Smart factory management system and method based on cloud platform

Abstract

The present invention discloses a cloud-based intelligent factory management system and method, which relates to the field of industrial Internet of Things technology. The method includes extracting features of pre-processed production equipment data and environmental data to form a feature data set, constructing an equipment failure prediction model based on the feature data set, obtaining failure prediction results, and generating a resource scheduling strategy based on the failure prediction results. The constructed equipment failure prediction model can detect potential equipment failure risks at an early stage, reduce downtime, and improve the overall operational efficiency and safety of the factory. The resource scheduling strategy realizes the dynamic optimization and scheduling of resources within the factory, which not only improves production efficiency, but also significantly reduces energy consumption and failure risks. It can also flexibly respond to changes in demand for different production tasks, thereby improving overall economic benefits.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial Internet of Things, and in particular to a cloud platform-based smart factory management system and method. Background Art

[0002] With the rapid development of information technology, the Industrial Internet of Things has gradually become an important part of modern manufacturing. By applying sensors, network communications and data analysis technologies to production equipment and environmental monitoring, the Industrial Internet of Things can achieve comprehensive perception and intelligent control of the production process.

[0003] Existing cloud-based factory management methods still have some shortcomings. First, during the feature extraction process, conventional manual feature engineering methods are not only time-consuming and labor-intensive, but also difficult to capture the deep-level characteristics of the data. Furthermore, existing equipment failure prediction models mostly rely on a single type of machine learning algorithm, which has limited performance when processing nonlinear data and cannot fully exploit the existing correlations in the data. Furthermore, when formulating resource scheduling strategies, the correlations and dynamic changes between devices are often ignored, resulting in scheduling solutions that are not flexible and efficient. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a cloud platform-based intelligent factory management method to solve the problems of poor accuracy of equipment failure prediction models and unreasonable resource scheduling in factory management methods.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a smart factory management method based on a cloud platform, which includes collecting production equipment data and environmental data and performing pre-processing;

[0008] Extract the features of pre-processed production equipment data and environmental data to form a feature data set;

[0009] Based on the feature data set, an equipment fault prediction model is constructed to obtain the fault prediction results;

[0010] Generate resource scheduling strategies based on fault prediction results.

[0011] As a preferred solution of the cloud platform-based smart factory management method of the present invention, wherein:

[0012] The production equipment data includes temperature, pressure, vibration, rotation speed, load and production rate;

[0013] The environmental data includes air quality and indoor temperature and humidity.

[0014] As a preferred solution of the cloud platform-based smart factory management method of the present invention, the pre-processing includes the following steps:

[0015] Use a mean filter to remove noise from production equipment data and perform normalization;

[0016] Outliers in the environmental data were identified and removed using the interquartile range, and missing data points were filled using linear interpolation.

[0017] As a preferred solution of the cloud platform-based smart factory management method of the present invention, the following steps are included: extracting the features of the pre-processed production equipment data and environmental data to form a feature data set:

[0018] Select the multilayer perceptron network as the feature extraction method;

[0019] The pre-processed production equipment data and environmental data are converted into a columnar storage file format and input into a multi-layer perceptron network for forward propagation to obtain features representing the status of the production equipment;

[0020] The characteristics of the production equipment status include indoor temperature and humidity, rotation speed, load and production rate;

[0021] Use principal component analysis to map the high-dimensional features of the production equipment status into a low-dimensional space to obtain the features of the production equipment status after dimensionality reduction;

[0022] The features of the production equipment status after dimensionality reduction are stored in a fast read and write format to form a structured feature data set.

