A Remote Printer Monitoring and Fault Prediction System Based on the Internet of Things

Abstract

The present invention discloses a remote printer monitoring and fault prediction system based on the Internet of Things, which relates to the technical fields of the Internet of Things and artificial intelligence. It includes a data acquisition module for real-time collecting the device status data of the remote printer and performing preprocessing; a data processing module for integrating the preprocessed device status data into a data matrix and a label vector; a model construction module for using a random forest model and combining with the Gumbel-Sigmoid activation function to output a final feature set; and a fault prediction module for using a bidirectional LSTM to construct a fault prediction model to output the predicted fault probability at the current time point and divide the fault interval data pool. By introducing the combination of the random forest algorithm and the Gumbel-Sigmoid activation function, the feature selection process can be automatically optimized, and the feature set crucial for fault prediction can be accurately screened out. The bidirectional LSTM model is used to capture the time series information, so that the fault prediction accuracy and real-time response ability are significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of the Internet of Things and artificial intelligence, especially a remote printer monitoring and fault prediction system based on the Internet of Things. Background Art

[0002] With the rapid development of Internet of Things technology, the remote monitoring and intelligent management of industrial equipment have become an important part of modern manufacturing. Especially in a large-scale distributed environment, remote monitoring technology enables real-time control of the operating status of equipment through data collection, processing, and analysis, thereby improving the system operation efficiency and reducing the operation and maintenance costs. As a typical representative of industrial production equipment, the reliability and efficiency of remote printers are crucial for printing quality and production continuity. In recent years, the combination of Internet of Things and big data analysis technology has enabled the status monitoring and fault prediction of remote printers to gradually develop from passive maintenance to predictive maintenance. By deploying sensors to obtain the operating status of equipment and combining machine learning algorithms to process and analyze data, enterprises can predict potential faults in advance and avoid production interruptions.

[0003] Currently, most technical solutions for remote printer monitoring rely on the collection of basic sensor data and single rule analysis methods. These solutions usually directly transmit the real-time collected device status data to the central processing system, which judges the device operating status through preset thresholds or fixed rules. This traditional method is prone to misjudgment in the face of complex operating scenarios and is difficult to accurately capture the potential trends of device faults. The traditional method lacks the ability to deeply mine high-dimensional feature data and it is difficult to achieve global fault risk prediction relying only on a single indicator. Summary of the Invention

[0004] In view of the above problems existing in the existing Internet-of-Things-based remote printer monitoring and fault prediction system, the present invention is proposed.

[0005] Therefore, the problems to be solved by the present invention are that the traditional monitoring system lacks flexibility in data processing, usually can only perform simple preprocessing on raw data, and fails to effectively integrate multi-dimensional device status information, which limits the model's understanding of complex data association characteristics, etc.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: A remote printer monitoring and fault prediction system based on the Internet of Things, which includes,

[0007] A data collection module, used to collect the device status data of the remote printer in real time and perform preprocessing;

[0008] A data processing module, used to integrate the preprocessed device status data into a data matrix and a label vector;

[0009] Model building module, which uses the random forest model and combines it with the Gumbel-Sigmoid activation function to output the final feature set;

[0010] The fault prediction module is used to build a fault prediction model using a bidirectional LSTM to output the predicted fault probability at the current time point and divide the fault interval data pool;

[0011] The classification processing module is used to screen important features and perform classification processing on data pools with different risk intervals through dynamic causal relationship analysis;

[0012] The data storage module is used to store and manage all data in a relational database.

[0013] As a preferred solution of the remote printing machine monitoring and fault prediction system based on the Internet of Things of the present invention, the real-time collection of the device status data of the remote printing machine includes:

[0014] Identify the remote printing presses that need to be monitored and predicted, install data collection devices in the remote printing presses, set the sampling frequency of the data collection devices, and collect the equipment status data of the printing presses in real time;

[0015] The device status data includes the paper status, vibration signal, ink remaining percentage, device temperature, network delay and number of tasks to be printed of the remote printer at each time point.

[0016] As a preferred solution of the remote printing machine monitoring and fault prediction system based on the Internet of Things of the present invention, the preprocessing includes:

[0017] Check the collected equipment status data and process missing values, detect and eliminate outliers, remove duplicate records, and standardize the equipment status data.

[0018] As a preferred solution of the remote printing machine monitoring and fault prediction system based on the Internet of Things of the present invention, wherein: the integration of the pre-processed device status data into a data matrix and a label vector includes:

[0019] The pre-processed device status data is sorted by timestamp and integrated into time series data. The vibration signal, ink remaining percentage, device temperature, network delay and number of tasks to be printed in the time series data are extracted as features and integrated into the feature vector corresponding to each time point. The feature vectors of all time points are divided into sliding windows with a window size of L. , the step length is S;

[0020] Count the number of occurrences of each paper state at each time point in the sliding window, and calculate the proportion of normal states in the sliding window as the paper state feature;

[0021] Calculate the mean of each feature dimension of the sliding window and the paper status feature as time series statistical features and combine them into a time series feature vector;

[0022] Use the mean and standard deviation to define the dynamic detection range of each feature dimension. Analyze all features within each sliding window and the dynamic detection range of the corresponding feature dimension. If any feature within the sliding window is larger than the dynamic detection range, the classification label of the sliding window is marked as a fault window. Otherwise, the classification label of the sliding window is marked as a normal window.

[0023] Construct a label vector Y, where each element corresponds to the classification label of a sliding window, the fault window is marked as 1, and the normal window is marked as 0;

[0024] The time series feature vectors of all sliding windows are combined into a global feature matrix, and the global feature matrix and label vector are combined into a data matrix. Each row represents a sliding window, and each column represents the time series feature vector and classification label of the sliding window.

