Double-scale dynamic trend prediction method and system applied to metallurgical manufacturing industry

Abstract

The application provides a double-scale dynamic trend prediction method and system applied to a metallurgical manufacturing industry. The method comprises the following steps: collecting vibration signals during normal operation of a bearing, and extracting time domain features of the vibration signals; constructing a local degradation trend by using time domain features of vibration signals of multiple working cycles, constructing a local trend prediction model, training the local trend prediction model by using the local degradation trend, reasoning the time domain features of the vibration signals of the current working cycle by using the local trend prediction model, and predicting a real-time running state of the bearing; extracting peak values of the time domain features of all the vibration signals, constructing a global degradation trend by smoothing the peak values, constructing a global trend prediction model, training the global trend prediction model by using the global degradation trend, reasoning the time domain features of all the vibration signals by using the global trend prediction model, and predicting a global performance degradation state of the bearing. The method can simultaneously predict the real-time running state and the global performance degradation state of the bearing.

Description

Technical Field

[0001] This invention relates to the field of bearing performance testing, and specifically to a dual-scale dynamic trend prediction method and system applied to the metallurgical manufacturing industry. Background Technology

[0002] Rolling bearings, characterized by low frictional damping and ease of maintenance, are widely used in the metallurgical manufacturing industry as key components of rotating machinery. The health of bearings directly affects the operation of production equipment. Fatigue damage that is not repaired promptly can trigger a chain reaction of failures, leading to equipment downtime, reduced production efficiency, and potentially significant economic losses. Therefore, real-time monitoring of bearing operating conditions, timely and accurate prediction of performance degradation trends, and rational planning of equipment maintenance schedules are beneficial for extending equipment lifespan, improving production efficiency, and preventing substantial economic losses.

[0003] Current trend prediction technologies target experimental bearings with stable operating conditions, good degradation monotonicity, and low noise interference. However, in actual service, especially in the metallurgical manufacturing industry, bearing operating conditions vary with the manufacturing process cycle due to the influence of equipment manufacturing cycle. This results in unstable operating conditions, weak degradation trend monotonicity, high noise interference, and a certain degree of periodicity. Figure 1 As shown, this leads to a decline in the reliability and accuracy of health indicators and prediction models, making it impossible to accurately predict performance degradation trends. Summary of the Invention

[0004] In order to overcome the defects in the prior art, the purpose of this invention is to provide a dual-scale dynamic trend prediction method and system for metallurgical manufacturing.

[0005] To achieve the above-mentioned objectives of this invention, this invention provides a dual-scale dynamic trend prediction method applied to the metallurgical manufacturing industry, comprising the following steps:

[0006] Vibration signals during normal bearing operation are collected, and the time-domain features of the vibration signals are extracted.

[0007] A local degradation trend is constructed using the time-domain characteristics of vibration signals from multiple working cycles. A local trend prediction model is then built. The local degradation trend is used to train the local trend prediction model. The local trend prediction model is then used to infer the time-domain characteristics of the vibration signal in the current working cycle to predict the real-time operating status of the bearing.

[0008] The peak values ​​of the time-domain features of all vibration signals are extracted, and the peak values ​​are smoothed to construct a global degradation trend. A global trend prediction model is constructed, and the global trend prediction model is trained using the global degradation trend. The global trend prediction model is then used to infer the time-domain features of all vibration signals to predict the global performance degradation state of the bearing.

[0009] This method combines the construction of local and global degradation trends, and uses local and global predictions respectively to achieve dynamic trend prediction based on the characteristics of degradation trends. It can simultaneously predict the real-time operating status and global performance degradation status of bearings, and realize dual-scale dynamic trend prediction.

[0010] In one alternative scheme of this dual-scale dynamic trend prediction method applied to the metallurgical manufacturing industry, the steps for constructing the local degradation trend are as follows:

[0011] Remove vibration signals that belong to the bearing stop state from the vibration signals to obtain vibration signals for each working cycle;

[0012] The local degradation trend is constructed by utilizing the temporal characteristics of vibration signals from several working cycles prior to the prediction time.

[0013] This alternative approach reduces the impact of downtime data on the local trend prediction model by deleting data collected when the bearing is down.

