A Deep Learning-Based Method for Recovering Missing Data from Monitoring Sensors

Abstract

The application discloses a kind of based on deep learning's monitoring sensor missing data recovery method, it is related to ocean platform health monitoring technical field, including: the strain data of acquisition sensor, constructs data set and data missing judgment strategy, data set is preprocessed and is divided into short-term and long-term data missing training set and test set;Bi-branch data processing model based on bidirectional recurrent neural network and GRU unit is constructed, two kinds of training set are input to train the model and obtain sensor data recovery model;Input test set obtains the missing data of prediction, according to the error of predicted value and true value the model is evaluated and optimized to obtain optimized model;Sensor data is input into the optimized sensor data recovery model, judges missing data type, and outputs the missing data of target sensor;The application can adaptively handle long-term and short-term data missing problem, ensure that missing data can be efficiently and accurately recovered under various data missing scenarios.

Description

Technical Field

[0001] This invention relates to the field of marine platform health monitoring technology, and in particular to a method for recovering missing data from monitoring sensors based on deep learning. Background Technology

[0002] Offshore platforms operate under complex and extreme sea conditions for extended periods, making them susceptible to structural failures and fatigue damage. Therefore, health monitoring of offshore platforms is crucial for status awareness, operational support, and safety decision-making. However, health monitoring and status assessment heavily rely on sensor data, which is frequently lost due to various factors such as malfunctions, power terminal issues, and signal interference. The loss of sensor data, especially at critical locations, can severely impact subsequent structural health assessments and significantly degrade the performance of the health monitoring system.

[0003] Structural health monitoring (SHM) of offshore platforms relies heavily on sensor data. In practical applications, offshore platforms operate for extended periods in harsh sea conditions. Therefore, sensors are prone to signal loss due to prolonged exposure to complex wave loads, aging, or power instability. Data loss directly prevents the SHM system from accurately assessing the health status of various structural components, severely impacting subsequent structural health evaluations. Missing data can be categorized into long-term fixed missing data and short-term random missing data. The former is typically caused by sensor damage, while the latter is often due to power failures or signal interference. Both types of missing data are frequent and equally important, but their recovery methods differ significantly. Existing methods mostly utilize sensor redundancy to recover long-term fixed missing data, neglecting to address and research the following issues:

[0004] (1) Insufficient utilization of sensor time correlation information: There is an important intrinsic relationship between the missing sensor signal and its own historical data (time correlation) and the historical data of other sensors (spatial correlation). However, existing data recovery methods often fail to fully consider these correlations, resulting in insufficient utilization of effective information and thus insufficient accuracy of the reconstructed signal.

[0005] (2) Insufficient accuracy in recovering short-term random data missing data: Short-term data missing data recovery is a challenging problem. Existing methods mostly rely on long-term historical data or other sensor data, ignoring the time dependence of the target sensor. Short-term missing data recovery should fully consider the dependence of the current data on its previous and subsequent time points, but most methods have failed to deeply analyze these short-term time correlations, resulting in poor recovery performance. This is especially true in high-frequency monitoring or rapid response application scenarios. The few data recovery methods that consider time correlation have not conducted in-depth research on short-term missing data recovery and have failed to optimize for the characteristics of short-term data missing data.

[0006] (3) Different methods and models are needed for different types of missing data: Existing methods lack the ability to adaptively handle the two types of missing data. They often require training multiple models for different types of missing data, resulting in high computational and operational complexity and difficulty in dealing with the variable missing data in complex environments. Therefore, although many methods have shown certain effects in their respective applicable scenarios, they have not been able to effectively solve the problem of simultaneously handling long-term and short-term missing data, which limits their widespread application in practical engineering. Summary of the Invention

[0007] This invention provides a deep learning-based method for recovering missing data from monitoring sensors. This method overcomes the technical problems of existing technologies, which mostly recover long-term fixed missing data during actual prediction, do not make sufficient use of sensor temporal correlation information, have insufficient accuracy in recovering short-term random missing data from sensors, and cannot adopt different methods and models for different types of missing data.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] A deep learning-based method for recovering missing data from monitoring sensors includes:

[0010] S1: Collect strain data from the sensors, construct a sensor dataset and a sensor data missing judgment strategy, and preprocess the dataset. The preprocessed dataset is divided into training and test sets for short-term and long-term data missing types according to the sensor data missing judgment strategy. The sensor dataset includes the non-failure data of the target sensor itself and the non-failure data of other sensors.

