An automatic early warning method for monitoring IoT sensing devices based on deep learning

Abstract

The present invention provides an automatic early warning method for monitoring IoT sensing devices based on deep learning. The method comprises: collecting sensor data generated by IoT sensing devices, and realizing data collection through sensor arrays, IoT interfaces, and cloud storage; utilizing deep neural networks to process and extract data features, and establishing a device status recognition model by learning data patterns of normal and abnormal device operation; monitoring device operation status in real time and predicting possible abnormal conditions based on the model, thereby improving prediction accuracy; automatically triggering an early warning mechanism and taking timely maintenance measures when the device operates abnormally or reaches early warning conditions; and receiving operator feedback and new data in real time, continuously optimizing model performance, and realizing self-learning. The present invention improves the accuracy of device failure prediction and significantly reduces the failure rate and maintenance costs of the device. This advantage ensures that the device can operate more stably during operation and reduces the economic losses caused by downtime due to failures.

Description

Technical Field

[0001] The present invention relates to the field of Internet of Things (IoT) and artificial intelligence (AI), and in particular to an automatic early warning method for monitoring IoT sensing devices based on deep learning. Background Art

[0002] In modern industrial production, the proper operation and maintenance of equipment are crucial for ensuring production efficiency and product quality. Traditional equipment monitoring methods typically rely on regular manual inspections and preventive maintenance, which not only consumes significant manpower and resources but also makes it difficult to promptly detect potential equipment failures, which can easily lead to sudden breakdowns, production halts, and economic losses. With the development of the Industrial Internet of Things (IoT), an increasing number of devices are equipped with various sensors that can collect real-time data on their operating status. This provides a rich source of data and new technical means for equipment status monitoring and fault prediction.

[0003] In recent years, deep learning technology has made significant progress in fields such as computer vision, natural language processing, and speech recognition, and is gradually being applied to industrial equipment monitoring. Deep learning models can automatically extract features from massive amounts of data and perform complex pattern recognition and predictive analysis, demonstrating strong generalization capabilities and predictive accuracy. By combining deep learning with the Internet of Things (IoT), real-time monitoring of equipment status and fault prediction can be achieved, enhancing the intelligence of equipment management.

[0004] Currently, deep learning-based equipment monitoring and early warning systems have seen initial application in some high-end manufacturing industries. For example, they are used to predict and manage the health of critical equipment such as aircraft engines and wind turbines. However, in traditional manufacturing industries like textile printing and dyeing, despite the diverse equipment types and complex operating environments, where equipment failures often lead to significant production interruptions and economic losses, the level of intelligent equipment monitoring remains relatively low. Summary of the Invention

[0005] In order to solve the technical problems in the prior art, the present invention provides an automatic early warning method for monitoring IoT sensing devices based on deep learning.

[0006] The technical solutions provided by the embodiments of the present invention are as follows:

[0007] First aspect:

[0008] An embodiment of the present invention provides an automatic early warning method for monitoring IoT sensing devices based on deep learning, including:

[0009] S1. Collect sensor data generated by IoT sensing devices. The sensor data is collected through a sensor array, IoT interface, and cloud storage. The sensor data includes temperature, humidity, voltage, current, and pressure generated by the monitored devices during operation.

[0010] S2. Processing and feature extracting the sensor data using a deep neural network, wherein the deep neural network includes an adaptive fusion convolutional recurrent neural network that fuses a convolutional neural network and a recurrent neural network to optimize feature selection and temporal dependency;

[0011] S3. Establishing a device state recognition model by learning data patterns of the normal operation and abnormal state of the monitored device. The device state recognition model determines the current state of the device based on the data patterns of the device operation state. The device state recognition model is established by learning and analyzing data of the device in normal and abnormal states through a deep learning algorithm. The device state recognition model uses a time-weighted attention deep convolutional neural network to enhance the ability to recognize sudden faults by weighting the time series features in the sensor data.

[0012] S4. Monitor the operating status of the monitored equipment in real time and make predictions based on the equipment status recognition model. The equipment status recognition model is trained using a time-enhanced loss function. By weighting the losses of historical data and real-time data to reduce overfitting, the adaptability of the equipment status recognition model to anomalies at different stages is improved. Sensor data is input into the equipment status recognition model to monitor the equipment operating status in real time and predict possible anomalies in the future. Further data analysis and processing are performed based on historical data and equipment characteristics to improve the accuracy of equipment predictions.