[0014] In one alternative scheme of this dual-scale dynamic trend prediction method applied to the metallurgical manufacturing industry, the step of extracting the peak values ​​of the time-domain features of all vibration signals is as follows:

[0015] The eigenvector F, composed of the time-domain features of all vibration signals, is differentially processed to obtain Fi. diff Extract F diff The coordinates of points with a mean difference of 0 are used to obtain Fzeros;

[0016] By performing a difference operation on Fzeros, we obtain F d-zeros Whether the waveform containing Fzeros is horizontal can be determined by whether the coordinate difference between two adjacent points in Fzeros is 1.

[0017] According to F d-zeros For F diff After correction, we get F d ′ iff The correction ensures that each horizontal waveform has only one peak.

[0018] Determine F d ′ iff Whether the characteristic value corresponding to each time step is a peak value;

[0019] By using periodic peak distance detection, redundant peaks are removed, ensuring that there is only one peak within each peak period, thus obtaining a peak sequence that contains the time-domain characteristics of all vibration signals.

[0020] In this optional scheme, peak extraction involves extracting a peak value for each work cycle (manufacturing cycle), representing the degradation state of that work cycle. The peak values ​​extracted from all cycles form a peak sequence, which is the global health indicator. The peak values ​​express the global degradation trend, providing an accurate data foundation for the accuracy of prediction.

[0021] In one alternative scheme of the dual-scale dynamic trend prediction method applied to the metallurgical manufacturing industry, the peak values ​​of the time-domain characteristics of all vibration signals are smoothed based on the following formula to obtain the global degradation trend.

[0022] s i =αf pi +(1-α)(s i-1 +Δs i-1 )

[0023] Δs i =β(s) i -s i-1 )+(1-β)Δs i-1

[0024] Where α and β are memory decay factors, s i Let time t i The peak value f of the time-domain characteristics of the vibration signal pi The smoothed value, Δs i This represents the trend.

[0025] Since the peaks extracted by the differential peak extraction method still have obvious local fluctuations, and since performance degradation mainly focuses on the monotonicity of the peak sequence, this alternative scheme can reduce the local fluctuations of the peak sequence and enhance the monotonicity by using smoothing processing.

[0026] In one alternative approach to the dual-scale dynamic trend prediction method applied to the metallurgical manufacturing industry, a GRU model is constructed, a dataset is obtained by sliding windowing of the local degradation trend, and the GRU model is trained based on the dataset to obtain a local trend prediction model.

[0027] Since the local trends exhibit obvious periodicity, the model only needs m periods of data to learn the characteristics of the local trends. Therefore, this embodiment uses sliding window processing to extract m periods of data for input into the model, thereby reducing the computational load of model parameters.

[0028] In one alternative scheme of this dual-scale dynamic trend prediction method applied to the metallurgical manufacturing industry, the steps for constructing the global trend prediction model are as follows:

[0029] The time series of the global degradation trend is tested for stationarity. If it is not stationary, the global degradation trend is subjected to difference transformation and then the stationarity test is performed again.

[0030] Examine whether the time series of the global degradation trend is white noise. If not, construct an ARI MA model and determine the order of the ARI MA model based on the global degradation trend.

[0031] A global trend prediction model is obtained by training the ARI MA model using the global degradation trend.

[0032] In this optional option, the ARI MA model can analyze the trend, periodicity, and seasonality characteristics of time series data, and can be used for global trend prediction with high accuracy.

[0033] In one alternative scheme of the dual-scale dynamic trend prediction method applied to the metallurgical manufacturing industry, the parameter d of the ARI MA model is determined based on the difference order of the stationarity test, and the autoregression order p and the moving average order q of the ARI MA model are determined according to the following formula.

[0034]

[0035] BIC(p0′,q0′)=minBIC(p,q),0≤p,q≤L

[0036] In the formula, The variance of the time series input to the ARI MA model is N, the number of time points in the time series input to the ARI MA model is L, and the maximum order of the moving average term q in the ARI MA model is set.

[0037] When p0′ and q0′ satisfy the above formula, p0′ and q0′ are the optimal autoregressive order and the optimal moving average order, respectively.

[0038] This option allows for quick and accurate order determination of the ARIMA model, enabling the construction of a global trend prediction model.

[0039] This application also proposes a bearing performance prediction system, including a processing module, a storage module, and a vibration acceleration sensor for acquiring bearing vibration signals;

[0040] The vibration acceleration sensor is connected to the processing module and sends the vibration signal it has collected to the processing module. The processing module and the storage module are interconnected. The storage module is used to store at least one executable instruction. The executable instruction causes the processing module to perform the operation corresponding to the dual-scale dynamic trend prediction method applied to the metallurgical manufacturing industry as described above, based on the vibration signal, to predict the bearing performance.

