Filling method for missing data of electric screw press based on gru and application and equipment thereof

Abstract

The present application belongs to the technical field of industrial equipment fault prediction, and discloses a filling method for missing data of an electric screw press based on GRU, application and equipment thereof, which comprises the following steps: (1) extracting continuous data in a striking cycle of the collected electric screw press, and then obtaining time series data of one striking cycle; (2) for the missing time series data in data collection, on the basis of the GRU-Attention network, position coding and the skip connection of the previous time series data are added to construct an interpolation model; the interpolation model respectively uses the time series data in the predetermined time window before the missing value for forward interpolation, uses the time series data in the predetermined time window after the missing value for backward interpolation, and processes the missing interval in the form of a sliding window to obtain a complete filling interval, and then completes the filling of the missing data. The present application solves the problem of online and rapid filling of missing data.

Description

Technical Field

[0001] This invention belongs to the technical field of industrial equipment fault prediction, and more specifically, relates to a method for filling missing data of an electric screw press based on GRU, and its application and equipment. Background Technology

[0002] As a crucial manufacturing equipment for metal forgings, the electric screw press has always held an irreplaceable position in industrial production. With the development of industrial manufacturing, in pursuit of high efficiency, high quality, and low cost, electric screw press systems have become increasingly large, with increased structural complexity and instability. Prolonged continuous operation leads to equipment deterioration, such as accelerated wear and poor lubrication. When this deterioration reaches a certain level, it can threaten personnel safety and cause significant damage to the equipment and production molds. Damaged equipment and molds are difficult to repair, maintenance costs are high, and equipment interruptions will affect the continuity of subsequent production, causing incalculable losses to enterprises. Therefore, maintaining the stable operation of the electric screw press is of paramount importance.

[0003] In production environments, the complex interrelationships of various components in electric screw presses and the uncertainties of the production process make equipment failure prediction more difficult, and data acquisition for the equipment often results in missing data. Traditional data imputation methods often fail to capture the complex patterns of equipment operation, while machine learning-based imputation methods such as LSTM and GRU-D, although providing good fits to missing samples, suffer from low computational efficiency, lack of adaptive adjustment, and long-term dependency issues. Furthermore, equipment operating states are typically complex and variable, and data gaps are unstable; imputation models need to be able to handle these situations, and the confidence level of the imputed data may not yet be incorporated. These factors can lead to misjudgments in equipment failure prediction.

[0004] Therefore, how to quickly fill in missing data under complex working conditions and further combine the confidence level of the data to give the current failure prediction results of the electric screw press is of great significance for the continuous health maintenance of the equipment and reducing the economic losses caused by equipment wear and tear and production stoppage. Summary of the Invention

[0005] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method and device for filling missing data of an electric screw press based on GRU, which solves the problem of online rapid filling of missing data.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for filling missing data in a GRU-based electric screw press is provided, the method comprising the following steps:

[0007] (1) Extract the continuous data within the impact cycle of the electric screw press, and then obtain the frequency components and amplitude spectrum through Fourier transform. Then, combine the impact energy, motor winding temperature and copper nut temperature to obtain the time series data of one impact cycle.

[0008] (2) To address the missing time series data in the data acquisition, a position encoding and a skip connection of the previous time series data are added to the GRU-Attention network to construct an interpolation model. The interpolation model uses the time series data within a predetermined time window before the missing value for forward interpolation and the time series data within a predetermined time window after the missing value for backward interpolation. The missing interval is processed in a sliding window manner to obtain a complete filling interval, thereby completing the missing data filling.

[0009] Furthermore, step (1) includes a step of integrating sensor data arranged on the components of the electric screw press.

[0010] Furthermore, multiple methods are employed for data collection, with different collection methods set for different data types. These methods include subscription collection, scheduled collection, and batch collection.

[0011] Furthermore, after the interpolation model completes the data filling of the missing interval, it will evaluate the confidence of the filled interval based on the distance and delay error of the filled data, and then perform weighted calculation and iterative update on the filled data of forward interpolation and backward interpolation.