[0025] As a preferred solution of the remote printing machine monitoring and fault prediction system based on the Internet of Things of the present invention, the random forest model is used in combination with the Gumbel-Sigmoid activation function to output the final feature set, which includes:

[0026] Divide the data matrix into training set and validation set, and initialize the parameters of the random forest model;

[0027] The parameters include the number of decision trees, the maximum tree depth, the number of features at each split, and the minimum number of sample splits;

[0028] Use the training set to fit the model, calculate the average Gini index drop value of all features in the training set in all trees, and calculate the importance value of each feature based on the average Gini index drop value ;

[0029] The importance value of each feature Normalize the features, set the importance threshold, traverse the training set and remove features whose importance values are lower than the importance threshold, sort the features whose importance values are greater than or equal to the importance threshold from high to low according to their importance values, and generate a candidate feature set. The candidate feature set includes the candidate features and the classification labels of the time points corresponding to the candidate features;

[0030] Retrain a new random forest model using the candidate feature set, evaluate the performance of the model on the validation set, calculate the AUC value, define the AUC threshold according to the actual scenario. If the AUC value of the validation set is greater than the AUC threshold, it is considered that the classification ability of the current candidate feature set meets the standard, and output the candidate feature set. Otherwise, reduce the importance threshold and regenerate the candidate feature set;

[0031] Initialize the initial selection probability, temperature parameter, maximum number of iterations, and the record of the number of unselected times;

[0032] Define the initial global weight of each candidate feature in the output candidate feature set according to the importance value of the features of the random forest and the initial local weight , the expression is:

[0033] , ,

[0034] where, represents the importance value of the i-th candidate feature in the candidate feature set, represents the maximum value of the importance values of all candidate features, represents the Pearson correlation coefficient between the i-th candidate feature in the candidate feature set and the classification label, represents the maximum value of the Pearson correlation coefficients between all candidate features and the corresponding classification labels;

[0035] For each candidate feature, generate a random variable, and based on the random variable, calculate the Gumbel noise of each candidate feature ;

[0036] Based on the initial global weight , the initial local weight and the Gumbel noise , use the Gumbel-Sigmoid activation function to generate the selection probability of each candidate feature in each iteration , the formula is:

[0037] ,

[0038] where, represents the global weight of the i-th candidate feature in the (h - 1)-th iteration, represents the local weight of the i-th candidate feature in the (h - 1)-th iteration, represents the temperature parameter in the (h - 1)-th iteration, which controls the distribution width of the probability;

[0039] According to the selection probability of each candidate feature in each iteration , set the probability threshold , determine the selection result of the candidate features. If , then select this candidate feature. If , then do not select this candidate feature;

[0040] In each iteration, dynamically adjust the global weight and local weight of the candidate features, and gradually reduce the temperature parameter through the temperature decay factor ;

[0041] The so-called dynamic adjustment of the global weight and local weight of the candidate features means that in each iteration, use the updated candidate feature set to train the random forest model, and calculate the difference between the AUC value of the candidate feature set in the previous round and the AUC value of the updated candidate feature set , and adjust the global weight of the selected candidate features through the difference;

[0042] If , indicating that the dynamically selected features have improved the performance, then increase the global weight of the selected candidate features. The formula is:

[0043] ,

[0044] Among them, represents the global learning rate;

[0045] If , indicating that the dynamically selected features have not improved the performance, then reduce the global weight of the selected candidate features. The formula is:

[0046] ,

[0047] Among them, represents the global learning rate;

[0048] Adjust the local weight of the unselected candidate features by introducing a decay factor, and decay according to the number of unselected times. The formula is:

[0049]

[0050] Among them, represents the decay rate control parameter, represents the cumulative number of unselected times from the 1st round to the hth round of iteration;

[0051] After dynamically adjusting the global weight and local weight of the candidate features, perform the next round of dynamic selection. When the maximum number of iterations is reached, terminate the iteration, output the finally selected features, and sort them in chronological order to integrate them into the final feature set.

[0052] As a preferred solution of the Internet of Things-based remote printing press monitoring and fault prediction system described in the present invention, where: the use of bidirectional LSTM to construct a fault prediction model to output the predicted fault probability at the current time point includes,

[0053] Dividing the obtained final feature set into a training set and a test set;

[0054] Using bidirectional LSTM to construct a fault prediction model, the feature input at each time point is respectively fed into two LSTM sub-networks, forward and backward, and the hidden states of the forward and backward are concatenated to form a complete hidden state;

[0055] Design the model structure of the fault prediction model, adopt the binary cross-entropy loss function to measure the error between the predicted fault probability output by the model and the classification label;

[0056] Adopt the Adam optimizer to update the weights, set the initial learning rate to 0.001, and dynamically adjust the learning rate according to the validation set loss;

[0057] Initialize the model parameters, input the training set for training. At each time point, the model calculates a probability of stopping prediction, maps it through a fully connected layer and uses the Sigmoid activation function. Set a stopping threshold. If the probability of stopping prediction is greater than the stopping threshold, it means that the current prediction result meets the standard, and the model stops calculating in advance. Otherwise, continue training;

[0058] By inputting the real-time device status data into the fault prediction model, output the predicted fault probability at the corresponding time point .

[0059] As a preferred solution of the Internet of Things-based remote printing press monitoring and fault prediction system described in the present invention, where: the division of the fault interval data pool includes,

[0060] According to the obtained predicted fault probability, set Divide the fault probability interval;

[0061] If , it means that the current remote printing press is in a normal state without fault signs, and it is determined as a low-risk interval;

[0062] If , it means that the current remote printing press has potential faults, and it is determined as a medium-risk interval;

[0063] If , it means that the current remote printing press has a high degree of abnormality, and it is determined as a high-risk interval;

[0064] Take the risk interval corresponding to each predicted failure probability as a data pool, which includes the vibration signal, ink remaining percentage, device temperature, network latency, and number of printing tasks to be printed of the device status data at the time point corresponding to each predicted failure probability.

[0065] As a preferred solution of the remote printer monitoring and fault prediction system based on the Internet of Things according to the present invention, wherein: for the data pools of different risk intervals, screening important features through dynamic causal relationship analysis includes,

[0066] Take each data in each data pool as a feature, analyze the causal weight of each feature in each data pool on the failure probability, and verify the influence of this feature on the status of the remote printer by dynamically adjusting the feature value;

[0067] In real-time analysis, by adjusting the feature value, observe the change of the predicted failure probability and update the causal weight;

[0068] Respectively set the causal weight thresholds for the low-risk interval and the high-risk interval. If the causal weight of a feature is greater than the causal weight threshold, extract this feature; otherwise, do not perform any operation.