[0023] As a preferred solution of the cloud platform-based smart factory management method of the present invention, the equipment failure prediction model is constructed based on the feature data set, including the following steps:

[0024] Divide the feature dataset into training set, test set and validation set;

[0025] Selecting convolutional neural networks as the architecture for building equipment failure prediction models;

[0026] Use the training set and Bayesian optimization method to randomly perturb the equipment failure prediction model to generate training samples;

[0027] Use the sliding window technique to adjust the subsequence labels in the training samples, generate fault category labels and combine them into training sample feature vectors;

[0028] Use the boosting ensemble learning framework to assign initial weights to each training sample feature vector and calculate the initial weight value of each training sample feature vector;

[0029] Use the weak learner to iterate the initial weight value of each training sample feature vector to obtain the updated weight value of the training sample feature vector, which is expressed as:

[0030] ;

[0031] ;

[0032] in, represents the initial weight value of each training sample feature vector, Indicates the The error rate of the feature vector of the training samples, Indicates the updated The feature vector of the training sample is The weight value in the round iteration, Indicates the updated The feature vector of the training sample is The weight value in the round iteration, Indicates the The fault category label value of the training sample feature vector, represents the exponential function, Indicates the Is the label category value of the training sample feature vector The exponential function value of the misclassification of the weak learner;

[0033] Use the test set to perform confusion matrix analysis on the weight values ​​of the updated training sample feature vectors to obtain the feature matrix of the training sample and calculate its anomaly score;

[0034] Draw the loss curve based on the feature matrix of the training sample, and use the local outlier factor to identify the wrong training sample type and make improvements;

[0035] The equipment failure prediction model is retrained based on the improved training sample type until the loss curve tends to be stable, and a trained equipment failure prediction model is obtained.

[0036] As a preferred solution of the cloud platform-based smart factory management method of the present invention, the method of obtaining the fault prediction result includes the following steps:

[0037] The validation set is input into the equipment failure prediction model and deployed on the factory's production equipment. The isolation forest algorithm is introduced to define the path length of the training sample feature matrix and calculate the comprehensive failure probability value of the factory's production equipment. The expression is:

[0038] ;

[0039] ;

[0040] ;

[0041] in, Represents the standardized failure probability value of factory production equipment, represents the activation function, represents the weight of the training sample feature matrix, represents the training sample feature matrix, represents the bias term, represents the abnormal score of the training sample feature matrix, represents the path length of the training sample feature matrix, Represents the number of training sample feature matrices The expected value of the path length is Represents the comprehensive failure probability value of factory production equipment, The weight coefficient representing the preliminary failure probability value of the factory production equipment, The weight coefficient representing the abnormality score of the training sample feature matrix;

[0042] Set alarm thresholds and determine the risk of equipment failure based on whether the comprehensive failure probability of factory equipment falls within the threshold.

[0043] The judgment result is a fault prediction result.

[0044] As a preferred solution of the cloud platform-based smart factory management method of the present invention, wherein: based on the fault prediction results, a resource scheduling strategy is generated, including the following steps:

[0045] Select industrial IoT devices as edge computing nodes;

[0046] Industrial IoT devices and actual production tasks of the current factory are listed as node features and edge features respectively;

[0047] According to the fault detection results, update the node features and edge features and build the graph structure;

[0048] Use a graph neural network model to encode the graph structure and extract the embedding vectors of each node feature and edge feature;

[0049] Based on the embedding vectors of each node feature and edge feature, define the optimization objective;

[0050] The optimization objectives include maximizing load balancing, minimizing response time, and minimizing failure risk;

[0051] According to the defined optimization goal, set the population size and generate the initial population;

[0052] Evaluate each initial population and introduce a fitness function to calculate the fitness value of each initial population on the optimization target;

[0053] According to the fitness value of each initial population on the optimization target, each population is non-dominated and sorted, and the crowding distance is calculated to obtain some excellent individuals as the parent population;

[0054] Select V crossover points from the parent population, use the binary crossover algorithm to exchange gene segments, and generate offspring populations;

[0055] Merge the parent population and the offspring population, select a set number of individual populations and iterate until the population evolution completes convergence to form the final population;

[0056] The Pareto front optimal solution is extracted from the final population as the resource scheduling strategy.

[0057] In a second aspect, the present invention provides a cloud-based smart factory management system, comprising:

[0058] Preprocessing module, collects production equipment data and environmental data, and performs preprocessing;

[0059] Feature extraction module, which extracts the features of pre-processed production equipment data and environmental data to form a feature data set;

[0060] The fault prediction module builds an equipment fault prediction model based on the feature data set and obtains the fault prediction results;

[0061] The policy generation module generates resource scheduling policies based on fault prediction results.

[0062] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the cloud platform-based smart factory management method as described in the first aspect of the present invention is implemented.