[0012] Furthermore, the hidden state of the GRU within the interpolation model window is input into the self-attention layer after cosine position encoding. The weighted score output by the self-attention layer is then processed by max pooling and average pooling, along with the GRU hidden state at the current last time step, and input into the feedforward neural network. Finally, the feedforward neural network outputs the final predicted value, which is the filling data for the next time step.

[0013] Furthermore, during the imputation process, firstly, the imputation model captures the real data and imputed data within the imputation window, and uses the pre-trained imputation model to perform preliminary forward and backward imputation on the missing time-series data. Then, after imputing the missing intervals, the imputed data for those intervals is evaluated using the real data following the missing intervals, with the delay error on the real data used as the error for the missing intervals. Next, based on the distance between the missing data time step and the previous real data and the delay error, a confidence assessment is given for each imputed data point. The confidence scores are then used to weight the imputed data from the forward and backward imputations to obtain the imputation result. Finally, the relationship between the delay error and the error threshold is determined, and iterative imputation is selectively performed based on the determination result to update the imputed data.

[0014] This invention also provides a fault prediction method for an electric screw press based on GRU interpolation. The fault prediction method uses the missing data filling method for electric screw presses based on GRU as described above to fill in the missing data to obtain complete data. The confidence level of the complete data is used as the input of the prediction model. Based on the MLP neural network layer, combined with the self-attention layer that introduces confidence level, the prediction model is obtained. Then, the prediction model is used to predict the faults of the electric screw press.

[0015] Furthermore, the self-attention layer of the prediction model incorporates the confidence information of the time series data, and the corresponding formula is:

[0016]

[0017] in, The confidence level for forward padding. The confidence level for backfilling;

[0018] In the calculation of attention, the confidence score is normalized, and the confidence score of the real data is marked as 1 to obtain the confidence score of the filled data:

[0019]

[0020] Introduce confidence weighting into the attention score:

[0021]

[0022] Where C is the confidence vector, which is a vector with the same dimensions as K and all values ​​are 0. The tensor.

[0023] The present invention also provides a fault prediction system for an electric screw press based on GRU interpolation. The system includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it performs the fault prediction method for an electric screw press based on GRU interpolation as described above.

[0024] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when called and executed by a processor, cause the processor to implement the GRU-based method for filling missing data in an electric screw press or the GRU-based method for predicting faults in an electric screw press as described above.

[0025] In summary, compared with the prior art, the method and application of the GRU-based electric screw press data filling method and equipment provided by the present invention have the following advantages:

[0026] 1. This invention can process time series data with missing values. The proposed imputation model adds position encoding and skip connections to the previous time series data on the basis of the GRU-Attention network. It takes into account the distance relationship of time series data and the autocorrelation between missing values ​​and the previous time series data. The accuracy and efficiency of the imputation results are improved by weighting and iterating forward and backward imputation.

[0027] 2. The confidence level output by the interpolation model can be used as input for subsequent prediction models, which can reduce the impact of interpolated data with low confidence on the prediction results and improve the accuracy of the prediction results.

[0028] 3. This invention reduces the computational and communication load on edge devices through a flexible data acquisition scheme. Attached Figure Description

[0029] Figure 1 This is a flowchart of a fault prediction method for an electric screw press based on GRU interpolation provided by the present invention;

[0030] Figure 2 This is a schematic diagram illustrating the interaction of data collection by integrated devices;

[0031] Figure 3 This is a schematic diagram of the GRU network structure;

[0032] Figure 4 This is a structural diagram of the interpolation model;

[0033] Figure 5 This is a model structure diagram of the fault model prediction unit;

[0034] Figure 6 This is a system structure diagram of an embodiment of the fault prediction method. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0036] Please see Figure 1This invention provides a method for filling missing data in an electric screw press based on GRU. The filling method mainly includes the following steps:

[0037] Step 1: Integrate sensor data arranged on the components of the electric screw press.