[0069] As a preferred solution of the remote printer monitoring and fault prediction system based on the Internet of Things according to the present invention, wherein: the classification processing includes,

[0070] Perform classification processing according to the features extracted from each interval:

[0071] For the low-risk interval, regularly monitor the fluctuation of this feature and give a reminder;

[0072] For the high-risk interval, generate fault optimization suggestions according to this feature.

[0073] As a preferred solution of the remote printer monitoring and fault prediction system based on the Internet of Things according to the present invention, wherein: storing all data in a relational database and managing it includes,

[0074] Select a relational database to store and manage data and its analysis results, design the database table structure to store different types of data, set a regular backup task to back up all data in the database, perform permission management on database users, and encrypt and store the data.

[0075] The beneficial effects of the present invention are as follows: By introducing the combination of the random forest algorithm and the Gumbel-Sigmoid activation function, the feature selection process can be automatically optimized, and the feature set crucial for fault prediction can be accurately screened out. The bidirectional LSTM model is used to capture time series information, significantly improving the fault prediction accuracy and real-time response ability. By dividing the risk intervals and implementing feature importance analysis and causal relationship analysis for the data pools in different risk intervals, refined management and precise response are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0077] Figure 1 FIG. is a schematic structural diagram of a remote printer monitoring and fault prediction system based on the Internet of Things.

[0078] Figure 2 FIG. is a schematic flowchart of remote printer monitoring and fault prediction. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0079] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the specific embodiments of the present invention in detail with reference to the drawings of the specification.

[0080] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0081] Secondly, the so-called "one embodiment" or "embodiment" refers to a specific feature, structure, or characteristic that can be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it an individual or selectively exclusive embodiment with other embodiments.

[0082] Embodiment 1

[0083] Refer to Figure 1 and Figure 2 , which is the first embodiment of the present invention. This embodiment provides a remote printer monitoring and fault prediction system based on the Internet of Things. The remote printer monitoring and fault prediction system based on the Internet of Things includes:

[0084] S1. The data acquisition module is used to collect the device status data of the remote printer in real time and perform preprocessing;

[0085] Specifically, the real-time collection of the device status data of the remote printer includes

[0086] Determine the remote printers that need to be monitored and predicted, install acquisition devices in the remote printers, set the sampling frequency of the acquisition devices, and collect the device status data of the printers in real time;

[0087] The device status data includes the paper status, vibration signal, ink remaining percentage, device temperature, network latency, and the number of pending print tasks of the remote printer at each time point;

[0088] The paper status includes normal status, paper jam status, paper shortage status, and multiple paper status;

[0089] The normal status means that the paper position and quantity meet the expectations, and this status is marked as 0;

[0090] The paper jam status means that the sensor detects that the paper is stagnant or the position is abnormal, and this status is marked as 1;

[0091] The paper shortage status means that the paper inlet sensor does not detect paper, and this status is marked as 2;

[0092] The multiple paper status means that the paper outlet sensor detects abnormal paper thickness or reflection signal, and this status is marked as 3.

[0093] By collecting the device status data of the remote printer in real time, including paper status, vibration signal, ink remaining percentage, device temperature, network latency, and the number of pending print tasks, the operation status and potential risks of the device can be comprehensively reflected. The classification of the paper status is marked with digital codes, providing standardized support for quickly identifying and handling physical abnormalities in print tasks. This real-time monitoring method helps to accurately capture key data during device operation, improve the problem diagnosis efficiency, reduce the downtime caused by device failures, and thus improve production continuity and operation reliability.

[0094] Furthermore, the preprocessing includes

[0095] Checking the collected device status data and handling missing values, detecting and removing outliers, removing duplicate records, and performing standardization processing on the device status data.

[0096] By preprocessing the collected device status data, including checking and handling missing values, detecting and removing outliers, removing duplicate records, and performing normalization, the quality and consistency of the data can be significantly improved. Handling missing values and outliers can avoid biases caused by invalid data during model training and prediction. Removing duplicate records can reduce the impact of redundant data on computational efficiency, while normalization helps eliminate the dimensional differences between different features, making the data more suitable for input into the fault prediction model, thereby enhancing the robustness and prediction accuracy of the model.

[0097] S2. A data processing module for integrating the preprocessed device status data into a data matrix and a label vector;

[0098] Specifically, integrating the preprocessed device status data into a data matrix and a label vector includes

[0099] Sorting the preprocessed device status data by timestamp and integrating it into time series data, extracting vibration signals, ink remaining percentage, device temperature, network latency, and the number of printing tasks to be processed as features at each time point and integrating them into a feature vector corresponding to each time point, and dividing the feature vectors of all time points into sliding windows with a window size of L , with a step size of S, determining the values of L and S through cross-validation, and the sliding window The expression is:

[0100] ,

[0101] where represents the feature vector at time point t;

[0102] Count the number of occurrences of each paper status at each time point within the sliding window, and calculate the proportion of the normal status within the sliding window as the paper status feature;

[0103] Calculate the mean of each feature dimension of the sliding window and the paper status feature as the time series statistical feature and combine them into a time series feature vector;

[0104] Define the dynamic detection range of each feature dimension using the mean and standard deviation , where represents the mean of the feature, represents the standard deviation of the feature, represents the sensitivity coefficient for controlling the dynamic detection range;

[0105] Analyze all features within each sliding window and the dynamic detection range of the corresponding feature dimension. If any feature within the sliding window is greater than the dynamic detection range, mark the classification label of this sliding window as a fault window; otherwise, mark the classification label of this sliding window as a normal window;

[0106] Construct a label vector Y, where each element corresponds to the classification label of a sliding window, the fault window is marked as 1, and the normal window is marked as 0;

[0107] Combine the time series feature vectors of all sliding windows into a global feature matrix, and combine the global feature matrix and the label vector into a data matrix. Each row represents a sliding window, and each column represents the time series feature vector and classification label of the sliding window.