[0013] S5. When the monitored equipment operates abnormally or reaches the warning condition, the warning mechanism is automatically triggered and maintenance measures are taken in time;

[0014] S6. Receive operator feedback information and newly connected IoT sensing device data in real time, and use the feedback information and IoT sensing device data to continuously optimize model performance and achieve self-learning.

[0015] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0016] (1) In the present invention, through deep learning technologies such as deep neural networks (DNN) and time-weighted attention deep convolutional neural networks (TWA-DCNN), the present invention can effectively process and analyze the sensor data of the equipment. In particular, the TWA-DCNN model, combined with the time series feature weighting mechanism, can assign different weights to the time series data of the equipment at different operating stages, especially paying more attention to important moments and sudden failure events. This method significantly enhances the ability to identify sudden failures, monitors the equipment status in a timely manner and makes accurate predictions, thereby achieving early fault warning and reducing the occurrence of equipment failures;

[0017] (2) In the present invention, by introducing the adaptive fusion convolutional recurrent neural network (AF-CRNN), combining the feature fusion mechanism of convolutional neural network (CNN) and recurrent neural network (RNN), and the dynamic gating mechanism, it is possible to effectively optimize feature selection and time series dependency. The dynamic gating mechanism enables the model to adaptively adjust information flow according to changes in the equipment operation mode, thereby improving the processing capability of time series data and enhancing the flexibility and accuracy of equipment status recognition. This innovation optimizes the performance of deep learning models in time series data processing and significantly improves the accuracy and sensitivity of fault prediction.

[0018] (3) In order to reduce overfitting and improve the model's adaptability to anomalies at different stages, the present invention introduces the time enhancement loss function (TELF). By weighted learning of historical data and real-time data, TELF can help the model better learn the changes in anomaly patterns, thereby avoiding the model's overfitting to outdated data and improving the model's stability and robustness in different operating environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0020] Figure 1 A system framework diagram of an automatic early warning method for monitoring IoT sensing devices based on deep learning provided by an embodiment of the present invention;

[0021] Figure 2 A data processing flow chart of a method for monitoring and automatically warning IoT sensing devices based on deep learning provided by an embodiment of the present invention;

[0022] Figure 3 A schematic diagram of a device status recognition model based on a deep learning IoT sensing device monitoring and automatic early warning method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0024] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as an "exemplary" in the present invention should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of the word "exemplary" is intended to present concepts in a concrete manner. Furthermore, in the embodiments of the present invention, "and / or" can mean both or either of the two.

[0025] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0026] Reference Manual Figure 1-3 , showing a system framework diagram, data processing flow chart and device status recognition model schematic diagram of a deep learning-based IoT perception device monitoring and automatic early warning method provided by an embodiment of the present invention.

[0027] An embodiment of the present invention provides an automatic early warning method for monitoring IoT sensing devices based on deep learning, the method comprising the following steps:

[0028] S1. Collect sensor data generated by IoT sensing devices. The collection of sensor data is achieved through sensor arrays, IoT interfaces, and cloud storage. The sensor data includes temperature, humidity, voltage, current, and pressure generated by the monitored devices during operation.

[0029] S2. Processing and feature extraction of sensor data using a deep neural network, wherein the deep neural network includes an adaptive fusion convolutional recurrent neural network that fuses a convolutional neural network and a recurrent neural network to optimize feature selection and temporal dependency;

[0030] S3. Establish a device state recognition model by learning the data patterns of the normal operation and abnormal state of the monitored equipment. The device state recognition model is a model that determines the current state of the equipment based on the data patterns of the equipment's operating state. The device state recognition model is established by learning and analyzing the data of the equipment in normal and abnormal states through a deep learning algorithm. The device state recognition model uses a time-weighted attention deep convolutional neural network to enhance the recognition ability of sudden faults by weighting the time series features in the sensor data;

[0031] S4. Monitor the operating status of the monitored equipment in real time and make predictions based on the equipment status recognition model. The equipment status recognition model is trained using a time-enhanced loss function. By weighting the losses of historical data and real-time data to reduce overfitting, the adaptability of the equipment status recognition model to anomalies at different stages is improved. Sensor data is input into the equipment status recognition model to monitor the equipment operating status in real time and predict possible anomalies in the future. At the same time, further data analysis and processing are performed based on historical data and equipment characteristics to improve the accuracy of equipment predictions.