[0038] Specifically, sensor data is integrated and deployed on key components of the electric screw press. Data such as slider displacement, motor current, motor voltage, motor winding temperature, impact energy, and copper nut temperature are integrated. A PLC is selected as the integration device, and the PLC provides access to the sensor data externally through OPCUA services.

[0039] Multiple data acquisition schemes are employed. Specifically, different acquisition schemes are set for different data types: 1) Subscription acquisition: For signal data, a subscription mechanism is used to monitor signal data; 2) Timed acquisition: For data with low change frequency, such as motor winding temperature, impact energy, and copper nut temperature, a timed reading method is used; 3) Batch acquisition: For data with high change frequency, such as slider displacement, motor current, and motor voltage, the data is stored in a local cache by the PLC and then read in batches.

[0040] Step 2: Extract the continuous data within the impact cycle of the electric screw press, and then obtain the frequency components and amplitude spectrum through Fourier transform. Combine this with the impact energy, motor winding temperature, and copper nut temperature to obtain the time series data of one impact cycle.

[0041] In this embodiment, continuous data within the impact cycle of the electric screw press is extracted. The data of slider displacement, motor current, and voltage are filtered and processed. The frequency components and amplitude spectrum are obtained through Fourier transform. Combined with standard data such as impact energy, motor winding temperature, and copper nut temperature, the time-series characteristic data of one impact cycle is obtained.

[0042] Step 3: To address the missing time-series data in the data acquisition, a position encoding and a skip connection to the previous time-series data are added to the GRU-Attention network to construct an interpolation model. This interpolation model performs forward interpolation using the time-series data within a predetermined time window before the missing value and backward interpolation using the time-series data within a predetermined time window after the missing value. The missing interval is processed in a sliding window manner to obtain the complete filling interval, thereby completing the missing data filling.

[0043] In this embodiment, after the interpolation model completes the data filling of the missing interval, the confidence level of the filled interval is evaluated based on the distance and delay error of the filled data. Based on this, the filled data of forward interpolation and the filled data of backward interpolation are weighted and iteratively updated.

[0044] Specifically, the hidden state of the GRU within the interpolation model window is input into the self-attention layer after cosine position encoding. The weighted score output of the self-attention layer is then processed by max pooling and average pooling, along with the GRU hidden state at the current last time step, and input into the feedforward neural network. Finally, the feedforward neural network outputs the final predicted value, which is the filling data for the next time step.

[0045] The imputation process is performed according to the following steps: 1) The imputation model captures the real data and imputed data within the imputation window, and uses the pre-trained imputation model to perform preliminary forward and backward imputation on the missing time series data respectively; 2) After the data imputation of the missing interval is completed, the error of the imputed data in the missing interval is evaluated using the real data after the missing interval, and the delay error on the real data is used as the error of the missing interval; 3) Based on the distance between the time step of the missing data and the previous real data and the delay error, a confidence assessment is given for each imputed data, and the confidence is used to perform weighted calculation on the imputed data of forward imputation and the imputed data of backward imputation to obtain the imputation result; 4) Determine the relationship between the delay error and the error threshold, and perform iterative imputation as appropriate to update the imputed data.

[0046] The initial interpolation process is as follows: the interpolation model window slides with the data flow. When new missing data is encountered, the known time-series data within the window is used as input, and the filler value for the missing data is used as output. The interpolation model is then used to fill in the missing time-series data. For example, based on the n time-series data (x1, x2, x3, ..., x...) preceding the missing interval... n Fill in the missing data intervals to obtain...

[0047]

[0048] The error evaluation process is as follows: when the system collects x n+m+1 When using real data, an interpolation model is used to calculate the filled data at that moment. Calculate delay error With relative error The relative error is taken as the error at each time step of the missing interval, i.e. Where t∈[n+1,n+m]. This process uses bidirectional interpolation, thus obtaining the forward interpolation value. and forward error Back interpolation value and backward error

[0049] The confidence level calculation and weighted evaluation process is as follows:

[0050]

[0051] Where t represents the time step, dir represents the padding direction, d represents the time distance between the padding data and the nearest true data in that direction, and err represents the error of that point in that direction. α and β are weight adjustment parameters, obtained through network search during training.