[0108] The above method integrates and extracts features from device status data through the sliding window technique, effectively capturing the dynamic characteristics during the operation of the remote printer. It constrains the feature dimensions using the statistical features within the window and the dynamic detection range, can accurately identify abnormal fluctuations of the device, improve the sensitivity and robustness of fault detection. Introducing the paper status features into the time series statistical features and combining the classification labels of the fault window and the normal window further enhances the comprehensive characterization of the device operation status. The finally generated data matrix not only retains the dynamic change information of multi-dimensional features, but also provides a complete input for model training and fault prediction in an efficient structural form, greatly improving the accuracy and efficiency of fault identification and prediction.

[0109] S3. A model construction module, which is used to output the final feature set using a random forest model combined with the Gumbel-Sigmoid activation function;

[0110] Specifically, using a random forest model combined with the Gumbel-Sigmoid activation function to output the final feature set includes:

[0111] Divide the data matrix into a training set and a validation set, and initialize the parameters of the random forest model;

[0112] The parameters include the number of decision trees, the maximum tree depth, the number of features at each split, and the minimum number of samples for splitting;

[0113] Use the training set for model fitting, calculate the average decrease value of the Gini index of all features in all trees in the training set, and calculate the importance value of each feature according to the average decrease value of the Gini index , the formula is:

[0114]

[0115] Among them, represents the number of decision trees in the random forest, represents the decrease amount of the Gini index of the j-th feature in the r-th decision tree in the training set;

[0116] The Gini index is an indicator for measuring the purity of a node. The smaller the Gini index, the purer the node (that is, the samples in the current node are more likely to belong to the same category).

[0117] Normalize the importance values of each feature, set an importance threshold, traverse the training set and remove the features whose importance values are lower than the importance threshold, sort the features whose importance values are greater than or equal to the importance threshold from high to low, and generate a candidate feature set. The candidate feature set includes candidate features and the classification labels of the corresponding time points of the candidate features;

[0118] Use the candidate feature set to retrain a new random forest model, evaluate the performance of the model on the validation set, calculate the AUC value, define an AUC threshold according to the actual scenario. If the AUC value of the validation set is greater than the AUC threshold, it is considered that the classification ability of the current candidate feature set meets the standard, and output the candidate feature set. Otherwise, reduce the importance threshold and regenerate the candidate feature set;

[0119] The AUC is the value of the area under the ROC curve, and its value range is [0, 1]. It represents the ability of the model to correctly classify positive and negative samples. The closer the AUC is to 1, the stronger the classification ability of the model. When AUC = 0.5, the performance of the model is no different from random guessing;

[0120] Initialize the initial selection probability, temperature parameter, maximum number of iterations, and the record of the number of unselected times;

[0121] According to the importance values of the features of the random forest, define the initial global weight and the initial local weight of each candidate feature in the output candidate feature set, and the expression is:

[0122] , ,

[0123] where represents the importance value of the i-th candidate feature in the candidate feature set, represents the maximum value of the importance values of all candidate features, represents the Pearson correlation coefficient between the i-th candidate feature in the candidate feature set and the classification label, represents the maximum value of the Pearson correlation coefficients between all candidate features and the corresponding classification labels;

[0124] The combination of the global weight (based on feature importance) and the local weight (based on the correlation between the feature and the label) can comprehensively evaluate the value of the feature:[[ID=K]]

[0125] The global weight reflects the overall contribution of the feature in the entire dataset; ​

[0126] The local weights further quantify the direct relationship between features and labels through the Pearson correlation coefficient;

[0127] This combined weight mechanism is more robust than simply relying on a single metric and can identify key features more accurately;

[0128] For each candidate feature, a random variable is generated, and based on the random variable, the Gumbel noise of each candidate feature is calculated , and the formula is:

[0129] ,

[0130] where, represents the random variable of the i-th candidate feature;

[0131] The temperature parameter mentioned refers to a key parameter in the Gumbel-Sigmoid activation function, which controls the smoothness of the feature selection probability;

[0132] Based on the initial global weights , the initial local weights and the Gumbel noise , the Gumbel-Sigmoid activation function is used to generate the selection probability of each candidate feature in each round of iteration , and the formula is:

[0133] ,

[0134] where, represents the global weight of the i-th candidate feature in the (h - 1)-th iteration, represents the local weight of the i-th candidate feature in the (h - 1)-th iteration, represents the temperature parameter in the (h - 1)-th iteration, which controls the distribution width of the probability;

[0135] After introducing the Gumbel-Sigmoid dynamic selection mechanism, each round of iteration can be dynamically adjusted according to the current model performance and the selection probability of features, avoiding the overly rigid limitations in traditional feature selection methods based on fixed thresholds;

[0136] The Gumbel-Sigmoid activation function can simulate the probability distribution, smooth the randomness of feature selection, and ensure the stability of the feature selection results;

[0137] According to the selection probability of each candidate feature in each round of iteration , a probability threshold is set to judge the selection result of the candidate feature. If , then select this candidate feature. If , then this candidate feature is not selected;

[0138] In each iteration, the global weight and local weight of the candidate features are dynamically adjusted, and the temperature parameter is gradually reduced through the temperature decay factor ;

[0139] The dynamic adjustment of the global weight and local weight of the candidate features means that in each iteration, the updated candidate feature set is used to train the random forest model, and the difference between the AUC value of the candidate feature set in the previous round and the AUC value of the updated candidate feature set is calculated , and the global weight of the selected candidate features is adjusted through the difference;

[0140] If , indicating that the dynamic feature selection has improved the performance, then the global weight of the selected candidate features is increased, and the formula is:

[0141] ,

[0142] where represents the global learning rate;

[0143] If , indicating that the dynamic feature selection has not improved the performance, then the global weight of the selected candidate features is decreased, and the formula is:

[0144] ,

[0145] where represents the global learning rate;

[0146] The local weight of the unselected candidate features is adjusted by introducing a decay factor, and the decay is based on the number of unselected times, and the formula is:

[0147]

[0148] where represents the decay rate control parameter, represents the cumulative number of unselected times from the 1st round to the hth round of iteration;

[0149] After dynamically adjusting the global weight and local weight of the candidate features, the next round of dynamic selection is performed. When the maximum number of iterations is reached, the iteration is terminated, and the finally selected features are output and sorted in chronological order to form the final feature set.