[0041] This bearing performance prediction system possesses all the advantages of the aforementioned dual-scale dynamic trend prediction method applied to the metallurgical manufacturing industry.

[0042] In one alternative embodiment of the bearing performance prediction system, an alarm module connected to the control module is also included. When the bearing performance exceeds a safety threshold, the control module controls the alarm module to issue an alarm.

[0043] The beneficial effects of this invention are as follows: This invention constructs a health index that reflects bearing performance degradation, and uses downtime processing and differential peak extraction to construct local and global degradation trends. Based on the characteristics of the degradation trends, a dual-scale trend prediction model composed of a local trend prediction model and a global trend prediction model is used to train and learn the local and global degradation trends, extracting the periodicity and monotonicity features of the time series, and expressing this monotonicity feature with peak values. Simultaneously, it predicts the real-time operating status history of the bearing and the global performance degradation history, achieving dual-scale dynamic trend prediction. Based on this, this invention can accurately predict bearing performance degradation trends, promptly detect potential safety hazards in production equipment and provide early warnings and alarms, assist in guiding the rational arrangement of maintenance cycles, improve production efficiency, and extend equipment service life.

[0044] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0045] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0046] Figure 1 This is a schematic diagram of the RMS full life cycle of the input shaft bearing of a roughing mill in the metallurgical industry.

[0047] Figure 2 This is a flowchart illustrating the present invention;

[0048] Figure 3 This is a schematic diagram of a local trend prediction process based on GRU.

[0049] Figure 4 This is a schematic diagram of the sliding window processing for local degradation trends;

[0050] Figure 5 This is a flowchart illustrating the global trend prediction method based on ARI MA. Detailed Implementation

[0051] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0052] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0053] Example 1

[0054] like Figure 2 As shown, this invention provides an embodiment of a dual-scale dynamic trend prediction method applied to metallurgical manufacturing equipment, comprising the following steps:

[0055] Vibration signals are collected during normal bearing operation.

[0056] When collecting vibration signals, multiple vibration acceleration sensors can be used simultaneously to collect vibration signals of critical bearings in metallurgical manufacturing equipment during normal operation. The vibration signal collected by the a-th sensor at time t is denoted as V. a t ={v1,v2,…v j …,v n}, where n is the number of sampling points, v j Let be the vibration signal collected by the a-th sensor at the j-th sampling point at time t.

[0057] Then, the time-domain features of the vibration signal are extracted.

[0058] In this embodiment, the root mean square value f of the time-domain features of the vibration signal collected by the vibration acceleration sensor is extracted. rms The root mean square (RMS) value reflects the energy level of the signal and is used to evaluate the vibration level; the larger the RMS value, the greater the vibration. rms The calculation formula is as follows:

[0059]

[0060] In actual testing, data preprocessing can also be performed, mainly including data integrity checks and outlier handling.

[0061] The data collected by the monitoring system also needs to be preprocessed to ensure its integrity and accuracy. This invention employs data preprocessing primarily to check data integrity and handle outliers.

[0062] Data integrity checks are performed on the collected vibration signals. If the length of the vibration signal N = sampling time Δt × sampling frequency Fs, then the data is intact.

[0063] Outlier handling is used to identify outliers in the collected vibration signals. If the signal value at a certain instant is greater than K times the average value, this point is considered an outlier, and the outlier is replaced with a neighboring value.

[0064] Then, based on the characteristics of the working environment and fixed manufacturing process cycle of rolling mills in the metallurgical industry, local and global degradation trends are constructed to express the real-time operating status and performance degradation status of the equipment, respectively. Details are as follows:

[0065] Construction of local degradation trends: Local degradation trends are constructed using the sensitive features of vibration signals from multiple operating cycles. In this embodiment, the shutdown data is first processed: after extracting the time-domain features of the vibration signals, a shutdown threshold F is used. S Vibration signals belonging to the bearing shutdown state are deleted from the vibration signals, i.e., the time domain characteristic does not satisfy f. i >F S The vibration signal was deleted, f i For t i The time-domain characteristics of the vibration signal at each working cycle are obtained. Then, windowing processing is performed: due to the strong seasonality and weak monotonicity of the local degradation trend, only m working cycles of data are needed to achieve high prediction accuracy when predicting the trend. Therefore, the vibration signal of each working cycle is windowed, retaining only the prediction time T. pre The temporal characteristics of the first m working cycles are used to construct the local degradation trend F. L :

[0066] F L =F[T pre -mT:T pre ], where T refers to the period. mT:T pre Indicates the prediction time T pre The first m work cycles.