[0032] S5. When the monitored equipment operates abnormally or reaches the warning condition, the warning mechanism is automatically triggered and maintenance measures are taken in time;

[0033] S6. Receive operator feedback and newly connected IoT sensing device data in real time, and use the feedback information and IoT sensing device data to continuously optimize model performance and achieve self-learning.

[0034] S2 specifically includes:

[0035] S21. Data preprocessing: Before inputting into the deep neural network, the sensor data needs to be preprocessed. The preprocessing includes standardization and normalization to adapt to the input requirements of the neural network.

[0036] S22. Construct a deep neural network structure. The deep neural network includes an input layer, a hidden layer, and an output layer. Each layer is composed of several neurons and takes the output of the previous layer as input. It performs nonlinear transformation through the activation function and generates new output. When transferring information from the input layer to the output layer, the forward propagation simplified formula is used.

[0037] In the above formula, is the activation function, is the weight, For input, is the bias term, is the output;

[0038] S23. Feature learning and extraction. During the training process, the weights and biases in the network are adjusted through the backpropagation algorithm so that the predicted output of the deep neural network is as close to the true label as possible. During the training process, the deep neural network automatically learns from the raw data and extracts features that help complete the task.

[0039] S22 specifically includes:

[0040] The input layer receives raw sensor data, including devices that monitor temperature, humidity, and vibration.

[0041] The hidden layer includes multiple fully connected layers, convolutional layers and recurrent layers, each layer is connected by a weight matrix and bias Transform the output of the previous layer into the input of the next layer using an activation function.

[0042] Perform nonlinear mapping, the fully connected layer is a feedforward neural network, the convolutional layer is suitable for image or sequence data, and the recurrent layer is suitable for time series data;

[0043] The output layer consists of one or more nodes for classification or regression. Classification uses the softmax function to output probability distribution, and regression directly outputs numerical values.

[0044] S3 specifically includes:

[0045] S31. Data preparation, including data collection, data labeling, and data cleaning and normalization. Data collection involves collecting data from sensors installed on the equipment under normal operating conditions and abnormal conditions. Data labeling involves annotating the collected data to distinguish normal data from abnormal data. Data cleaning and normalization involves removing noise, filling missing values, and normalizing or standardizing the data to facilitate model learning.

[0046] S32. Feature engineering, including feature extraction. Feature extraction is to extract meaningful features from raw data. Meaningful features include vibration amplitude and temperature changes at specific frequencies. For image data, features are directly learned by convolutional neural networks (CNNs).

[0047] S33. Construct a deep learning algorithm model. The deep learning algorithm model includes a deep learning algorithm. The deep learning algorithm includes a convolutional neural network (CNN) and a recurrent neural network (RNN). If the device status can be identified through image or spectrum analysis, a convolutional neural network is used. For time series data, a recurrent neural network is more suitable. The simplified formula of the convolution operation of the convolutional neural network is:

[0048] In the above formula, Indicates that the output feature map after the convolution operation is located at Eigenvalue of position ， represents the filter, represents the input image, represents the convolution operation, Represents the convolution kernel Located in the middle The weight of the position, Represents the input feature map Located in the middle The input value of the position, represents the size (width and height) of the convolution kernel, and It is the index variable inside the convolution kernel, which is used to traverse all positions of the convolution kernel. In the construction of RNN, a basic RNN unit is represented as,

[0049] In the above formula, Represents the time step The hidden state of represents the input of the current time step, represents the model parameters, Represents the activation function, which includes tanh or ReLU. Long short-term memory (LSTM) or gated recurrent unit (GRU) are variants of RNN and are more suitable for learning long sequence data. The basic unit of the LSTM model is expressed as the following four formulas, which are used to update the unit state and hidden state , and output predicted values ,

[0050] The forget gate determines the information that needs to be forgotten in the previous unit state.

[0051] ;

[0052] The input gate determines the new information that should be updated to the cell state.

[0053] ;

[0054] Unit state update, combining forget and input operations,

[0055] ;

[0056] The output gate determines the cell state information that should be output as the new hidden state.