[0011] S2: A bidirectional recurrent neural network and GRU units are introduced to obtain the temporal relationship of the sensors, and a fully connected layer is introduced to obtain the spatial relationship between the sensor data. A two-branch data processing model based on the bidirectional recurrent neural network and GRU units is constructed. The two-branch data processing model is used to determine different data missing types according to the sensor data missing type judgment strategy, and select the corresponding branch to process the input data.

[0012] S3: Input the training sets for the two types of missing data into the dual-branch data processing model for training to obtain the sensor data recovery model; the sensor data recovery model is used to recover missing data using the corresponding sub-model according to different types of missing data.

[0013] S4: Input the test set into the sensor data recovery model to obtain the predicted missing data. Evaluate and optimize the sensor data recovery model based on the error between the predicted value and the true value to obtain the optimized sensor data recovery model.

[0014] S5: Input the non-failure data of the target sensor itself and the non-failure data of other sensors into the optimized sensor data recovery model, determine the missing data type, and output the missing data of the target sensor.

[0015] Furthermore, the dual-branch data processing model includes:

[0016] The dual-branch data processing model includes an input layer, a missing data type discriminator, a network call module, a short-term missing data recovery model, a long-term missing data recovery model, and an output layer.

[0017] The missing data type discriminator is used to determine whether the missing data type is short-term or long-term based on the sensor data missing judgment strategy, and inputs the judgment result to the network call module to call the corresponding branch for training.

[0018] The network call module is used to call the short-term missing data recovery model or the long-term missing data recovery model to process the input data based on the judgment result of the missing data type discriminator.

[0019] The short-term missing data recovery model includes m fully connected layers, 1 GRU layer, and r Bi-GRU layer connected in sequence, used to recover short-term missing data;

[0020] The long-term missing data recovery model comprises m fully connected layers and n GRU layers connected in sequence, used to recover long-term missing data.

[0021] Furthermore, the training sets for the two types of missing data are respectively input into the dual-branch data processing model for training to obtain the sensor data recovery model, including:

[0022] S31. Randomly determine a set of hyperparameters, namely m, n and r, and randomly initialize the weights and biases of each network layer in the dual-branch data processing model under these parameters.

[0023] S32. Input the short-term missing data training set into the input layer and pass it to the missing data type discriminator for judgment. If it is a short-term missing data type, the judgment result is input into the network call module, and proceed to S33 to call the short-term missing data recovery model for training. If it is a long-term missing data type, the judgment result is input into the network call module, and proceed to S34 to call the long-term missing data recovery model for training.

[0024] S33. Input the short-term missing data training set into m fully connected layers, and perform dimensionality reduction and feature extraction on the input data through multi-layer linear and non-linear transformations to capture the spatial correlation between different sensors.

[0025] The features extracted through m fully connected layers are input into a single GRU layer to obtain the short-term temporal dependencies in the input features.

[0026] Define the time acquisition range for the past time period and the time acquisition range for the future time period. Input the features processed by one GRU layer into r Bi-GRU layers to obtain the context information of the feature sequence within the time acquisition range of the past time period and the time acquisition range of the undetermined time period.

[0027] The features extracted by the Bi-GRU layer are integrated and fused through the output layer to predict the missing data at the time point corresponding to the current data and use it as the missing data at the first time point.

[0028] Add the missing data from the first time point to the training set, repeat the training process, predict the missing data from the current time point, and use it as the missing data for the second time point.

[0029] The missing data corresponding to the predicted time point of the current data is continuously added to the training set for training, until the missing data of the (x-1)th time point is added to the training set for training. The missing data corresponding to the current data is then predicted and used as the missing data of the xth time point; x represents the time point at the end of the missing data.

[0030] S34. Input the long-term missing data training set into m fully connected layers, and perform dimensionality reduction and feature extraction on the input data through multi-layer linear and non-linear transformations to capture the spatial correlation between different sensors.

[0031] The features extracted through m fully connected layers are input into n GRU layers to obtain the long-term temporal dependencies in the input features.