[0063] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the cloud platform-based smart factory management method as described in the first aspect of the present invention.

[0064] The present invention has the following beneficial effects: extracting features from pre-processed production equipment data and environmental data to form a feature dataset, constructing an equipment failure prediction model, obtaining failure prediction results, and generating a resource scheduling strategy based on the failure prediction results. The constructed equipment failure prediction model can detect potential equipment failure risks at an early stage, reducing downtime and improving the overall operational efficiency and safety of the factory. The resource scheduling strategy realizes the dynamic optimization and scheduling of resources within the factory, which not only improves production efficiency, but also significantly reduces energy consumption and failure risks. It can also flexibly respond to changes in demand for different production tasks, thereby improving overall economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the technical solutions of 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 paying any creative work.

[0066] Figure 1 This is a flow chart of the cloud platform-based smart factory management method in Example 1;

[0067] Figure 2 This is a diagram for determining the risk level of equipment failure in Example 1. DETAILED DESCRIPTION

[0068] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0069] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0070] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0071] Example 1, with reference to Figure 1 and Figure 2 , which is the first embodiment of the present invention, provides a smart factory management method based on a cloud platform, comprising the following steps:

[0072] S1. Collect production equipment data and environmental data and perform preprocessing.

[0073] The following steps are included:

[0074] Production equipment data includes temperature, pressure, vibration, speed, load, and production rate;

[0075] Environmental data includes air quality and indoor temperature and humidity;

[0076] The production equipment and environmental data collected in this step can help to promptly identify and resolve potential problems (such as overheating, overpressure, and excessive vibration) in subsequent analysis, reducing the time equipment operates in abnormal conditions and extending its service life. This also maintains appropriate temperature, humidity, and air quality, reducing energy consumption and protecting employee health.

[0077] Use a mean filter (smoothing technique) to remove noise from production equipment data and perform normalization;

[0078] Select several data points (such as 3 or 5 data points) of the production equipment data and slide the time window. Calculate the average value of the data points in each time window and replace it with the data point in the center of the window.

[0079] The purpose of normalization is to scale the production equipment data to a specific range (such as [-1,1]), so that data from different production equipment have similar scales;

[0080] Use the interquartile range (i.e., the difference between the first quartile Q1 and the third quartile Q3 of the data, IQR) to identify and remove outliers in the environmental data, and use linear interpolation to fill in missing data points;

[0081] Determine the upper and lower boundaries of environmental data, and treat values ​​below the lower boundary or above the upper boundary as outliers and remove them;

[0082] Detect the missing parts in the environmental data, and calculate the estimated value according to the linear relationship based on the values ​​of the two known data points before and after, and finally fill the missing position with the estimated value.

[0083] S2. Extract features of the pre-processed production equipment data and environmental data to form a feature data set.

[0084] The following steps are included:

[0085] Multilayer Perceptron (MLP) network was selected as the feature extraction method;

[0086] The reason for choosing MLP is that compared with common feature extraction methods, MLP can better extract implicit features in high-dimensional data and can capture nonlinear relationships in the data;

[0087] The pre-processed production equipment data and environmental data are converted into a columnar storage format (Parquet format) and input into a multi-layer perceptron network (including input layer, hidden layer, and output layer) for forward propagation to obtain features representing the status of the production equipment;

[0088] Initialize the bias terms of the multilayer perceptron network with small random values ​​(to avoid gradient disappearance);

[0089] The preprocessed production equipment data and environmental data are passed through the input layer of the MLP (each row represents a data point, each column represents a feature) and then into the hidden layer. The weighted sum of each neuron is calculated and the ReLU activation function is applied. The output of the hidden layer is obtained and passed to the second layer. The same calculation process is repeated. Finally, the output of the second layer is passed to the output layer, and the linear activation function is used to obtain the feature representation of the final production equipment status.

[0090] The characteristics of production equipment status include indoor temperature and humidity, rotation speed, load, and production rate;

[0091] Use principal component analysis (PCA) to map the high-dimensional features of the production equipment status into a low-dimensional space to obtain the reduced-dimensional features of the production equipment status.