[0057] ;

[0058] In the above formula, represents the output of the forget gate, represents the sigmoid activation function, represents the weight matrix, represents the bias term, represents the concatenation of the hidden state of the previous time step and the input of the current time step, represents the hyperbolic tangent function, represents the model parameters, represents element-wise multiplication;

[0059] S34, training the model, including dividing the data set, selecting a loss function and an optimizer, and a training process, wherein dividing the data set is to divide the data into a training set, a validation set, and a test set, and the training process of the training model is to adjust the model parameters through a back propagation algorithm to minimize the loss function, and iterate multiple rounds until the model no longer improves significantly or reaches a predetermined stopping condition;

[0060] S35. Evaluation and tuning, including performance evaluation and model tuning. Performance evaluation is to evaluate the model performance on the test set using accuracy, recall rate and F1 score. Model tuning is to adjust the model structure and hyperparameters based on the evaluation results or use deep CNN and bidirectional RNN for multiple iterative training to optimize performance.

[0061] S4 specifically includes:

[0062] S41. Data collection and preprocessing: Collecting parameters of the equipment during operation through sensors installed on the equipment. The sensors include temperature sensors, vibration sensors, and current and voltage sensors. The collected data is preprocessed. The preprocessing includes noise removal, missing value processing, and data standardization to facilitate subsequent analysis.

[0063] S42. Feature engineering: extracting useful features from raw data. Features include time series trends, frequency components, statistical properties of signals, and complex features. Statistical properties of signals include mean, variance, and kurtosis. Complex features include descriptions of waveform shapes. The goal of feature selection is to find the set of variables that best characterizes the device state.

[0064] S43. Select or build a device status recognition model. Models include supervised learning models, unsupervised learning models, and deep learning models. Supervised learning models include support vector machines (SVMs), random forests (RFs), and gradient boosted trees (GBTs), which are used for classification or regression problems to predict whether a device is about to fail or estimate its remaining useful life. Unsupervised learning models include clustering algorithms (K-means, DBSCAN), which are used to discover patterns in data and identify normal and abnormal operating modes of the device. Deep learning models include recurrent neural networks (RNNs), long short-term memory networks (LSTMs), or gated recurrent units (GRUs), which are suitable for processing time series data and capturing how device status changes over time.

[0065] S44, real-time detection and prediction, inputs real-time sensor data into the trained model, and the model evaluates the operating status of the equipment based on the characteristics of the current data. If the model predicts that the status deviates from the normal range, it can issue an early warning to prompt maintenance personnel to take preventive measures.

[0066] By continuously collecting data and feeding it back into the model, we optimize model performance and improve prediction accuracy through online learning or regular retraining. At the same time, we combine expert knowledge and actual maintenance feedback to continuously adjust and optimize feature engineering and model parameters.

[0067] When equipment is operating abnormally or reaches a warning condition, the warning mechanism is automatically triggered, prompting timely maintenance measures. When the equipment status detection model detects an abnormality or reaches a warning condition, the warning mechanism is automatically triggered, notifying relevant personnel to address the situation. At the same time, appropriate maintenance measures, such as component replacement and equipment repair, are designed to ensure normal equipment operation.

[0068] In one possible implementation, the device state recognition model introduces a temporal weighted attention deep convolutional neural network (TWA-DCNN) algorithm to perform weighted processing on temporal features in sensor data. The weight is calculated by the following formula:

[0069] in, For time point The corresponding weight value, is the current time point, is the weighted attenuation coefficient, is the weight adjustment factor, The weighted function can dynamically adjust the sampling frequency, assign higher weights to the sampling data at important moments, and enhance the model's ability to identify sudden faults.

[0070] The weighted attenuation coefficient and weight adjustment factor dynamically weight the sampled data at different time points. This weighting function dynamically adjusts the sampling frequency to assign higher weights to data sampled at critical moments, thereby improving the model's responsiveness to sudden failures. This approach enables timely identification of potential failure modes in response to sudden changes in device status or abnormalities, without compromising model accuracy due to interference from standard data. By optimizing the weighting strategy, the model can more effectively process sensor data in high-noise environments.

