a data fusion unit and a fault prediction unit; the data denoising unit is used to filter and denoise the original signal; The data fusion unit is used to fuse the vibration, sound and temperature data of belt conveyor parts using dimensionality reduction technology to obtain one-dimensional time series data; the fault prediction unit uses the ECTimeMixer model to predict the time and type of fault.
10. The belt conveyor fault prediction method based on ECTimeMixer according to claim 9, characterized in that: The fault prediction system also includes a database unit for storing the original signal input by the data acquisition unit.
Citation Information
Cited By
Servo motor control method, control device and control system
CN120415211A
Time sequence set prediction method and system and storage medium
CN120849827A