[0032] The features extracted by n GRU layers are integrated and fused through the output layer to predict the missing data at the time point corresponding to the current data and use it as the missing data at the first time point.

[0033] Add the missing data from the first time point to the training set, repeat the training process, predict the missing data from the current time point, and use it as the missing data for the second time point.

[0034] The missing data corresponding to the predicted time point of the current data is continuously added to the training set for training, until the missing data of the (y-1)th time point is added to the training set for training. The missing data corresponding to the current data is then predicted and used as the missing data of the yth time point; y represents the time point of the endpoint of the missing data.

[0035] S35. Set the loss function and optimizer: Set MSE as the loss function and use the Adam optimizer for optimization.

[0036] S36. Calculate the loss function between the predicted data and the real data, calculate the gradient based on the loss function, and use the Adam optimizer to update the weights and biases of the model based on the calculated gradient.

[0037] S37. Set the numerical range of hyperparameters, and adjust the number of fully connected layers, GRU layers and Bi-GRU layers of the model according to the set numerical range based on the grid search method to obtain the sensor data recovery model.

[0038] Furthermore, the sensor data missing determination strategy includes:

[0039] Determine the start and end times of the missing data periods in the training set, and calculate the duration of the missing data periods.

[0040] Determine the length of the available data period after the missing data. The available data period is the length of time between the end time of the missing data period and the start time of the next missing data period.

[0041] Set a time threshold for missing data and a time threshold for available data. When the duration of the missing data period is less than the time threshold for missing data and the duration of the available data period is greater than the time threshold for available data, the missing data of the target sensor is classified as short-term missing data.

[0042] When the duration of the missing data period is greater than the missing data time threshold, or the duration of the available data period is less than the available data time threshold, the missing data of the target sensor is classified as long-term missing data.

[0043] Furthermore, the test set is input into the sensor data recovery model to obtain the predicted missing data. The sensor data recovery model is then evaluated and optimized based on the error between the predicted and actual values ​​to obtain an optimized sensor data recovery model, including:

[0044] The test set is input into the sensor data recovery model, and the evaluation index of the predicted value and the true value is calculated. That is, the root mean square error and the mean absolute error of the predicted value and the true value are calculated. It is determined whether the two evaluation indexes meet the preset standards. If either index fails to meet the standard, the amount of training data is increased and re-input into the sensor data recovery model for a new round of training until all evaluation indexes meet the preset standards, and the optimized sensor data recovery model is obtained.

[0045] Beneficial Effects: This invention provides a deep learning-based method for recovering missing data from monitoring sensors. The missing data is categorized into short-term and long-term missing data. Based on the characteristics of bidirectional recurrent neural networks (BRNNs) capturing the temporal relationships between sensor data and GRU units capturing the spatial relationships, a dual-branch data processing model based on BRNNs and GRU units is constructed. By simultaneously utilizing the correlation between temporal and spatial data, the inherent relationships between data can be effectively captured, significantly improving the accuracy of data recovery. For short-term missing data, the constructed model effectively improves the accuracy of short-term recovery, enabling the model to more accurately predict missing values ​​when handling short-term missing data. Different branches of sub-models are constructed for different types of missing data, enabling adaptive handling of both long-term and short-term missing data problems, ensuring efficient and accurate recovery of missing data in various data loss scenarios. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 A flowchart of a method for recovering missing data from a monitoring sensor based on deep learning, provided by the present invention;

[0048] Figure 2 This is a structural diagram of the sensor data recovery model designed for this invention;

[0049] Figure 3 A structural diagram of the short-term data loss recovery model designed for this invention;

[0050] Figure 4 This is a structural diagram of the long-term missing data recovery model designed for this invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] This embodiment provides a method for recovering missing data from monitoring sensors based on deep learning, such as... Figure 1As shown, it includes:

[0053] S1: Collect strain data from the sensors, construct a sensor dataset and a sensor data missing judgment strategy, and preprocess the dataset. The preprocessed dataset is divided into training and test sets for short-term and long-term data missing types according to the sensor data missing judgment strategy. The sensor dataset includes the non-failure data of the target sensor itself and the non-failure data of other sensors.