[0071] In one possible implementation, the device state recognition model specifically includes: using an adaptive fused convolutional recurrent neural network (AF-CRNN), combining the feature fusion mechanisms of convolutional neural networks (CNNs) and recurrent neural networks (RNNs), and introducing a self-developed dynamic gating mechanism to optimize feature selection and timing dependencies. The dynamic gating mechanism is calculated using the following formula:

[0072] in, is the dynamic gate value, is the hidden state at the previous moment, Input features for the current moment, is the gating weight matrix, is the bias term, The Sigmoid activation function can adaptively adjust the information flow and enhance the model's ability to process time series data by learning changes in device operation modes.

[0073] The dynamic gating mechanism adaptively adjusts the flow of information. This is achieved through a gating formula in which the input features at each moment are adjusted based on the hidden state at the previous moment and the gating weight matrix. Through this mechanism, the model can adjust the emphasis it places on input data from different time periods based on the changing operating state of the device, thereby improving the model's ability to handle complex time series data. In particular, it can quickly detect anomalies and make accurate predictions when the device state undergoes sudden changes.

[0074] In one possible implementation, the deep convolutional neural network specifically includes: the weighted deep convolutional neural network (TWA-DCNN) model combined with a time-enhanced loss function (TELF), which weights the loss functions of historical data and real-time data so that the model can differentiate between historical fault modes and real-time anomalies when learning abnormal modes. The loss function is defined as:

[0075] in, is the loss value, is the actual value, is the predicted value, is the weighting coefficient, The error between the model prediction and the actual situation is reduced by weighting, and the adaptability of the model to abnormalities at different stages is improved.

[0076] This loss function weights the losses of historical and real-time data, enabling the model to differentiate learning across different time periods when learning anomaly patterns. This effectively reduces overfitting and improves the model's generalization. By weighting the loss function, the model can better balance the influence of historical data on learning with the sensitivity of real-time data to changes in device status. This approach ensures that the model is more adaptable to both historical failure patterns and real-time anomalies, enabling it to promptly detect potential failures and provide early warnings. During implementation, precise adjustment of the weighting coefficients further optimizes the model's learning process, preventing overfitting during training.

[0077] In one possible implementation, the data preprocessing step specifically includes: introducing a time window weighted outlier detection algorithm (TWA-OUD), and detecting and correcting outliers during the preprocessing process using the following formula:

[0078] in, is the normalized abnormality metric, is the mean value of the data at the previous moment, is the standard deviation of the data at the previous moment, is the corrected data value, To correct the coefficient, the above algorithm is used to detect sudden changes or anomalies in the device status data in real time, and the accuracy of subsequent model learning is improved by effectively correcting the outliers.

[0079] This algorithm detects and corrects outliers in real time during data preprocessing. This correction method is based on data standardization and correction coefficients. The standardized outlier metric is compared with the mean and standard deviation of the previous data moment to detect sudden changes or anomalies in device status data in real time. Correcting outliers effectively removes noise and errors from sensor data, ensuring the accuracy of subsequent model learning. This method reduces errors introduced by factors such as environmental changes and sensor drift in device monitoring systems, improving the accuracy and reliability of device status identification.

[0080] In one possible implementation, the deep neural network specifically includes: further introducing an ensemble learning-based fault prediction module (EL-FPM), integrating multiple weak classifiers with a deep learning model for fusion, and using the following weighted fusion formula:

[0081] in, is the final prediction probability after fusion, For the The predicted probability of a classifier, is the weight coefficient of the corresponding classifier, is the number of classifiers.

[0082] The fault prediction module (EL-FPM) integrates multiple weak classifiers with a deep learning model and uses weighted fusion to improve prediction accuracy. By weighting the prediction results of multiple classifiers, the robustness of the overall system can be improved and the deviation that may be caused by a single classifier can be reduced. Specifically, the final prediction probability is obtained by weighting the output results of each classifier. This weighting strategy is adjusted based on the prediction accuracy and weight coefficient of each classifier. The role of this method is to effectively improve the prediction accuracy under various complex fault modes, while reducing the prediction errors caused by the limitations of a single model. During implementation, by adjusting the number of classifiers and the weight coefficient, the performance of the fault prediction system can be further optimized and its adaptability to different types of faults can be enhanced.