[0052] This formula combines the time distance between the filled data and the real data and the delay error of that interval as a confidence index, representing the credibility of the filled data at that time point in a certain direction.

[0053] like Figure 4 As shown, the imputation model needs to incorporate confidence levels to weight the imputed values, and the final imputed result is:

[0054]

[0055]

[0056]

[0057] The iterative filling process involves the interpolation model completing forward and backward weighted filling, and then determining whether the filling delay error in a certain direction of the missing interval is large (i.e., err). dir When the value is greater than θ (delay error threshold, default 0.2), the model can be imputed a second time, and the formula updated.

[0058]

[0059] Where γ represents the update rate, which defaults to 0.4.

[0060] The confidence level and the filled data are recalculated and updated using this method until the delay error no longer exceeds the threshold θ or the number of iterations is greater than the maximum number of iterations.

[0061] This invention also provides a fault prediction method for an electric screw press based on GRU interpolation. The fault prediction method uses the missing data filling method for electric screw presses based on GRU as described above to fill in the missing data to obtain complete data. The confidence level of the complete data is used as the input of the prediction model. Based on the MLP neural network layer, combined with the self-attention layer that introduces confidence level, the prediction model is obtained. Then, the prediction model is used to predict the faults of the electric screw press.

[0062] In this embodiment, the confidence information of the preprocessed padding data is used as the input to the prediction model. Based on the MLP neural network layer, a self-attention layer that introduces confidence is combined to obtain the prediction model.

[0063] Specifically, the self-attention layer of the prediction model incorporates the confidence information of the time-series data, and the corresponding formula is:

[0064]

[0065] in, The confidence level for forward padding. The confidence level for backfilling.

[0066] In the calculation of attention, the confidence score is normalized, and the confidence score of the real data is marked as 1 to obtain the confidence score of the filled data:

[0067]

[0068] Introduce confidence weighting into the attention score:

[0069]

[0070] Where C is the confidence vector, which is a vector with the same dimensions as K and all values ​​are 0. The tensor.

[0071] The self-attention layer uses two heads to capture different deep relationships between data.

[0072] Specifically, the MLP neural network layer of the prediction model takes the data processed by the attention layer as input to perform fault prediction and classification of the electric screw press. The current maximum possible prediction state is obtained through softmax processing, which represents the fault type that the electric screw press may fail in the next 2 hours.

[0073] The present invention also provides a fault prediction system for an electric screw press based on GRU interpolation. The system includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it performs the fault prediction method for an electric screw press based on GRU interpolation as described above.

[0074] A computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement either the GRU-based method for filling missing data in an electric screw press or the GRU-based method for predicting faults in an electric screw press, as described above.

[0075] In another embodiment, the fault prediction method is implemented using a fault prediction system for an electric screw press based on GRU interpolation. The fault prediction system includes a data acquisition module, a data processing module, a fault prediction processing module, and a fault prediction application module.

[0076] like Figure 6 As shown, the system of the present invention is divided into a perception layer, a network layer, a processing layer, and an application layer according to the Internet of Things structure, which demonstrates the working principle of the system and the hardware facilities on which it operates.

[0077] The data acquisition module includes a data integration unit and an edge communication unit.

[0078] The data integration unit is used to integrate sensor data from key components of the electric screw press. This system focuses on the J58ZK electric screw press, and the data integration unit collects real-time data from components prone to failure, such as the nut and slider in the transmission system, and the motor in the electrical control system. The specific data acquisition steps are as follows: A PLC is connected to the sensors on the equipment. The PLC performs basic order-of-magnitude processing on the data and provides services externally as OPCUA data nodes. For continuous data, the PLC alternately stores the impact cycle data using two structure arrays and uses an integer value as the clock signal for the impact; this integer is a globally unique ID.