[0150] By combining the random forest model and the Gumbel-Sigmoid activation function, the present invention realizes the dynamic selection and optimization of candidate features, significantly improving the flexibility and accuracy of feature selection. The random forest model quantifies the importance of features by calculating the average decrease in the Gini index, and globally selects a candidate feature set that contributes more to fault prediction. During the dynamic selection process, the Gumbel-Sigmoid activation function further optimizes the feature selection result, balances randomness and stability, by integrating global weights and local weights, combining a dynamic adjustment mechanism and a temperature decay factor. This method avoids the limitations of traditional fixed-threshold feature selection methods, can adaptively adjust the feature selection strategy, and optimize the importance ranking of features in real time according to the model performance. The finally generated feature set is more robust and interpretable, providing high-quality input for the subsequent fault prediction model, and effectively improving the prediction ability and applicability of the model.

[0151] S4. A fault prediction module, which is used to construct a fault prediction model using bidirectional LSTM to output the predicted fault probability at the current time point and divide the fault interval data pool;

[0152] Specifically, constructing a fault prediction model using bidirectional LSTM to output the predicted fault probability at the current time point includes,

[0153] Dividing the obtained final feature set into a training set and a test set;

[0154] Constructing a fault prediction model using bidirectional LSTM. The feature input at each time point is separately fed into two LSTM sub-networks, the forward network and the backward network. The forward network processes the data at each time point in sequence from the start to the end of the time series, and the backward network processes the data in reverse from the last time point of the time series. At each time point, the hidden states of the forward and backward directions are concatenated to form a complete hidden state. This structure can capture the bidirectional dependencies in the time series;

[0155] Design the number of hidden units in the LSTM layer to be 60. After the output of the LSTM layer, connect a fully connected layer to perform feature transformation on the hidden state. The output dimension of the fully connected layer is 32, and the activation function is ReLU to ensure that the output value is non-negative and enhance the non-linear expression ability of the model. The fully connected layer is connected to the output layer of a single node, and the Sigmoid activation function is used to map the output of the fully connected layer to the interval [0,1], representing the predicted fault probability;

[0156] Adopt the binary cross-entropy loss function to measure the error between the predicted fault probability output by the model and the classification label;

[0157] Adopt the Adam optimizer to update the weights, set the initial learning rate to 0.001, and dynamically adjust the learning rate according to the validation set loss;

[0158] Initialize the model parameters, input the training set for training. At each time point, the model calculates the probability of stopping prediction, maps it through a fully connected layer and uses the Sigmoid activation function. Set the stopping threshold. If the probability of stopping prediction is greater than the stopping threshold, it means that the current prediction result meets the standard, and the model stops calculating in advance; otherwise, continue training.

[0159] By inputting the real-time device status data into the fault prediction model, the predicted fault probability at the corresponding time point is output. .

[0160] This bidirectional LSTM fault prediction model utilizes the bidirectional time series processing ability, combines the hidden states of the forward and backward LSTM networks to capture the bidirectional dependencies in the data, thereby comprehensively mining the time series characteristics of the device operating state. By designing the number of hidden units, fully connected layers and activation functions, the model has strong non-linear expression ability and accurate fault probability prediction ability. Using the binary cross-entropy loss function and the Adam optimizer with dynamic learning rate adjustment, the model can converge quickly and adapt to different data distributions, optimizing the prediction accuracy. By introducing the stopping prediction mechanism, the model can dynamically judge whether to terminate the calculation in advance according to the current prediction result, significantly improving the calculation efficiency. This design not only improves the accuracy and real-time performance of the prediction, but also effectively saves computing resources, providing efficient and reliable technical support for intelligent monitoring and fault warning.

[0161] Furthermore, the divided fault interval data pool includes

[0162] According to the obtained predicted fault probability, set Divide the fault probability intervals.

[0163] If , it means that the current remote printer status is normal, there are no fault signs, and it is determined as a low-risk interval;

[0164] If , it means that there are potential faults in the current remote printer, and it is determined as a medium-risk interval;

[0165] If , it means that there is a high degree of abnormality in the current remote printer, and it is determined as a high-risk interval;

[0166] Take the risk interval corresponding to each predicted fault probability as the data pool. The data pool includes the vibration signal, ink remaining percentage, device temperature, network delay and number of printing tasks to be printed of the device status data at the time point corresponding to each predicted fault probability.

[0167] By dividing the predicted failure probability into three intervals of low risk, medium risk, and high risk, and storing the device status data at relevant time points into different data pools according to the risk intervals corresponding to the failure probability, refined management of the device status can be achieved. This division method not only helps to accurately distinguish normal operation, potential failures, and highly abnormal states, but also provides targeted data support for subsequent feature importance analysis and dynamic causal relationship analysis, thereby significantly improving the efficiency and accuracy of failure prediction, and providing a scientific basis for risk intervention and optimization strategies.

[0168] S5. Classification processing module, which is used to screen important features through dynamic causal relationship analysis and perform classification processing for data pools in different risk intervals;

[0169] Specifically, screening important features through dynamic causal relationship analysis for data pools in different risk intervals includes

[0170] Regarding each data in each data pool as a feature, analyzing the causal weight of each feature in each data pool on the failure probability, and verifying the impact of this feature on the status of the remote printer by dynamically adjusting the feature value. The formula is:

[0171]

[0172] Where represents the causal weight of the feature on the prediction result, represents the change in the predicted failure probability after the extracted feature value is adjusted, represents the amplitude of the feature value adjustment;

[0173] In real-time analysis, by adjusting the feature value, observing the change in the predicted failure probability, and updating the causal weight;

[0174] Respectively set the causal weight thresholds for the low-risk interval and the high-risk interval. If the causal weight of the feature is greater than the causal weight threshold, extract this feature; otherwise, do not perform any operation.