[0054] S2: A bidirectional recurrent neural network and GRU units are introduced to obtain the temporal relationship of the sensors, and a fully connected layer is introduced to obtain the spatial relationship between the sensor data. A two-branch data processing model based on the bidirectional recurrent neural network and GRU units is constructed. The two-branch data processing model is used to determine different data missing types according to the sensor data missing type judgment strategy, and select the corresponding branch to process the input data.

[0055] S3: Input the training sets for the two types of missing data into the dual-branch data processing model for training to obtain the sensor data recovery model; the sensor data recovery model is used to recover missing data using the corresponding sub-model according to different types of missing data.

[0056] S4: Input the test set into the sensor data recovery model to obtain the predicted missing data. Evaluate and optimize the sensor data recovery model based on the error between the predicted value and the true value to obtain the optimized sensor data recovery model.

[0057] S5: Input the target sensor's own data and data from other sensors into the optimized sensor data recovery model, determine the missing data type, and output the missing data of the target sensor.

[0058] Specifically, firstly, strain data from sensors is collected to construct a sensor dataset and a sensor data missing judgment strategy. The dataset is then preprocessed, and the preprocessed dataset is divided into training and testing sets based on the sensor data missing judgment strategy, which distinguishes between short-term and long-term data missing types. By differentiating the duration of missing data, different matching and recovery algorithms can be more scientifically developed during the model building stage, resource allocation can be optimized, and the overall system efficiency and reliability can be improved.

[0059] Secondly, a bidirectional recurrent neural network and GRU unit are introduced to obtain the temporal relationship of the sensor, and a fully connected layer is introduced to obtain the spatial relationship between the sensor data. A two-branch data processing model based on bidirectional recurrent neural network and GRU unit is constructed. The introduction of bidirectional recurrent neural network and GRU unit can effectively process time series data, mine the temporal correlation between sensor signals, significantly improve its robustness and accuracy when recovering missing data with temporal correlation, and enable the model to support the reconstruction of long-term fixed and short-term random missing data.

[0060] Next, the training sets for the two types of missing data are respectively input into the dual-branch data processing model for training to obtain the sensor data recovery model. The sensor data recovery model is used to recover missing data using corresponding sub-models according to different types of missing data. The test set is input into the sensor data recovery model to obtain the predicted missing data. The sensor data recovery model is evaluated and optimized based on the error between the predicted value and the true value to obtain the optimized sensor data recovery model. This model can select different sub-models to process the input data according to the type of missing sensor data, so that the output predicted data is more consistent with the actual results.

[0061] Finally, the target sensor's own data and data from other sensors are input into the optimized sensor data recovery model to determine the type of missing data and output the missing data of the target sensor. The optimized sensor data recovery model not only relies on the historical data of a single sensor, but also makes full use of the historical data of other sensors and their interrelationships, which can more accurately recover missing data. Especially when the data is missing for a long time or there is noise interference, it can significantly improve the accuracy and stability of data recovery.

[0062] In a specific embodiment, the following scheme involves collecting strain data from sensors, constructing a sensor dataset and a sensor data missing detection strategy, and preprocessing the dataset. The preprocessed dataset is then divided into training and test sets for short-term and long-term data missing types based on the sensor data missing detection strategy:

[0063] Randomly select a normal sensor as the target sensor for data recovery, treat undamaged sensors as other sensors, collect the non-missing data of the target sensor and the non-missing data of other sensors, combine all other sensor data together and label them as "data_intact", label the target sensor data as "data_missing", store the data in the required database in dictionary format, arrange them in chronological order, perform mean-variance normalization on the data in the database to avoid data overload and the impact of different dimensions on model training, and divide the dataset into training set and test set in a 7:3 ratio;

[0064] The sensor data missing detection strategy includes:

[0065] Determine the start and end times of the missing data periods in the training set, and calculate the duration of the missing data periods.

[0066] Determine the length of the available data period after the missing data, which is the length of time between the end time of the available data period and the start time of the next missing data period.

[0067] Set a time threshold for missing data and a time threshold for available data. When the duration of the missing data period is less than the time threshold for missing data and the duration of the available data period is greater than the time threshold for available data, the missing data of the target sensor is classified as short-term missing data.

[0068] When the duration of the missing data period is greater than the missing data time threshold or the duration of the available data period is less than the available data time threshold, the missing data of the target sensor is classified as long-term missing data.