[0079] like Figure 2 As shown, the edge communication unit provides three automated acquisition schemes for different data types: 1) Subscription acquisition: For signal data, a subscription mechanism is used to monitor signal data; 2) Timed acquisition: For slowly changing data, a timed reading method is used; 3) Batch acquisition: For rapidly changing data, batch acquisition and storage are performed, and then the acquisition unit is asynchronously notified to read the data in batches using a subscription mechanism. This module greatly reduces the computational and communication load of the edge device through flexible acquisition rules. In specific implementation, the edge communication unit obtains the PLC's connection information, data address, and acquisition rules from a local or remote server, and performs acquisition. The acquired real-time data is uploaded to the server for subsequent processing. The unit uses a global message ID and upload response to implement data retransmission functionality.

[0080] Meanwhile, for typical faults such as slider overtravel, slider jamming, slider slippage, unstable impact energy, excessive motor winding temperature, motor main inverter failure, power supply damage, etc., fault signals will be generated in the data integration unit, and the edge communication unit will subscribe to perceive the real faults and upload the fault signal information.

[0081] The data processing module includes a data identification unit, a periodic data extraction unit, and a missing data filling unit. The data identification unit distributes data to other units according to the type of uploaded data. Ordinary data is transmitted to the periodic data extraction unit for further processing, while fault signal data is transmitted to the model update unit to mark the prediction results.

[0082] The periodic data extraction unit is used to extract continuous data generated during the impact cycle of the electric screw press. This unit filters the collected slider displacement, velocity, motor current, and voltage, and obtains their frequency components and amplitude spectrum through Fourier transform. Combined with scalar data such as impact energy, motor winding temperature, output torque, and copper nut temperature, it obtains the time-series data for one impact cycle. This time-series data is then transmitted to the missing data filling unit and marked with a confidence level of 1 for the true data.

[0083] The missing data imputation unit is used to fill in data gaps caused by multiple factors during the production process. It utilizes an improved GRU-Attention network to perform forward and backward imputation of missing data using data before and after the missing interval. The confidence level of the imputation interval is assessed based on the imputation distance and delay error, and weighted calculations are performed accordingly, with iterative updates.

[0084] like Figure 3 As shown, the GRU model uses a reset gate and an update gate. The reset gate is used to forget data that is far from the time point where the missing value is located, and the update gate determines the weight of the state information of the previous time step. Therefore, the hidden layer state of the model represents the dynamic information encoding of the time step in the sequence data.

[0085] In the improved GRU-Attention model, the output sequence of the GRU hidden layer state is input into the self-attention layer after cosine position encoding. Cosine position encoding can further extract the periodic patterns of the state in the sequence. After data extraction by the self-attention layer, the weighted scores are subjected to max pooling and average pooling. Combined with the GRU hidden layer state at the current last time step, the three are output by the feedforward neural network to obtain the final predicted value. The output of this model is the padding data for the next time point with the same dimension as the input time series data.

[0086] The imputation process is performed according to the following steps: 1) GRU-Attention captures the real data and imputed data within the imputation window and uses the pre-trained model to initially imput the missing time series data; 2) When new real data is encountered, the imputation value of the missing interval is evaluated for error, and the delay error on the real data is used as the error of the missing interval; 3) Combining the distance and error between the current time step and the previous real data, a confidence evaluation is given for each imputation value, and the data obtained through bidirectional interpolation is weighted to obtain the imputation result; 4) According to the magnitude of the delay error, iterative imputation is performed to update the imputation value.

[0087] The initial imputation process involves the imputation window sliding along the data stream. When new missing data is encountered, known time-series data within the window are used as input, and the imputation value for the missing data is used as output. The GRU-Attention model is then used to impute the missing time-series data. For example, based on the n time-series data (x1, x2, x3, ..., x...) preceding the missing interval... n Fill in the missing data intervals to obtain...