[0092] First, the covariance matrix of the characteristics of the production equipment status is calculated and the variance eigenvalue decomposition of the principal components is performed. After the decomposition is completed, the principal components with a variance contribution rate of 95% are selected based on the size of the principal component variance eigenvalues. The principal components are sorted from high to low according to the variance eigenvalues ​​to form a new low-dimensional feature space. At the same time, the low-dimensional feature space is projected onto the selected principal components to obtain the feature representation of the production equipment status after dimensionality reduction.

[0093] The features of the production equipment status after dimensionality reduction are stored in a fast read-write format (i.e., HDF5 format) to form a structured feature data set;

[0094] The HDF5 format is used in this step, which significantly speeds up data reading, especially when processing large-scale data sets.

[0095] S3. Build an equipment failure prediction model based on the feature dataset.

[0096] The following steps are included:

[0097] The feature dataset is divided into training set, test set and validation set (with the allocation ratios of 70%, 15% and 15% respectively);

[0098] Selecting convolutional neural networks as the architecture for building equipment failure prediction models;

[0099] The reason for choosing convolutional neural networks here is that they are suitable for processing data with spatial structure (such as images or time series) and can better capture complex patterns in the data;

[0100] This equipment failure prediction model consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. It is designed for binary classification problems (is the equipment operating status normal or abnormal) and regression problems (failure probability).

[0101] Input layer: For image data, the input is a three-dimensional tensor (height, width, number of color channels). For example, a 28x28 pixel grayscale image is (28, 28, 1), while an RGB color image is (28, 28, 3). For time series data, the shape of the input layer is (time step, number of features).

[0102] Convolutional layer: Use convolution kernel filters to extract local features of the input spatial structure data (i.e., the depth of the local feature map), and use the ReLU activation function to increase the nonlinearity of the convolutional layer (to enable the model to learn more complex patterns);

[0103] Pooling layer: Based on the depth of the local feature map of the convolutional layer, it defines the maximum value of the local area (i.e., reduces the size of the feature map). For example, a 2x2 maximum pooling will reduce the height and width of the feature map by half.

[0104] Fully connected layer: The reduced feature map size is converted into a one-dimensional vector and input into each neuron of the fully connected layer. Dropout (regularization technique) is used to randomly discard some neurons (to reduce overfitting);

[0105] Output layer: Select the appropriate activation function according to the specific task requirements to obtain the output result (Sigmoid activation function is used for binary classification problems, Softmax activation function is used for multi-classification problems);

[0106] Use the training set and Bayesian optimization method to randomly perturb the equipment failure prediction model (uniform distribution perturbation and normal distribution perturbation) to generate training samples;

[0107] Determine the hyperparameters of the equipment failure prediction model (learning rate, batch size, and convolution kernel size), and set a reasonable value range for each hyperparameter. For example, the learning rate can vary between 10 to the power of -5 and 10 to the power of -2, and the batch size can vary between 32 and 256. The convolution kernel size needs to be discretized to define its value range.

[0108] Uniformly distributed perturbation: randomly select new hyperparameter values ​​from the current range of hyperparameter values. For example, if the current learning rate is 0.001, you can randomly select a new learning rate in the range of [0.0005, 0.0015];

[0109] Normal distribution perturbation: Use normal distribution to perturb the current hyperparameter value. For example, assuming the current batch size is 128 and the standard deviation is 46, the new batch size after normal distribution perturbation is 174;

[0110] Use Bayesian optimization (hyperparameter value tuning) to iterate the hyperparameter values ​​obtained after uniform distribution perturbation and normal distribution perturbation. After the iteration is completed, high-performance training samples are generated;

[0111] Use the sliding window technique to adjust the subsequence labels in the training samples, generate fault category labels and combine them into training sample feature vectors;

[0112] Use the sliding window technique to segment the training samples, generate multiple continuous subsequence labels, and select an appropriate time window size and step size (determined according to task requirements. For example, if you need to capture the changing trend of device status within an hour, you can choose a window size of 60, and the same applies to the step size).