[0069] The training set is divided into a short-term missing data training set and a long-term missing data training set based on the types of data required for short-term and long-term missing data:

[0070] The training set for short-term missing data mainly consists of the target sensor's own historical and future data, supplemented by non-missing data from sensors located close to the target sensor. The scope of data collection is determined based on the actual collection time and duration of the missing data.

[0071] The training set with long-term missing data mainly consists of data from other sensors that are not missing, supplemented by historical data from periods similar to the time period when the target sensor's data is missing. The scope of data collection is determined by the actual collection time and the length of time the missing data is missing.

[0072] In this solution, a sensor data missing judgment strategy is constructed, which divides missing data into long-term data and short-term data. By distinguishing the duration of missing data, different matching and recovery algorithms can be more scientifically applied during the model building stage, resource allocation can be optimized, and the overall system efficiency and reliability can be improved.

[0073] In a specific embodiment, a scheme is implemented by introducing a bidirectional recurrent neural network to obtain the temporal relationship between sensor data and a GRU unit to obtain the spatial relationship between sensor data, thus constructing a two-branch data processing model based on the bidirectional recurrent neural network and the GRU unit:

[0074] like Figure 2As shown, the dual-branch data processing model includes an input layer, a missing data type discriminator, a network call module, a short-term missing data recovery model, a long-term missing data recovery model, and an output layer.

[0075] The missing data type discriminator is used to determine whether the missing data type is short-term or long-term based on the sensor data missing judgment strategy, and inputs the judgment result to the network call module to call the corresponding branch for training.

[0076] The network call module is used to call the short-term missing data recovery model or the long-term missing data recovery model to process the input data based on the judgment result of the missing data type discriminator.

[0077] like Figure 3 As shown, the short-term missing data recovery model includes m fully connected layers, 1 GRU layer, and r Bi-GRU layer connected in sequence, which are used to recover short-term missing data.

[0078] like Figure 4 As shown, the long-term missing data recovery model includes m fully connected layers and n GRU layers connected in sequence, which are used to recover long-term missing data.

[0079] This scheme introduces a Bidirectional Recurrent Neural Network (Bi-RNN) module and a Gated Recurrent Unit (GRU) neural network unit. The GRU network is the core component of the data recovery model, responsible for mining temporal and spatial correlations from large amounts of time-series data. Through its unique gating mechanism, the GRU can effectively process time-series data and mine the temporal correlations between sensor signals. Specifically, the GRU network selectively retains key information valuable for reconstruction and discards irrelevant redundant parts and useless noise information by controlling the flow of information and the forgetting mechanism, retaining only the signals useful for data recovery. This not only avoids the gradient explosion or gradient vanishing problems common in traditional recurrent neural networks (RNNs), but also effectively suppresses noise interference and reduces the negative impact of noise on data processing.

[0080] Furthermore, to fully utilize the spatiotemporal correlations between sensor data, a bidirectional recurrent neural network (Bi-RNN) is introduced into the model. The structure of the Bi-RNN allows the model to utilize both "past" and "future" information from time-series data simultaneously. This means that not only is the data at the current moment useful for prediction, but the model can also learn valuable information about missing data from future points in time, thus providing a more comprehensive perspective on the data recovery process. This fusion of bidirectional information enables the model to fully consider and utilize time-series data from different times, significantly improving its robustness and accuracy in recovering time-correlated missing data, and allowing the model to support the reconstruction of both long-term fixed and short-term random missing data.

[0081] In a specific embodiment, the scheme for training the sensor data recovery model by inputting the training sets of the two types of missing data into the dual-branch data processing model respectively is as follows:

[0082] S31. Randomly determine a set of hyperparameters, namely m, n and r, and randomly initialize the weights and biases of each network layer in the dual-branch data processing model under these parameters.

[0083] S32. Input the training data set into the missing data type discriminator through the input layer for judgment. If it is a short-term missing data type, input the judgment result into the network call module and enter S33 to call the short-term missing data recovery model for training. If it is a long-term missing data type, input the judgment result into the network call module and enter S34 to call the long-term missing data recovery model for training.

[0084] S33. Input the short-term missing data training set into m fully connected layers, and perform dimensionality reduction and feature extraction on the input data through multi-layer linear and non-linear transformations to capture the spatial correlation between different sensors.