[0088]

[0089] The error evaluation process is as follows: when the system collects x n+m+1 The actual data is calculated using a filling model to obtain the filling data at that moment. Calculate delay error With relative error This error is taken as the error for each time step in the missing interval, i.e. Where t∈[n+1,n+m]. This process uses bidirectional interpolation, thus obtaining the forward interpolation value. and forward error Back interpolation value and backward error

[0090] The confidence level calculation and weighted evaluation process is as follows:

[0091]

[0092] Where t represents the time step, dir represents the filling direction, d represents the time distance between the filling value and the nearest true value in that direction, err represents the error of the point in that direction, and α and β are weight adjustment parameters obtained through network search during training.

[0093] This formula combines the time distance between the filled data and the real data and the delay deviation of the interval as a confidence index, representing the credibility of the filled data at a certain time point in a certain direction.

[0094] The model needs to incorporate confidence levels to weight the imputed values, resulting in the following imputed outcome:

[0095]

[0096]

[0097]

[0098] The iterative filling process involves the model completing bidirectional weighted filling and determining if the filling delay error in a certain direction of the missing interval is large (i.e., err).dir When θ (delay error threshold, default 0.2) is greater than or equal to 0.2, the model can be imputed a second time, and the update formula is:

[0099]

[0100] Where γ represents the update rate, which defaults to 0.4.

[0101] The confidence level and filler value are recalculated and updated using this method until the delay error no longer exceeds the threshold θ or the number of iterations is greater than the maximum number of iterations.

[0102] The fault prediction processing module includes a fault prediction unit and a model update unit.

[0103] like Figure 5 As shown, the fault prediction unit refers to a self-attention layer combining confidence and an MLP neural network model. The self-attention layer combining confidence is used to discover deep correlations in time series data and introduces confidence weighting to the attention, reducing the impact of data points with low confidence on the model output.

[0104] Specifically, the self-attention layer introduces confidence information from the time-series data:

[0105]

[0106] in, Let be the confidence level of the forward padding at time t. The confidence level for backfilling.

[0107] In the calculation of attention, the confidence score is normalized, and the confidence score of the real data is marked as 1 to obtain the confidence score of the filled data:

[0108]

[0109] Introduce confidence weighting into the attention score:

[0110]

[0111] Where C is the confidence vector, which is a vector with the same dimensions as K and all values ​​are 0. The tensor.

[0112] The self-attention layer uses two heads to capture different deep relationships between data.

[0113] The MLP neural network layer takes the data processed by the attention mechanism as input to perform fault prediction and classification on the electric screw press. It obtains the current most likely prediction state through softmax processing, which represents the fault type that the model may encounter in the next 2 hours.

[0114] During implementation, offline datasets were used to train the network model during the model training phase. Types with fewer fault occurrences were resampled to balance the sample size, and the training and test sets were divided in a 7:3 ratio. Dropout was used during training to prevent overfitting, resulting in a pre-warmed model. In the online prediction phase, a sliding window was used to obtain padded data for fault prediction.

[0115] The model update unit is used to acquire equipment fault signals and manually input fault markers. The original input and marked output of the prediction record will be used for subsequent offline updates of the MLP prediction model. It accepts fault signals from the data identification unit and fault markers from the fault marking unit. Based on the fault timestamp matching the corresponding strike period, fault prediction records within 2 hours before the fault are extracted. The original input and marked output of this record are used as a new dataset to perform offline iterative updates to the model to adapt to changes in equipment operating modes.

[0116] The model-based fault prediction application module includes a fault record query unit, a fault record statistics unit, and a fault marking unit. This module implements interactive functionality via a web application.

[0117] The fault record query unit is used to query fault prediction results and provides filtering conditions based on dimensions such as equipment, time, and fault type, which facilitates viewing records and maintaining equipment.

[0118] The fault record statistics unit is used to provide statistical information on fault prediction results. It performs statistics on faults according to region, production line, equipment range, and fault type, and specifically provides statistical charts for year, month, and day to observe the characteristics of fault types.