[0113] Set the initial index according to the selected window size and step size, and traverse from the first subsequence label to the last subsequence label. After completing the update of the subsequence label, determine whether a fault occurs within the time range of each subsequence label. If any fault occurs, generate a fault category label, that is, 1 for fault and 0 for normal, and combine them into a training sample feature vector;

[0114] Use the Boosting ensemble learning framework to assign initial weights to each training sample feature vector and calculate the initial weight value and error rate of each training sample feature vector.

[0115] The reason for assigning an initial weight to each training sample feature vector here is to ensure that all training sample feature vectors have the same influence before calculating the error rate;

[0116] This step can prevent the equipment fault prediction model from overfitting or biasing to abnormal training samples in the early stages by calculating the initial weight values ​​and error rates of the feature vectors of all training samples.

[0117] Use the weak learner to iterate the initial weight value of each training sample feature vector to obtain the updated weight value of the training sample feature vector, which is expressed as:

[0118] ;

[0119] ;

[0120] in, represents the initial weight value of each training sample feature vector, Indicates the The error rate of the training sample feature vector. The lower the error rate, the better the performance of the training sample feature vector. Indicates the updated The feature vector of the training sample is The weight value in the round iteration, Indicates the updated The feature vector of the training sample is The weight value in the round iteration, Indicates the Whether the feature vector of the training sample is a normal label category value, represents the exponential function, Indicates the Is the label category value of the training sample feature vector The exponential function value of the weak learner's misclassification (the exponential function value is equal to 0 when the classification is correct and the exponential function value is equal to 1 when the classification is wrong);

[0121] Use the test set to perform confusion matrix analysis on the weight values ​​of the updated training sample feature vectors to obtain the feature matrix of the training sample and calculate its anomaly score;

[0122] The confusion matrix is ​​a classification evaluation tool that can be used to show the prediction accuracy of each training sample label category;

[0123] Use the confusion matrix to divide the test set into predicted labels and true labels (in numerical form) for comparison, and calculate the specific value of the confusion matrix;

[0124] The specific values ​​of the confusion matrix include the number of actually negative classes that are correctly predicted as negative classes, the number of actually negative classes that are incorrectly predicted as positive classes (false positives), the number of actually positive classes that are incorrectly predicted as negative classes (false negatives), and the number of actually positive classes that are correctly predicted as positive classes;

[0125] The negative class refers to the class where a certain event or state did not occur. In the confusion matrix here, the negative class means that the weight value of the training sample feature vector is normal, while the positive class refers to the class where a certain event or state occurred. Here, the weight value of the training sample feature vector is abnormal.

[0126] The specific value of the confusion matrix (taking the number of negative classes that are incorrectly predicted as positive classes as an example) can be expressed by the following formula:

[0127] ;

[0128] in, Indicates the specific value of the number of negative classes that are actually predicted to be positive. represents the number of training sample feature vectors, represents the exponential function, Indicates the The true label of the training sample feature vector is the negative class (normal), express The predicted label of the training sample feature vector is positive (abnormal);

[0129] Combined with the specific values ​​of the confusion matrix, a heatmap is used to visualize the weight values ​​of the updated training sample feature vectors (showing the diversity and completeness of the weight values ​​of the training sample feature vectors) to generate a feature matrix for the training sample.

[0130] Based on the feature matrix of the training samples, the local outlier factor is used to identify the wrong training sample types and make improvements (using the k-fold cross validation method);

[0131] The types of erroneous training samples here include training samples with outliers, repeated training samples, training samples containing noise, etc.

[0132] In this step, by identifying error samples, the prediction accuracy of the equipment failure prediction model is gradually improved. Improving error samples makes the equipment failure prediction model more stable and reliable in actual applications.

[0133] Retrain the equipment failure prediction model based on the improved training sample type until the loss curve stabilizes, thereby obtaining a trained equipment failure prediction model.

[0134] The loss curve tends to be stable, which means that the curve gradually decreases without overfitting, that is, the equipment failure prediction model has converged.

[0135] S4. Obtain fault prediction results.