[0175] Through dynamic causal relationship analysis, the actual impact of features in different risk intervals on the status of the remote printer can be accurately evaluated. By gradually adjusting the feature value, observing the change in the predicted failure probability in real time, and dynamically updating the causal weight, the contribution of each feature to the failure risk can be quantified. This method effectively solves the problems of static and limited feature importance in traditional analysis, realizes the dynamic association verification between features and failure prediction results, can provide a scientific basis for the precise optimization and failure prevention of equipment, and helps to improve the stability and reliability of the remote printer.

[0176] Furthermore, performing classification processing includes

[0177] Classify and process the features extracted according to each interval:

[0178] Low-risk interval, regularly monitor the fluctuations of this feature and give reminders;

[0179] High-risk interval, generate fault optimization suggestions based on this feature.

[0180] By classifying and processing the remote printer through interval-based causal analysis, precise status management and fault prevention and control can be achieved. The feature analysis in the low-risk interval helps to identify the key factors for maintaining the stable operation of the equipment, while the in-depth analysis in the high-risk interval focuses on the key fault features to generate targeted optimization plans, thereby effectively reducing the probability of faults and extending the service life of the equipment. This classification and processing method significantly improves the reliability and intelligence level of the equipment operation, and at the same time optimizes the operation and maintenance efficiency.

[0181] S6. Data storage module, used to store all data in a relational database and manage it;

[0182] Storing all data in a relational database and managing it includes

[0183] Selecting a relational database to store and manage data and its analysis results, designing the database table structure to store different types of data, setting regular backup tasks to back up all data in the database, performing permission management on database users, and encrypting the storage of data.

[0184] Adopting a relational database to store and manage data and its analysis results, through designing a reasonable database table structure, can efficiently organize and store different types of data. Setting regular backup tasks ensures the security and recoverability of data. Implementing permission management on database users can restrict unauthorized access and ensure the security of data use. At the same time, by encrypting the storage of sensitive information, it effectively prevents data leakage and tampering, thereby enhancing the security, reliability, and availability of data management and providing strong support for subsequent analysis and decision-making.

[0185] Embodiment 2

[0186] This is the second embodiment of the present invention. This embodiment provides the experimental steps of the remote printer monitoring and fault prediction system based on the Internet of Things, including:

[0187] The experimental steps include

[0188] Data acquisition module, data processing module, model construction module, and classification processing module;

[0189] I. In the data acquisition module, the real-time device status data of the remote printer is collected by setting the sampling frequency. As shown in Table 1, the collected device status data is checked and missing values are processed, outliers are detected and removed, duplicate records are removed, and the device status data is standardized;

[0190] The parameters of the sliding window are set, the window size is 5, and the window step size is 1;

[0191] Table 1 Device status data for each time window

[0192] Timestamp Vibration signal Ink remaining (%) Device temperature Network latency Number of printing tasks pending 2023-11-20 10:00:00 0.15 80 45 50 2 2023-11-20 10:00:01 0.16 79 45.2 48 2 2023-11-20 10:00:02 0.30 78 46 60 3 2023-11-20 10:00:03 0.12 77 47 55 1 2023-11-20 10:00:04 0.40 76 48.5 70 2

[0193] II. In the data processing module, statistical features are calculated for vibration signals, ink remaining, device temperature, network latency, and the number of printing tasks to be processed. The number of occurrences of each paper status within the sliding window is counted. As shown in Table 2, the proportion of the normal status is calculated as the paper status feature, and the paper status feature is obtained as , and the statistical features and paper status features are summarized;

[0194] Table 2 Collected paper status data

[0195] Timestamp Paper status 2023-11-20 10:00:00 0 2023-11-20 10:00:01 0 2023-11-20 10:00:02 3 2023-11-20 10:00:03 2 2023-11-20 10:00:04 0

[0196] The features of each sliding window are analyzed. For each feature dimension within each sliding window, the feature value is checked according to the dynamic detection range. If the sliding window is determined to be a fault window, the classification label is 1. The time series feature vectors of all sliding windows are combined into a global feature matrix, and the global feature matrix and the label vector are combined into a data matrix;

[0197] III. In the model construction module, the obtained data matrix is subjected to feature iteration using a random forest model combined with the Gumbel-Sigmoid activation function;

[0198] The data matrix is divided into a training set, and the training set is used for model fitting. As shown in Table 3;

[0199] Table 3 Training set

[0200] Sliding window sequence number Feature 1 Feature 2 Feature 3: ... <![CDATA[O1]]> 0.90 40.0 60.0 ...

[0201] The average decrease value of the Gini index of all features in all trees in the training set is calculated. Based on the average decrease value of the Gini index, the importance value MDG of each feature is calculated. The importance values of all features are normalized, and the Pearson correlation coefficient between the features and the classification label is calculated , as shown in Table 4:

[0202] Table 4 MDG and the Pearson correlation coefficient between the features and the classification label

[0203] #timg# (Normalized) C #timg# #timg# Feature 1 0.08 0.60 0.25 0.9 Feature 2 0.06 0.50 0.25 0.9 Feature 3 0.15 0.8 0.25 0.9

[0204] Calculate the initial global weight and local weight, and the formula is:

[0205] , ,

[0206] Feature 1:

[0207] , 0.67,

[0208] Feature 2:

[0209] , 0.56,

[0210] Feature 3:

[0211] , 0.89,

[0212] Set the temperature parameter to 0.1, and the Gumbel noises of Feature 1, Feature 2 and Feature 3 are -0.05, 0.01 and 0.01 respectively. Calculate the feature selection probability according to the global weight and local weight, and the formula is:

[0213] ,

[0214] Feature 1:

[0215] ,

[0216] Feature 2:

[0217] ,

[0218] Feature 3:

[0219] ,

[0220] According to the selection probability of each candidate feature in each round of iteration , set the probability threshold , and judge the selection result of the candidate feature. If , then select this candidate feature. If , then do not select this candidate feature;

[0221] The dynamic adjustment of the global weight and local weight of the candidate feature means that in each round of iteration, use the updated candidate feature set to train the random forest model, and calculate the difference between the AUC value of the candidate feature set in the previous round and the AUC value of the updated candidate feature set Adjust the global weights of the selected candidate features through difference adjustment;

[0222] Set the of Feature 1, Feature 2, and Feature 3 to 0.01, 0.02, -0.03 respectively, and the learning rate

[0223] If indicates that the dynamically selected features have improved performance, then increase the global weights of the selected candidate features. The formula is:

[0224] ,

[0225] where represents the global learning rate;

[0226] If indicates that the dynamically selected features have not improved performance, then decrease the global weights of the selected candidate features. The formula is:

[0227] ,

[0228] Feature 1:

[0229] 0.3201,

[0230] Feature 2:

[0231] ,

[0232] Feature 3:

[0233] ,

[0234] Adjust the local weights of the unselected candidate features by introducing a decay factor, decay according to the number of unselected times. After dynamically adjusting the global and local weights of the candidate features, perform the next round of dynamic selection. When the maximum number of iterations is reached, terminate the iteration, output the finally selected features, and sort them in chronological order to integrate into the final feature set.