[0085] The features extracted through m fully connected layers are input into a single GRU layer to obtain the short-term temporal dependencies in the input features.

[0086] Define the time acquisition range for the past time period and the time acquisition range for the future time period. Input the features processed by one GRU layer into r Bi-GRU layers to obtain the context information of the features within the time acquisition range of the past time period and the time acquisition range of the future time period.

[0087] This solution does not limit the time collection range of the past time period or the time collection range of the future time period. Those skilled in the art can select specific values ​​according to actual needs, such as obtaining information within 5 seconds or information within 10 seconds.

[0088] The features extracted by the Bi-GRU layer are integrated and fused through the output layer to predict the missing data at the time point corresponding to the current data and use it as the missing data at the first time point.

[0089] Add the missing data from the first time point to the training set, repeat the training process, predict the missing data from the current time point, and use it as the missing data for the second time point.

[0090] The missing data corresponding to the predicted time point of the current data is continuously added to the training set for training, until the missing data of the (x-1)th time point is added to the training set for training. The missing data corresponding to the current data is then predicted and used as the missing data of the xth time point; x represents the time point at the end of the missing data.

[0091] S34. Input the long-term missing data training set into m fully connected layers, and perform dimensionality reduction and feature extraction on the input data through multi-layer linear and non-linear transformations to capture the spatial correlation between different sensors.

[0092] The features extracted through m fully connected layers are input into n GRU layers to obtain the long-term temporal dependencies in the input features.

[0093] The features extracted by n GRU layers are integrated and fused through the output layer to predict the missing data at the time point corresponding to the current data and use it as the missing data at the first time point.

[0094] Add the missing data from the first time point to the training set, repeat the training process, predict the missing data from the current time point, and use it as the missing data for the second time point.

[0095] The missing data corresponding to the predicted time point of the current data is continuously added to the training set for training, until the missing data of the (y-1)th time point is added to the training set for training. The missing data corresponding to the current data is then predicted and used as the missing data of the yth time point; y represents the time point of the endpoint of the missing data.

[0096] S35. Set the loss function and optimizer: Set MSE as the loss function and use the Adam optimizer for optimization.

[0097] S36. Calculate the loss function between the predicted data and the real data, calculate the gradient based on the loss function, and use the Adam optimizer to update the weights and biases of the model based on the calculated gradient.

[0098] S37. Set the numerical range of hyperparameters, and adjust the number of fully connected layers, GRU layers and Bi-GRU layers of the model according to the set numerical range based on the grid search method to obtain the sensor data recovery model.

[0099] In this scheme, the sensor data recovery model includes short-term missing data recovery models and long-term missing data recovery models, respectively. The short-term missing data training set consists of normal signals from the target sensor and other sensors, and the sensors most relevant to the target sensor or closest to it are selected, with a limited number of them. Therefore, it is more necessary to learn the temporal relationship before and after the missing data. A Bi-GRU layer is used to fully learn the temporal relationship between the data and obtain the past and future contextual information in the features to obtain more accurate prediction data. The long-term missing data training set mainly consists of data from other sensors. Therefore, it is more necessary to learn the spatial relationship between the missing data and other data. If a Bi-GRU layer is used alone, it cannot provide high recovery accuracy. Therefore, a GRU layer and a linear layer are used together to learn the spatial relationship between the data to obtain more accurate results for long-term missing data.

[0100] GRU can capture long-term and short-term time dependencies in data. Time dependency is also known as time dependence. The term time dependence comes from the "Temporal Dependency" of recurrent neural networks, which refers to the mutual influence between different time steps in sequence data, that is, the past state will affect the future state.

[0101] In this invention, time dependence refers to the temporal characteristics of sensor monitoring data, that is, the data at the current time step can be inferred from past data. Based on the time span of the dependency relationship, time dependence can be divided into long-term dependence and short-term dependence. The time span is determined according to the length of the actual missing data.

[0102] Long-term dependency refers to the impact of early sensor data on future sensor data; short-term dependency refers to the impact of sensor data at adjacent time steps on current sensor data; early sensor data refers to the historical data of the sensors that need to be collected by the staff based on the missing data of the sensors, which is the total length of the training data of the input data;

[0103] In GRU, long-term dependencies are mainly controlled by the update gate, which determines the extent to which historical information is retained in the current state; while short-term dependencies are controlled by the reset gate, which determines the degree of integration between the current input and the past state. Through the gating mechanism, the missing data recovery model can more effectively capture the temporal dependencies in long-term series, thereby improving the accuracy of missing data recovery.