[0067] Construct a local trend prediction model and utilize the local degradation trend F L The local trend prediction model is trained. Since local health indicators (i.e., data that expresses the real-time operating status of the bearing after removing downtime data) have weak monotonicity and strong periodicity, this embodiment preferably, but is not limited to, using a GRU model to construct the local trend prediction model. Specifically:

[0068] like Figure 3 As shown, first consider the local degradation trend F L Sample acquisition using sliding window processing: The local prediction model of this invention selects direct prediction, using N L Predict N data points P There are N data points, therefore a window with a length of N is required. L +N P The windows are fitted with sliding windows, such as... Figure 4 As shown, this facilitates model training.

[0069] Then, the samples were divided: after applying a sliding window to the local degradation trend data, the obtained samples were divided into training set and validation set according to the proportion, and the model was trained and the prediction effect of the model was verified.

[0070] Model Training and Prediction: The local prediction model based on the GRU model is trained using the training set, and the model's performance is validated using the validation set. If the model's prediction performance is poor, the model is retrained. Construction of the Global Degradation Trend: The peak values ​​of the time-domain sensitive features of all vibration signals are extracted, and the peak values ​​are smoothed to construct the global degradation trend.

[0071] In this embodiment, the step of extracting the peak values ​​of the time-domain features of all vibration signals is as follows:

[0072] First, assume that the time-domain characteristics of all vibration signals constitute an eigenvector F, F = {f1, f2, ..., f i ,…,f n}, where f i To represent at time t i The eigenvalues ​​of F. By differencing the eigenvector F, we obtain Fi. diff ={f d2 ,f d3 ,…,f di ,…,f dn}, where f di f i Extract the difference value of F. diff The coordinates of the points where the mean difference is 0 are obtained to obtain F. zeros ={x1,x2,…,x z ,…,x Z}, x z It refers to the coordinates of the z-th point with a difference value of 0.

[0073] Then, Fzeros are differentially processed to obtain F d-zerosTo determine whether the waveform containing Fzeros is horizontal, we can check if the coordinate difference between two adjacent points in Fzeros is 1. For example, if the coordinate difference between f2 and f3 is 1, then the two points are adjacent and the waveform between them is horizontal.

[0074] Then according to F d-zeros For F diff Make corrections to ensure that each horizontal waveform has only one peak value, and obtain the corrected F. d ′ iff ={f d ′2,f d ′3,…,f d ′ i ,…,f d ′ n}, where f d ′ i f di The correction value. During differential processing, the horizontal waveform is calculated as a string of 0s. In the next step of peak determination, the points in the horizontal waveform show the same behavior. In this embodiment, to ensure that each horizontal waveform has only one peak, it is necessary to use one value from the horizontal waveform to correct the other 0 values. For example, if the input feature vector is F = {0, 1, 1, 1, 4, 3}, the differential processing formula is diff = f i+1 -f i F is calculated after difference processing. diff ={1,0,0,3,-1}, the points of the horizontal waveform need to be corrected to obtain diff ={1,3,3,3,-1}.

[0075] If the waveform containing Fzeros is not horizontal, the value obtained from the data difference processing of the non-horizontal waveform will not affect the subsequent peak extraction, so no correction is needed and the next step of processing can be directly carried out.

[0076] Determine F d ′ iff Whether the characteristic value corresponding to each time step is a peak value. In this embodiment, it is assumed that: if f d ′ i >0 and f d ′ i+1 If < 0, then time t i The characteristic value is the peak value.

[0077] The difference formula is diff = f i+1 -f i For example, the input feature vector F = {0, 1, 1, 1, 4, 3} corresponds to time T = {1, 2, 3, 4, 5, 6}; after difference processing, F is calculated. diff={1,0,0,3,-1}, whose corresponding time series T ={2,3,4,5,6}, after point correction of the horizontal waveform, we obtain F. d ′ iff ={1,3,3,3,-1}, corresponding to time T ={2,3,4,5,6}; if we want to determine whether the eigenvalue corresponding to time t = 5 is the peak value, we need to determine whether diff5 = 3 > 0 and diff6 = -1 < 0, i.e., the peak value.