[0235] IV. In the classification processing module, input the real-time device status data into the fault prediction model to output the predicted fault probability at the corresponding time point , set Map the fault probability at each time point to the normal interval, low, and high-risk intervals, as shown in Table 4;

[0236] Table 4 Real-time device status data and risk interval division

[0237] Time point Vibration signal Ink remaining percentage Temperature Network latency Number of tasks Predicted failure probability Risk interval <![CDATA[t1]]> 0.12 80% 45℃ 20 2 0.10 Normal interval <![CDATA[t2]]> 0.15 75% 46 25 3 0.20 Normal interval <![CDATA[t3]]> 0.35 60% 50 40 4 0.50 Low-risk interval <![CDATA[t4]]> 0.70 30% 60 70 5 0.85 High-risk interval <![CDATA[t5]]> 1.00 20% 65 100 6 0.95 High-risk interval

[0238] 4. Analysis of experimental results

[0239] 1. The experiment achieved risk classification based on the failure probability output by the fault prediction model. The experimental data showed that the present invention can accurately divide the risk intervals based on the failure probability. The division logic of each interval is clear, and the characteristic distribution is highly consistent with the risk level.

[0240] 2. The variation of features in different risk intervals shows that the distribution of each feature in the risk interval is significantly different, which verifies the rationality of feature selection. The present invention can effectively capture key features related to faults through dynamic feature extraction and sensitive feature analysis.

[0241] 3. Compared with traditional feature selection technology, this paper introduces the Gumbel-Sigmoid activation function to achieve dynamic feature optimization, which significantly improves the prediction accuracy, recall rate and precision of the model;

[0242] If the functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0243] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0244] More specific examples (non-exhaustive list) of computer-readable media include the following: electrical connections (electronic devices) having one or more wirings, portable computer disk cartridges (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber devices, and portable compact disc read-only memory (CDROM). Additionally, the computer-readable media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or otherwise processing as appropriate, and then storing it in a computer memory.

[0245] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having suitable combinational logic gate circuits, programmable gate arrays (PGA), field programmable gate arrays (FPGA), etc.

[0246] 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 preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A remote printing press monitoring and fault prediction system based on the Internet of Things, characterized by: include, Data acquisition module, used to collect the equipment status data of remote printing machines in real time and perform pre-processing; Data processing module, used to integrate the pre-processed device status data into data matrix and label vector; Model building module, which uses the random forest model and combines it with the Gumbel-Sigmoid activation function to output the final feature set; The fault prediction module is used to build a fault prediction model based on the final feature set using a bidirectional LSTM to output the predicted fault probability at the current time point and divide the fault interval data pool; The classification processing module is used to screen important features and perform classification processing on data pools with different risk intervals through dynamic causal relationship analysis; Data storage module, used to store all data in a relational database and manage it; The random forest model is combined with the Gumbel-Sigmoid activation function to output the final feature set, including: Divide the data matrix into training set and validation set, and initialize the parameters of the random forest model; The parameters include the number of decision trees, the maximum tree depth, the number of features at each split, and the minimum number of sample splits; Use the training set to fit the model, calculate the average Gini index drop value of all features in the training set in all trees, and calculate the importance value of each feature based on the average Gini index drop value ; The importance value of each feature Normalize the features, set the importance threshold, traverse the training set and remove features whose importance values are lower than the importance threshold, sort the features whose importance values are greater than or equal to the importance threshold from high to low according to their importance values, and generate a candidate feature set. The candidate feature set includes the candidate features and the classification labels of the time points corresponding to the candidate features; Retrain a new random forest model using the candidate feature set, evaluate the model's performance on the validation set, calculate the AUC value, and define the AUC threshold based on the actual scenario. If the AUC value of the validation set is greater than the AUC threshold, the classification ability of the current candidate feature set is considered to meet the standard and the candidate feature set is output. Otherwise, lower the importance threshold and regenerate the candidate feature set. Initialize the initial selection probability, temperature parameter, maximum number of iterations and number of unselected records; According to the importance value of the feature of the random forest, the initial global weight of each candidate feature in the output candidate feature set is defined and initial local weights , the expression is: ， ， in, represents the importance value of the i-th candidate feature in the candidate feature set, Represents the maximum importance value among all candidate features, represents the Pearson correlation coefficient between the i-th candidate feature in the candidate feature set and the classification label, Represents the maximum value of the Pearson correlation coefficient between all candidate features and the corresponding classification labels; For each candidate feature, generate a random variable, and based on the random variable, calculate the Gumbel noise of each candidate feature ; Based on the initial global weight , initial local weight and Gumbel noise , using the Gumbel-Sigmoid activation function to generate the selection probability of each candidate feature in each round of iteration , the formula is: ， in, represents the global weight of the i-th candidate feature in the h-1-th iteration, represents the local weight of the i-th candidate feature in the h-1-th iteration, represents the temperature parameter in the h-1th iteration, which controls the distribution width of the probability; According to the selection probability of each candidate feature in each iteration , set the probability threshold , judge the selection result of candidate features, if , then select the candidate feature if , then the candidate feature is not selected; In each round of iteration, the global weight and local weight of the candidate features are dynamically adjusted, and the temperature parameter is gradually reduced through the temperature attenuation factor. ; The dynamic adjustment of the global and local weights of the candidate features refers to using the updated candidate feature set to train the random forest model in each iteration, and calculating the difference between the AUC value of the candidate feature set in the previous round and the AUC value of the updated candidate feature set. , adjust the global weight of the selected candidate features by difference; like , indicating that dynamic feature selection improves performance, then the global weight of the selected candidate features is increased. The formula is: ， in, represents the global learning rate, represents the global weight of the i-th candidate feature in the h-th iteration; like , indicating that the dynamic selection feature does not improve the performance, then the global weight of the selected candidate feature is reduced. The formula is: ， The local weight of the unselected candidate features is adjusted by introducing a decay factor, which is decayed according to the number of times they are not selected. The formula is: ， in, represents the decay rate control parameter, Indicates the cumulative number of unselected times from the 1st to the hth round of iterations; After dynamically adjusting the global and local weights of the candidate features, the next round of dynamic selection is carried out. When the maximum number of iterations is reached, the iteration is terminated, the final selected features are output, and they are sorted in chronological order and integrated into the final feature set.