[0136] The following steps are included:

[0137] The validation set is input into the equipment failure prediction model and deployed on the factory's production equipment. The isolation forest algorithm is introduced to define the path length of the training sample feature matrix (in the isolation forest algorithm, the path length refers to the distance from the root node to the leaf node. The shorter the distance, the higher the anomaly score value will be, and the sample is more likely to be identified as an anomaly point. For each training sample feature matrix, starting from the root node, traversing down along the tree structure of the isolation forest until reaching the final leaf node, the path length of the training sample feature matrix is ​​obtained). The comprehensive failure probability value of the factory's production equipment is calculated, and the expression is:

[0138] ;

[0139] ;

[0140] ;

[0141] in, Represents the standardized failure probability value of factory production equipment, represents the activation function, represents the weight of the training sample feature matrix, represents the training sample feature matrix, Represents the bias term, which is used to adjust the baseline level of the training sample feature matrix. It represents the abnormal score of the training sample feature matrix, which is used to measure whether the score of the training sample feature matrix belongs to the normal range or the abnormal range. represents the path length of the training sample feature matrix, Represents the number of training sample feature matrices The expected value of the path length (i.e., the average value of the path length of the training sample feature matrix), Represents the comprehensive failure probability value of factory production equipment, The weight coefficient representing the preliminary failure probability value of the factory production equipment, The weight coefficient representing the abnormality score of the training sample feature matrix;

[0142] Set an alarm threshold (0.2-0.8, based on the false positives and false negatives of the confusion matrix) and determine the risk of equipment failure based on whether the comprehensive failure probability of the factory equipment falls within the threshold.

[0143] When the comprehensive failure probability value of factory production equipment is greater than or equal to 0.2 and less than 0.4, the risk level of equipment failure is low. Technicians are arranged to conduct regular equipment inspections and use MATLAB (mathematical software developed by MathWorks) to continuously record and analyze the equipment's operating data during operation. The data is also uploaded to a cloud database for real-time data synchronization.

[0144] When the comprehensive failure probability value of factory production equipment is greater than or equal to 0.4 and less than 0.6, the risk level of equipment failure is medium. The inspection frequency is adjusted from once a week to once every two days to strengthen the monitoring of equipment status. At this time, the equipment is operating normally, but some problems may occur. For example, mechanical parts such as bearings and gears may show slight wear. Use a thermal imager to check the condition of these mechanical parts and the lubrication status. If the lubricant turns black or contains impurities, the lubricant should be replaced immediately.

[0145] When the comprehensive failure probability value of a factory's production equipment is greater than or equal to 0.6 and less than 0.8, the risk level of equipment failure is high. At this point, the equipment has shown serious signs of failure and all production equipment must be immediately stopped and a comprehensive inspection conducted. The equipment may have electrical component failures (such as main control boards and power interfaces) or structural damage (such as cracks in the equipment's frame or supporting structure). In this case, the equipment must be fully disassembled (inspecting windings, fans, and other locations). A vibration analyzer should be used to detect the equipment's vibration patterns to identify unbalanced or loose parts. Oil samples should be collected and sent to a laboratory for analysis (to assess the metal particle content and other contaminants in the oil). Repairs and replacements should be made based on all diagnostic results. If the main production equipment requires extended maintenance, backup equipment should be started to continue operation.

[0146] The determination result is a fault prediction result;

[0147] In this step, the risk level of equipment failure is determined by setting alarm thresholds, effectively managing and maintaining production equipment to ensure safe and stable operation at different risk levels. This approach not only detects and resolves failures in a timely manner, but also further improves equipment reliability and safety through a graded early warning mechanism and targeted maintenance measures.

[0148] S5. Generate a resource scheduling strategy based on the fault prediction results.

[0149] The following steps are included:

[0150] Select Industrial Internet of Things (IoT) devices as edge computing nodes (including sensors, controllers, and gateways);

[0151] The IoT selected in this step can achieve real-time data synchronization with the cloud platform (AWS IoT) through the MQTT transmission protocol. Combined with the powerful computing power of the cloud platform, it can meet the production needs of factories of different sizes and provide a reliable foundation for subsequent resource scheduling.

[0152] Industrial IoT devices and actual production tasks of the current factory are listed as node features and edge features respectively;

[0153] Node characteristics include task load (recording the number and complexity of tasks currently being processed by each edge computing node), CPU usage, memory usage, and network bandwidth;

[0154] Edge features include task dependencies (e.g., one production task requires the result of another production task to continue execution) and communication requirements (evaluating communication overhead, i.e., quantifying the amount of data that needs to be exchanged between tasks);

[0155] In this step, detailed node and edge features can be used to fully understand the status of the entire factory's production equipment. Precise feature definition reduces unnecessary resource waste.