[0083] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0084] (1) In the present invention, through deep learning technologies such as deep neural networks (DNN) and time-weighted attention deep convolutional neural networks (TWA-DCNN), the present invention can effectively process and analyze the sensor data of the equipment. In particular, the TWA-DCNN model, combined with the time series feature weighting mechanism, can assign different weights to the time series data of the equipment at different operating stages, especially paying more attention to important moments and sudden failure events. This method significantly enhances the ability to identify sudden failures, monitors the equipment status in a timely manner and makes accurate predictions, thereby achieving early fault warning and reducing the occurrence of equipment failures;

[0085] (2) In the present invention, by introducing the adaptive fusion convolutional recurrent neural network (AF-CRNN), combining the feature fusion mechanism of convolutional neural network (CNN) and recurrent neural network (RNN), and the dynamic gating mechanism, it is possible to effectively optimize feature selection and time series dependency. The dynamic gating mechanism enables the model to adaptively adjust information flow according to changes in the equipment operation mode, thereby improving the processing capability of time series data and enhancing the flexibility and accuracy of equipment status recognition. This innovation optimizes the performance of deep learning models in time series data processing and significantly improves the accuracy and sensitivity of fault prediction.

[0086] (3) In order to reduce overfitting and improve the model's adaptability to anomalies at different stages, the present invention introduces the time enhancement loss function (TELF). By weighted learning of historical data and real-time data, TELF can help the model better learn the changes in anomaly patterns, thereby avoiding the model's overfitting to outdated data and improving the model's stability and robustness in different operating environments.

[0087] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

[0088] There are a few points to note:

[0089] (1) The drawings of the embodiments of the present invention only relate to the structures related to the embodiments of the present invention. Other structures may refer to conventional designs.

[0090] (2) For the sake of clarity, the thickness of layers or regions in the drawings used to describe the embodiments of the present invention are exaggerated or reduced, that is, these drawings are not drawn to scale. It is understood that when an element such as a layer, film, region, or substrate is referred to as being "on" or "under" another element, the element may be "directly on" or "under" the other element or intervening elements may be present.

[0091] (3) In the absence of conflict, the embodiments of the present invention and the features therein may be combined with each other to form new embodiments.

[0092] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. The protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for automatic early warning monitoring of IoT sensing devices based on deep learning, characterized in that: include: S1. Collect sensor data generated by IoT sensing devices. The sensor data is collected through a sensor array, IoT interface, and cloud storage. The sensor data includes temperature, humidity, voltage, current, and pressure generated by the monitored devices during operation. S2. Processing and feature extracting the sensor data using a deep neural network, wherein the deep neural network includes an adaptive fusion convolutional recurrent neural network that fuses a convolutional neural network and a recurrent neural network to optimize feature selection and temporal dependency; S3. Establishing a device state recognition model by learning data patterns of the normal operation and abnormal state of the monitored device. The device state recognition model determines the current state of the device based on the data patterns of the device operation state. The device state recognition model is established by learning and analyzing data of the device in normal and abnormal states through a deep learning algorithm. The device state recognition model uses a time-weighted attention deep convolutional neural network to enhance the ability to recognize sudden faults by weighting the time series features in the sensor data. S4. Monitor the operating status of the monitored equipment in real time and make predictions based on the equipment status recognition model. The equipment status recognition model is trained using a time-enhanced loss function. By weighting the losses of historical data and real-time data to reduce overfitting, the adaptability of the equipment status recognition model to anomalies at different stages is improved. Sensor data is input into the equipment status recognition model to monitor the equipment operating status in real time and predict possible anomalies in the future. Further data analysis and processing are performed based on historical data and equipment characteristics to improve the accuracy of equipment predictions. S5. When the monitored equipment operates abnormally or reaches the warning condition, the warning mechanism is automatically triggered and maintenance measures are taken in time; S6. Receive operator feedback information and newly connected IoT sensing device data in real time, and use the feedback information and IoT sensing device data to continuously optimize model performance and achieve self-learning.

2. The method for automatic early warning based on deep learning IoT sensing device monitoring according to claim 1 is characterized in that: The S2 specifically includes: S21, data preprocessing: Before inputting into the deep neural network, the sensor data needs to be preprocessed, and the preprocessing includes standardization and normalization to adapt to the input requirements of the neural network; S22. Construct a deep neural network structure, which includes an input layer, a hidden layer, and an output layer. Each layer is composed of a number of neurons, and takes the output of the previous layer as input, performs nonlinear transformation through an activation function, and generates a new output. When information is transmitted from the input layer to the output layer, a simplified forward propagation formula is used. In the above formula, is the activation function, is the weight, For input, is the bias term, is the output; S23. Feature learning and extraction. During the training process, the weights and biases in the network are adjusted through the back-propagation algorithm so that the predicted output of the deep neural network is as close to the true label as possible. During the training process, the deep neural network automatically learns from the raw data and extracts features that help complete the task.