2. The remote printing press monitoring and fault prediction system based on the Internet of Things according to claim 1, characterized in that: The real-time collection of device status data of the remote printing machine includes: Identify the remote printing presses that need to be monitored and predicted, install a data acquisition device in the remote printing press, set the sampling frequency of the data acquisition device, and collect the equipment status data of the printing press in real time; The device status data includes the paper status, vibration signal, ink remaining percentage, device temperature, network delay and number of tasks to be printed of the remote printer at each time point.

3. The remote printing press monitoring and fault prediction system based on the Internet of Things according to claim 2, characterized in that: The pre-processing comprises: Check the collected equipment status data and process missing values, detect and eliminate outliers, remove duplicate records, and standardize the equipment status data.

4. The remote printing press monitoring and fault prediction system based on the Internet of Things according to claim 3, characterized in that: The integration of the pre-processed device status data into a data matrix and a label vector includes: The pre-processed device status data is sorted by timestamp and integrated into time series data. The vibration signal, ink remaining percentage, device temperature, network delay and number of tasks to be printed in the time series data are extracted as features and integrated into the feature vector corresponding to each time point. The feature vectors of all time points are divided into sliding windows with a window size of L. , the step length is S; Count the number of occurrences of each paper state at each time point in the sliding window, and calculate the proportion of normal states in the sliding window as the paper state feature; Calculate the mean of each feature dimension of the sliding window and the paper status feature as time series statistical features and combine them into a time series feature vector; Use the mean and standard deviation to define the dynamic detection range of each feature dimension. Analyze all features within each sliding window and the dynamic detection range of the corresponding feature dimension. If any feature within the sliding window is larger than the dynamic detection range, the classification label of the sliding window is marked as a fault window. Otherwise, the classification label of the sliding window is marked as a normal window. Construct a label vector Y, where each element corresponds to the classification label of a sliding window, the fault window is marked as 1, and the normal window is marked as 0; The time series feature vectors of all sliding windows are combined into a global feature matrix, and the global feature matrix and label vector are combined into a data matrix. Each row represents a sliding window, and each column represents the time series feature vector and classification label of the sliding window.

5. The remote printing press monitoring and fault prediction system based on the Internet of Things according to claim 4, characterized in that: The method of using bidirectional LSTM to construct a fault prediction model to output the predicted fault probability at the current time point includes: The final feature set is divided into training set and test set; A bidirectional LSTM is used to build a fault prediction model. The feature input at each time point is fed into the forward and backward LSTM sub-networks respectively. The forward and backward hidden states are concatenated to form a complete hidden state. Design the model structure of the fault prediction model and use the binary cross entropy loss function to measure the error between the predicted fault probability output by the model and the classification label; The Adam optimizer is used for weight update, the initial learning rate is set to 0.001, and the learning rate is dynamically adjusted according to the validation set loss; Initialize the model parameters and input the training set for training. At each time point, the model calculates a probability of stopping prediction, maps it through a fully connected layer and uses the Sigmoid activation function. Set a stopping threshold. If the probability of stopping prediction is greater than the stopping threshold, it means that the current prediction result meets the standard and the model stops calculating in advance. Otherwise, continue training. By inputting real-time equipment status data into the fault prediction model, the predicted failure probability at the corresponding time point is output. .

6. The remote printing press monitoring and fault prediction system based on the Internet of Things according to claim 5, characterized in that: The division of the fault interval data pool includes: According to the predicted failure probability, set Divide the failure probability interval; like , it means that the current remote printer status is normal, with no signs of failure, and is determined to be in the normal range; like , it means that the current remote printer has a potential fault and is judged to be in the low-risk range; like , it means that the current remote printing machine is highly abnormal and is determined to be in a high-risk range; The medium-risk interval and high-risk interval are used as data pools respectively. The data pools include the vibration signal, ink remaining percentage, device temperature, network delay and number of tasks to be printed of the equipment status data at the time point corresponding to each predicted failure probability.

7. The remote printing press monitoring and fault prediction system based on the Internet of Things according to claim 6, characterized in that: The data pool for different risk intervals is screened for important features through dynamic causal relationship analysis, including: Taking each data in each data pool as a feature, the causal weight of each feature in each data pool on the failure probability is analyzed, and the influence of the feature on the status of the remote printing machine is verified by dynamically adjusting the feature value; In real-time analysis, by adjusting the eigenvalues, we observe the changes in the predicted failure probability and update the causal weights; The causal weight thresholds for the low-risk interval and the high-risk interval are set respectively. If the causal weight of a feature is greater than the causal weight threshold, the feature is extracted; otherwise, no operation is performed.

8. The remote printing press monitoring and fault prediction system based on the Internet of Things according to claim 7, characterized in that: The classification process includes: Classification is performed based on the features extracted from each interval: In the low-risk range, regularly monitor the fluctuation of this feature and issue reminders; High-risk interval, based on which fault optimization suggestions are generated.

9. The remote printing press monitoring and fault prediction system based on the Internet of Things according to claim 8, characterized in that: The storing and managing of all data in a relational database includes: Choose a relational database to store and manage data and its analysis results, design the database table structure to store different types of data, set up regular backup tasks, back up all data in the database, manage database user permissions and encrypt data storage.

Images