[0078] After extracting the peaks, a peak sequence is obtained and provided for the next step of periodic peak distance detection. In this embodiment, periodic peak distance detection is preferred, but not limited to, to ensure that there is only one peak within each peak period. Redundant peaks are removed, and finally, the peak value F of the entire feature vector F is obtained. P ={f p1 ,f p1 ,…,f pi ,…,f pn}, that is, to obtain the peak sequence containing the time-domain characteristic peaks of all vibration signals, where f pi Let time t i The peak value of the time-domain characteristics of the vibration signal.

[0079] The periodic peak distance detection is as follows: the bearing vibration signal is periodic under the influence of the manufacturing cycle. Therefore, it is determined whether the interval between two peaks is greater than or equal to the manufacturing cycle. If it is less than the manufacturing cycle, the redundant peaks are deleted in combination with the intervals of other adjacent peaks.

[0080] After obtaining the peak values ​​of the entire feature vector F, the peak values ​​of the time-domain features of all vibration signals (i.e., the peak sequence mentioned above) are smoothed to obtain the global degradation trend F. G .

[0081] Since the global degradation trend focuses more on the monotonicity of the degradation process, a quadratic exponential smoothing method is used to smooth the features in this embodiment. The formula is as follows:

[0082] s i =αf pi +(1-α)(s i-1 +Δs i-1 )

[0083] Δs i =β(s) i -s i-1 )+(1-β)Δs i-1

[0084] Where α and β are memory decay factors, and in this embodiment, α = β = 0.9 is preferred, but not limited to. i Let time t iThe peak value f of the time-domain characteristics of the vibration signal pi The smoothed value, Δs i This represents the trend.

[0085] After the above processing, the global degradation trend F is obtained. G This reflects the overall performance degradation state of the bearing. A global trend prediction model is constructed, and the global degradation trend is used to train the global trend prediction model.

[0086] Since global health indicators are characterized by strong monotonicity and no obvious periodicity, this embodiment preferably, but is not limited to, constructing a global trend prediction model based on the ARIMA model.

[0087] like Figure 5 As shown, first consider the global degradation trend F G The time series is then tested for stationarity. If the time series is not stationary, the difference method is used to perform a difference transformation on the global degradation trend to remove the trend (monotonicity) of the series, and then the stationarity test is performed again until the time series is stationary.

[0088] Then, check whether the time series of the global degradation trend is white noise. If not, construct an ARIMA model and determine the order of the ARIMA model based on the global degradation trend. If the time series is white noise, the time series analysis is meaningless, and stop constructing the model.

[0089] In this embodiment, Ljung-Box is preferably used, but not limited to, for white noise testing. LB The statistic is calculated using the following formula:

[0090]

[0091] In the formula, N is the number of time points in the time series input to the ARIMA model, and also the number of peaks in the peak sequence; The correlation coefficient is the result of the k-th lag in the sample. It follows a chi-square distribution with m degrees of freedom.

[0092] Next, the order of the ARIMA model is determined: The ARIMA model has three key parameters: p, autoregressive order d, and moving average order q. The difference order during the stationarity test is used as the parameter d of the ARIMA model. Using the Bayesian information criterion, the autoregressive order p and the moving average order q of the ARIMA model are determined according to the following formula;

[0093]

[0094] BIC(p0′,q0′)=minBIC(p,q),0≤p,q≤L

[0095] In the formula, The time series variance input to the ARIMA model is the peak series variance processed above. N is the number of time points in the time series input to the ARIMA model, which is also the number of peaks in the peak series. For example, assuming there are 100 peaks in the peak series, N is 100. L is the maximum order of the moving average term q in the ARIMA model. If p0′ and q0′ satisfy the above formula, then they are the optimal autoregressive order and the optimal moving average order. Finally, the global degradation trend is used to train the ARIMA model to obtain the global trend prediction model. If the model's prediction performance is poor, the model order is redefined, and the model is retrained.

[0096] In actual production, when predicting bearing performance, a trained local trend prediction model is used to infer the time-domain characteristics of the vibration signal in the current working cycle to predict the real-time operating status of the bearing; a trained global trend prediction model is used to infer the time-domain characteristics of all vibration signals to predict the global performance degradation state of the bearing.