3. The automatic early warning method for monitoring IoT sensing devices based on deep learning according to claim 2 is characterized in that: The S22 specifically includes: The input layer is used to receive raw sensor data, and the raw sensors include devices that monitor temperature, humidity and vibration; The hidden layer includes multiple fully connected layers, convolutional layers and recurrent layers, each layer is connected by a weight matrix and bias Transform the output of the previous layer into the input of the next layer using an activation function. Perform nonlinear mapping, the fully connected layer is a feedforward neural network, the convolutional layer is suitable for image or sequence data, and the recurrent layer is suitable for time series data; The output layer includes one or more nodes for classification or regression. The classification uses a softmax function to output a probability distribution, and the regression directly outputs a numerical value.

4. The method for automatic early warning based on deep learning IoT sensing device monitoring according to claim 1 is characterized in that: The S3 specifically includes: S31. Data preparation, including data collection, data labeling, and data cleaning and normalization. The collected data is data collected from sensors installed on the device under normal operating conditions and abnormal conditions. The labeled data is annotated to distinguish normal data from abnormal data. The data cleaning and normalization is to remove noise, fill missing values, and normalize or standardize the data to facilitate model learning. S32. Feature engineering, including feature extraction, wherein the feature extraction is to extract meaningful features from the raw data. The meaningful features include vibration amplitude and temperature change at a specific frequency. For image data, the features are directly learned by a convolutional neural network (CNN); S33. Construct a deep learning algorithm model, wherein the deep learning algorithm model includes the deep learning algorithm, and the deep learning algorithm includes the convolutional neural network (CNN) and the recurrent neural network (RNN). If the device status can be identified by image or spectrum analysis, the convolutional neural network is used. For time series data, the recurrent neural network is more suitable. The simplified formula of the convolution operation of the convolutional neural network is: In the above formula, Indicates that the output feature map after the convolution operation is located at The eigenvalues ​​of the position, represents the filter, represents the input image, represents the convolution operation, Represents the convolution kernel Located in the middle The weight of the position, Represents the input feature map Located in the middle The input value of the position, represents the size (width and height) of the convolution kernel, is the index variable inside the convolution kernel, which is used to traverse all positions of the convolution kernel. In the construction of the RNN, a basic RNN unit is represented as, In the above formula, Represents the time step The hidden state of represents the input of the current time step, represents the model parameters, Represents the activation function, which includes tanh or ReLU. Long short-term memory (LSTM) or gated recurrent unit (GRU) are variants of the RNN and are more suitable for learning long sequence data. The basic unit of the LSTM model is expressed as the following four formulas for updating the unit state and hidden state , and output predicted values , The forget gate determines the information that needs to be forgotten in the previous unit state. ; The input gate determines the new information that should be updated to the cell state. ; Unit state update, combining forget and input operations, ; The output gate determines the cell state information that should be output as the new hidden state. ; In the above formula, represents the output of the forget gate, represents the sigmoid activation function, represents the weight matrix, represents the bias term, represents the concatenation of the hidden state of the previous time step and the input of the current time step, represents the hyperbolic tangent function, represents the model parameters, represents element-wise multiplication; S34, training the model, including dividing the data set, selecting a loss function and an optimizer, and a training process, wherein the data set is divided into a training set, a validation set, and a test set, and the training process of the training model is to adjust the model parameters through a back propagation algorithm to minimize the loss function, and iterate multiple rounds until the model no longer improves significantly or reaches a predetermined stopping condition; S35. Evaluation and tuning, including performance evaluation and model tuning. The performance evaluation is to evaluate the model performance on the test set using accuracy, recall rate and F1 score. The model tuning is to adjust the model structure and hyperparameters according to the evaluation results or use deep CNN and bidirectional RNN for multiple iterative training to optimize performance.