[0104] The features passed in the model are arranged in chronological order. During model training, only one missing data point is output at each collection time. This missing data point is then added to the training set to predict the missing data point for the next collection time.

[0105] In a specific embodiment, the test set is input into the sensor data recovery model to obtain the predicted missing data. The sensor data recovery model is then evaluated and optimized based on the error between the predicted and actual values. The optimized sensor data recovery model is obtained as follows:

[0106] The test set is input into the sensor data recovery model, and the evaluation index of the predicted value and the true value is calculated. That is, the root mean square error and the mean absolute error of the predicted value and the true value are calculated. It is determined whether the two evaluation indexes meet the preset standards. If either index fails to meet the standard, the amount of training data is increased and re-input into the sensor data recovery model for a new round of training until all evaluation indexes meet the preset standards, and the optimized sensor data recovery model is obtained.

[0107] In a specific embodiment, the scheme of inputting the non-failure data of the target sensor itself and the non-failure data of other sensors into the optimized sensor data recovery model, determining the missing data type, and outputting the missing data of the target sensor is as follows:

[0108] Based on the actual missing data, the required sensor data is obtained. The data from the target sensor itself and other sensors are input into the optimized sensor data recovery model. If the missing data is determined to be short-term data, the data is processed through a fully connected layer, a GRU layer, and a Bi-GRU layer to recover the short-term missing data. If the missing data is determined to be long-term data, the input data is processed through a fully connected layer and a GRU layer to recover the long-term missing data. The result can complete the missing data with high accuracy.