[0119] The fault labeling unit is used for input of actual fault conditions. It provides a labeling service for real fault conditions. By manually labeling the accuracy of fault prediction results, it facilitates the model update unit to update the model.

[0120] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is 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 within the scope of protection of the present invention.

Claims

1. A method for filling missing data in an electric screw press based on GRU, characterized in that, The filling method includes the following steps: (1) Extract the continuous data within the impact cycle of the electric screw press, and then obtain the frequency components and amplitude spectrum through Fourier transform. Then, combine the impact energy, motor winding temperature and copper nut temperature to obtain the time series data of one impact cycle. (2) To address the missing time series data in the data collection, a position encoding and a skip connection of the previous time series data are added to the GRU-Attention network to construct an interpolation model. The interpolation model uses the time series data within a predetermined time window before the missing value for forward interpolation and the time series data within a predetermined time window after the missing value for backward interpolation. The missing interval is processed in a sliding window manner to obtain a complete filling interval, thereby completing the missing data filling. The hidden state of the GRU within the interpolation model window is input into the self-attention layer after cosine position encoding. The weighted score output by the self-attention layer is then processed by max pooling and average pooling, and the GRU hidden state at the last time step is input into the feedforward neural network. Finally, the feedforward neural network outputs the final predicted value, which is the filling data for the next time step. During imputation, firstly, the imputation model captures the real data and imputed data within the imputation window, and uses the pre-trained imputation model to perform preliminary forward and backward imputation on the missing time series data. Then, after imputing the missing intervals, the imputed data for those intervals is evaluated using the real data following the missing intervals, with the delay error on the real data used as the error for the missing intervals. Next, based on the distance between the missing data time step and the previous real data and the delay error, a confidence assessment is given for each imputed data point. The confidence scores are then used to weight the imputed data from forward and backward imputation to obtain the imputation result. Finally, the relationship between the delay error and the error threshold is determined, and iterative imputation is selectively performed based on the determination result to update the imputed data.

2. The method for filling missing data in an electric screw press based on GRU as described in claim 1, characterized in that: Step (1) is preceded by a step of integrating sensor data arranged on the components of the electric screw press.

3. The method for filling missing data in an electric screw press based on GRU as described in claim 2, characterized in that: Multiple methods are used for data collection, with different collection methods set for different data types. The collection methods include subscription collection, scheduled collection, and batch collection.

4. The method for filling missing data in an electric screw press based on GRU as described in claim 1, characterized in that: After the interpolation model completes the data filling of the missing interval, it will evaluate the confidence of the filled interval based on the distance and delay error of the filled data, and then perform weighted calculation and iterative update on the filled data of forward interpolation and backward interpolation.

5. A fault prediction method for an electric screw press based on GRU interpolation, characterized in that: The fault prediction method uses the missing data filling method for electric screw press based on GRU as described in any one of claims 1-4 to fill in the missing data to obtain complete data, and uses the confidence level of the complete data as the input of the prediction model. Based on the MLP neural network layer, combined with the self-attention layer that introduces confidence level, the prediction model is obtained, and then the prediction model is used to predict the fault of the electric screw press.

6. The fault prediction method for electric screw press based on GRU interpolation as described in claim 5, characterized in that: The self-attention layer of the prediction model incorporates confidence information from the time-series data, and the corresponding formula is: in, The confidence level for forward padding. The confidence level for backfilling; In the calculation of attention, the confidence score is normalized, and the confidence score of the real data is marked as 1 to obtain the confidence score of the filled data: Introduce confidence weighting into the attention score: Where C is the confidence vector, which is a vector with the same dimensions as K and all values ​​are 0. The tensor.

7. A fault prediction system for an electric screw press based on GRU interpolation, characterized in that: The system includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it performs the fault prediction method for an electric screw press based on GRU interpolation as described in any one of claims 5-6.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the GRU-based method for filling missing data in an electric screw press as described in any one of claims 1-4, or the GRU-based method for predicting faults in an electric screw press as described in any one of claims 5-6.

Images