5. The method for automatic early warning based on deep learning IoT sensing device monitoring according to claim 1 is characterized in that: The S4 specifically includes: S41, data collection and preprocessing: collecting parameters of the device during operation through sensors installed on the device, including temperature sensors, vibration sensors, and current and voltage sensors, and preprocessing the collected data. The preprocessing includes noise removal, missing value processing, and data standardization to facilitate subsequent analysis; S42. Feature Engineering: Extracting useful features from raw data. These features include time series trends, frequency components, statistical properties of signals, and complex features. The statistical properties of signals include mean, variance, and kurtosis. Complex features include descriptions of waveform shapes. The goal of feature selection is to find the set of variables that best characterizes the device state. S43. Select or construct a device status recognition model, including supervised learning models, unsupervised learning models, and deep learning models. Supervised learning models include support vector machines (SVMs), random forests (RFs), and gradient boosted trees (GBTs), which are used for classification or regression problems to predict whether a device is about to fail or estimate its remaining useful life. Unsupervised learning models include clustering algorithms, including K-means and DBSCAN, which are used to discover patterns in data and identify normal and abnormal operating modes of the device. Deep learning models include recurrent neural networks (RNNs), long short-term memory networks (LSTMs), or gated recurrent units (GRUs), which are suitable for processing time series data and capturing how device status changes over time. S44, real-time detection and prediction, inputs real-time sensor data into the trained model, and the model evaluates the operating status of the equipment based on the characteristics of the current data. If the model predicts that the status deviates from the normal range, it can issue an early warning to prompt maintenance personnel to take preventive measures.

6. The method for automatic early warning based on deep learning IoT sensing device monitoring according to claim 5 is characterized in that: The device state recognition model specifically includes: By introducing the temporal weighted attention deep convolutional neural network (TWA-DCNN) algorithm, the temporal features in the sensor data are weighted. The weight is calculated by the following formula: in, For time point The corresponding weight value, is the current time point, is the weighted attenuation coefficient, is the weight adjustment factor, The weighted function can dynamically adjust the sampling frequency, assign higher weights to the sampling data at important moments, and enhance the model's ability to identify sudden faults.

7. The method for automatic early warning based on deep learning IoT sensing device monitoring according to claim 6 is characterized in that: The device state recognition model specifically includes: We use an adaptive fused convolutional recurrent neural network (AF-CRNN) that combines the feature fusion mechanisms of convolutional neural networks (CNNs) and recurrent neural networks (RNNs) and introduces a self-developed dynamic gating mechanism to optimize feature selection and temporal dependencies. The dynamic gating mechanism is calculated using the following formula: in, is the dynamic gate value, is the hidden state at the previous moment, Input features for the current moment, is the gating weight matrix, is the bias term, The Sigmoid activation function can adaptively adjust the information flow and enhance the model's ability to process time series data by learning changes in device operation modes.

8. The method for automatic early warning based on deep learning IoT sensing device monitoring according to claim 6 is characterized in that: The deep convolutional neural network specifically includes: The Temporal Weighted Attention Deep Convolutional Neural Network (TWA-DCNN) model combines the Time Enhancement Loss Function (TELF) to weight the loss functions of historical data and real-time data, enabling the model to differentiate between historical failure modes and real-time anomalies when learning anomaly patterns. The loss function is defined as: in, is the loss value, is the actual value, is the predicted value, is the weighting coefficient, The error between the model prediction and the actual situation is reduced by weighting, and the adaptability of the model to abnormalities at different stages is improved.

9. The method for automatic early warning based on deep learning IoT sensing device monitoring according to claim 2 is characterized in that: The data preprocessing step specifically includes: The time window weighted outlier detection algorithm (TWA-OUD) is introduced to detect and correct outliers during the preprocessing process using the following formula: in, is the normalized abnormality metric, is the mean value of the data at the previous moment, is the standard deviation of the data at the previous moment, is the corrected data value, To correct the coefficient, the above algorithm is used to detect sudden changes or anomalies in the device status data in real time, and the accuracy of subsequent model learning is improved by effectively correcting the outliers.

10. The method for automatic early warning based on deep learning IoT sensing device monitoring according to claim 2 is characterized in that: The deep neural network specifically includes: We further introduce an ensemble learning-based fault prediction module (EL-FPM), which integrates multiple weak classifiers and combines them with a deep learning model, using the following weighted fusion formula: in, is the final prediction probability after fusion, For the The predicted probability of a classifier, is the weight coefficient of the corresponding classifier, is the number of classifiers.

Images