[0097] Example 2

[0098] This application also provides a bearing performance prediction system, including a processing module, a storage module, a vibration acceleration sensor for collecting bearing vibration signals, and an alarm module. The vibration acceleration sensor is connected to the processing module and sends the collected vibration signals to the processing module. The processing module and the storage module are communicatively connected. The storage module stores at least one executable instruction, which causes the processing module to perform operations corresponding to the dual-scale dynamic trend prediction method applied to the metallurgical manufacturing industry as described in Embodiment 1, based on the vibration signals, to predict bearing performance. The alarm module is controlled by a control module; when the bearing performance exceeds a safety threshold, the control module controls the alarm module to issue an alarm.

[0099] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0100] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A dual-scale dynamic trend prediction method applied to the metallurgical manufacturing industry, characterized in that, Includes the following steps: Vibration signals during normal bearing operation are collected, and the time-domain features of the vibration signals are extracted. A local degradation trend is constructed using the time-domain characteristics of vibration signals from multiple working cycles. A local trend prediction model is then built. The local degradation trend is used to train the local trend prediction model. The local trend prediction model is then used to infer the time-domain characteristics of the vibration signal in the current working cycle to predict the real-time operating status of the bearing. Extracting the peak values ​​of the time-domain features of all vibration signals involves the following steps: The eigenvector F, composed of the time-domain features of all vibration signals, is differentially processed to obtain... ,extract The coordinates of the points where the mean difference is 0 are obtained. ; right Perform difference processing to obtain ,according to Determine if the coordinate difference between two adjacent points is 1 Is the waveform horizontal? according to right After correction The correction ensures that each horizontal waveform has only one peak. judge Whether the characteristic value corresponding to each time point is a peak value; By using periodic peak distance detection, redundant peaks are removed, ensuring that there is only one peak in each peak period, thus obtaining a peak sequence containing the time-domain characteristic peaks of all vibration signals. The peak values ​​of the time-domain characteristics of all vibration signals are smoothed based on the following formula to obtain the global degradation trend; ， ,in and For memory decay factor, For a moment Peak value of the time-domain characteristics of the vibration signal Smoothed values This is a trend item; The steps to construct a global trend prediction model are as follows: The time series of the global degradation trend is tested for stationarity. If it is not stationary, the global degradation trend is subjected to difference transformation and then the stationarity test is performed again. Examine whether the time series of the global degradation trend is white noise. If not, construct an ARIMA model and determine the order of the ARIMA model based on the global degradation trend. A global trend prediction model is obtained by training the ARIMA model using the global degradation trend. The global trend prediction model is trained using the global degradation trend, and the global trend prediction model is used to infer the time-domain characteristics of all vibration signals to predict the global performance degradation state of the bearing.

2. The dual-scale dynamic trend prediction method for metallurgical manufacturing industry according to claim 1, characterized in that, The steps to construct a local degradation trend are as follows: Remove vibration signals that belong to the bearing stop state from the vibration signals to obtain vibration signals for each working cycle; The local degradation trend is constructed by utilizing the temporal characteristics of vibration signals from several working cycles prior to the prediction time.

3. The dual-scale dynamic trend prediction method for metallurgical manufacturing industry according to claim 1, characterized in that, A GRU model is constructed, and a dataset is obtained by sliding windowing of the local degradation trend. The GRU model is then trained based on this dataset to obtain a local trend prediction model.

4. The dual-scale dynamic trend prediction method for metallurgical manufacturing industry according to claim 1, characterized in that, The parameters d of the ARIMA model are determined based on the difference order during the stationarity test. The autoregression order p and the moving average order q of the ARIMA model are determined according to the following formulas. ， In the formula, The variance of the time series input to the ARIMA model is N, the number of time points in the time series input to the ARIMA model is L, and the maximum order of the moving average term q in the ARIMA model is set. when and When the above formula is satisfied, , These are the optimal autoregression order and the optimal moving average order, respectively.

5. A bearing performance prediction system, characterized in that, It includes a processing module, a storage module, and a vibration acceleration sensor for acquiring bearing vibration signals; The vibration acceleration sensor is connected to the processing module and sends the vibration signal it has collected to the processing module. The processing module and the storage module are interconnected. The storage module is used to store at least one executable instruction. The executable instruction causes the processing module to perform the operation corresponding to the dual-scale dynamic trend prediction method applied to the metallurgical manufacturing industry as described in any one of claims 1-4 based on the vibration signal, and to predict the bearing performance.

6. The bearing performance prediction system according to claim 5, characterized in that, It also includes an alarm module connected to the control module. When the bearing performance exceeds the safety threshold, the control module controls the alarm module to issue an alarm.

Images