[0109] The optimized sensor data recovery model not only relies on historical data from a single sensor, but also makes full use of historical data from other sensors and their interrelationships, enabling more accurate recovery of missing data. Especially when the data is missing for a long time or there is noise interference, it can significantly improve the accuracy and stability of data recovery.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1.A method for recovering missing data of a monitoring sensor based on deep learning, characterized in that, The method comprises the following steps: S1: collecting strain data of a sensor, constructing a sensor data set and a sensor data missing judgment strategy, and preprocessing the data set, and dividing the preprocessed data set into a training set and a test set of short-term data missing type and long-term data missing type according to the sensor data missing judgment strategy; the sensor data set comprises non-failed data of a target sensor and non-failed data of other sensors; S2: introducing a bidirectional recurrent neural network and a GRU unit to obtain the time relationship of the sensor, introducing a full connection layer to obtain the spatial relationship between the sensor data, and constructing a double-branch data processing model based on the bidirectional recurrent neural network and the GRU unit; the double-branch data processing model is used to judge different data missing types according to the sensor data missing judgment strategy, and select the corresponding branch to process the input data; The double-branch data processing model comprises an input layer, a missing data type discriminator, a network calling module, a short-term missing data recovery model, a long-term missing data recovery model and an output layer; The missing data type discriminator is used to judge whether the data missing type is a short-term data missing type or a long-term data missing type according to the sensor data missing judgment strategy, and input the judgment result into the network calling module to call the corresponding branch for training; The network calling module is used to call the short-term missing data recovery model or the long-term missing data recovery model to process the input data according to the judgment result of the missing data type discriminator; The short-term missing data recovery model comprises m layers of full connection layers, 1 layer of GRU layers and r layers of Bi-GRU layers connected in sequence, and is used for recovering short-term missing data; The long-term missing data recovery model comprises m layers of full connection layers and n layers of GRU layers connected in sequence, and is used for recovering long-term missing data; S3: inputting the training sets of the two data missing types into the double-branch data processing model for training to obtain a sensor data recovery model; the sensor data recovery model is used to recover missing data using the corresponding sub-model according to different data missing types; S4: inputting the test set into the sensor data recovery model to obtain predicted missing data, evaluating and optimizing the sensor data recovery model according to the error value of the predicted value and the true value, and obtaining an optimized sensor data recovery model; S5: inputting the non-failed data of the target sensor and the non-failed data of other sensors into the optimized sensor data recovery model, judging the missing data type, and outputting the missing data of the target sensor. 2.The method of claim 1, wherein, Inputting the training sets of the two data missing types into the double-branch data processing model for training to obtain a sensor data recovery model, comprising: S31, randomly determine a group of hyperparameters, namely m, n and r, and randomly initialize the weights and biases of each network layer in the double-branch data processing model under the parameters; S32, input the data training set to the missing data type discriminator through the input layer for judgment, if it is a short-term data missing type, input the judgment result to the network calling module, enter S33 to call the short-term missing data recovery model for training, if it is a long-term data missing type, input the judgment result to the network calling module, enter S34 to call the long-term missing data recovery model for training; S33, input the short-term missing data training set to m full connection layers, reduce the dimension and extract the features of the input data through multi-layer linear and nonlinear transformation, and capture the spatial correlation between different sensors; Input the features extracted through the m full connection layers to a GRU layer, and obtain the short-term time dependence in the input features; Set the time collection range of the past time period and the time collection range of the future time period, input the features processed by the GRU layer to r Bi-GRU layers, and obtain the context information in the time collection range of the past time period and the time collection range of the future time period in the features; Integrate and fuse the features extracted through the Bi-GRU layer through the output layer, predict the missing data corresponding to the time point of the current data as the missing data of the first time point; Add the missing data of the first time point to the data training set, repeat the training process, predict the missing data corresponding to the time point of the current data as the missing data of the second time point; Continuously add the predicted missing data corresponding to the time point of the current data to the data training set for training, until the missing data of the x-1 time point is added to the data training set for training, the missing data corresponding to the time point of the current data is predicted as the missing data of the x time point; x represents the time point of the missing data end position; S34, input the long-term missing data training set to m full connection layers, reduce the dimension and extract the features of the input data through multi-layer linear and nonlinear transformation, and capture the spatial correlation between different sensors; Input the features extracted through the m full connection layers to n GRU layers, and obtain the long-term time dependence in the input features; Integrate and fuse the features extracted through the n GRU layers through the output layer, predict the missing data corresponding to the time point of the current data as the missing data of the first time point; Add the missing data of the first time point to the data training set, repeat the training process, predict the missing data corresponding to the time point of the current data as the missing data of the second time point; Continuously add the predicted missing data corresponding to the time point of the current data to the data training set for training, until the missing data of the x-1 time point is added to the data training set for training, the missing data corresponding to the time point of the current data is predicted as the missing data of the x time point; x represents the time point of the missing data end position; S35, set the loss function and the optimizer: set MSE as the loss function, and use Adam optimizer for optimization; S36, calculate the loss function of the predicted data and the real data, calculate the gradient according to the loss function, and update the weight and bias of the model according to the calculated gradient using the Adam optimizer; S37, set the numerical range of the hyperparameters, adjust the number of layers of the full connection layer, GRU layer and Bi-GRU layer of the model according to the set numerical range based on the grid search method, and obtain the sensor data recovery model. 3.The method of claim 1, wherein, The sensor data missing judgment strategy comprises: determining the start time point and the end time point of the missing data time period in the training set, and calculating the time length of the missing data time period; determine the time length of the available data time period after the missing data, that is, the time length from the end time point of the missing data time period to the start time point of the next missing data time period; set the time threshold of the missing data and the time threshold of the available data, when the time length of the missing data time period is less than the time threshold of the missing data, and the time length of the available data time period is greater than the threshold of the available data, the missing data of the target sensor is short-term data missing type; when the time length of the missing data time period is greater than the time threshold of the missing data or the time length of the available data time period is less than the time threshold of the available data, the missing data of the target sensor is long-term data missing type. 4.The method of claim 1, wherein input the test set into the sensor data recovery model, input the test set into the sensor data recovery model, obtain the predicted missing data, evaluate and optimize the sensor data recovery model according to the error value of the predicted value and the real value, and obtain the optimized sensor data recovery model, comprising: input the test set into the sensor data recovery model, calculate the evaluation index of the predicted value and the real value, that is, calculate the root mean square error value and the average absolute error value of the predicted value and the real value, judge whether the two evaluation indexes meet the preset standard, if any index fails to meet the standard, increase the training data amount and input the sensor data recovery model for a new round of training, until the evaluation indexes meet the preset standard, and obtain the optimized sensor data recovery model.

Images