[0156] According to the fault detection results, update the node features and edge features and build the graph structure;

[0157] The reason for updating node and edge features is that once it is detected that certain devices have a higher risk of failure, the task load will be redistributed. For example, if a machine is predicted to fail, some or all tasks on the machine will be transferred to other healthy machines to avoid production interruption.

[0158] In this step, by updating node features and edge features in real time, we can better respond to emergencies and maintain the continuity and efficient operation of equipment production;

[0159] Use the graph neural network model (GNN model) to encode the graph structure and extract the embedding vectors of each node feature and edge feature;

[0160] Encoding process: Each node receives messages from neighboring nodes (including feature information of neighboring nodes);

[0161] Use Gated Recurrent Unit (GRU) to calculate the message between each edge (i.e. the connection between two nodes);

[0162] The messages of neighbor nodes and the messages between each edge are combined and aggregated using the maximum pooling aggregation method to obtain the embedding vector of each node feature and edge feature, which is input into the GNN model for update iteration to generate the final embedding vector of each node feature and edge feature.

[0163] Based on the embedding vectors of each node feature and edge feature, define the optimization objective;

[0164] The reason for defining the optimization objective based on the embedding vectors of each node feature and edge feature is that ordinary feature-based methods cannot fully capture these complex dependencies and contextual information. The GNN model, through multiple rounds of message passing and aggregation, can capture the complex interactions between production equipment and production tasks. For example, the state change of one device will affect the task allocation of other devices, or the change of one production task will also affect the execution order of other production tasks.

[0165] Optimization goals include maximizing load balance (whether tasks are evenly distributed across all devices), minimizing response time (the time it takes from production task submission to production task completion), and minimizing failure risk (minimizing the possibility of device failure).

[0166] According to the defined optimization goal, set the population size and generate the initial population;

[0167] Population size refers to the number of individuals included in each generation of an evolutionary algorithm. Its selection requires a balance between computational complexity and solution diversity. Larger populations provide greater diversity but at a higher computational cost, while smaller populations can lead to premature convergence (i.e., prematurely reaching the optimal solution). An appropriate population size is selected based on the problem scale and computing resources (in the case of resource scheduling, a population of 50 to 200 individuals is selected, with each individual population representing a resource scheduling strategy). These populations are then combined through random initialization to form the initial population.

[0168] This step ensures that the population has sufficient diversity at the initial stage by setting an appropriate population size and using random initialization, thereby increasing the possibility of finding the optimal solution;

[0169] Evaluate each initial population and introduce a fitness function to calculate the fitness value of each initial population on the optimization target;

[0170] The fitness function introduced here is used to evaluate the performance of each individual population on the optimization target (that is, the quality of the individual on a given optimization target), and the fitness function is divided into single-objective fitness function and multi-objective fitness function;

[0171] If we only focus on one optimization goal, we can directly design the corresponding fitness function. For example, for the goal of minimizing response time, the fitness function is the inverse of the task completion time.

[0172] For multiple optimization objectives, a multi-objective fitness function is defined and the weighted summation method is used to calculate the fitness value of each initial population on the optimization objective. The expression is:

[0173] ;

[0174] ;

[0175] ;

[0176] in, represents the fitness value of each initial population in maximizing load balance, represents the number of node features, Indicates the The load of the node feature, represents the average load of all node features, represents the fitness value of each initial population in minimizing the response time, Indicates the total number of production tasks, Indicates the The completion time of a task, represents the fitness value of each initial population in minimizing the failure risk, Indicates the comprehensive failure probability value of factory production equipment;

[0177] According to the fitness value of each initial population on the optimization target, each population is non-dominated sorted (sorted according to the performance of individual populations on each optimization target) and the crowding distance is calculated (measures the density of individual populations in their non-dominated layer), and some excellent individuals are obtained as parent populations;

[0178] Use Pareto dominance to perform non-dominated sorting: Suppose there are two individual populations, one is X and the other is H. If individual X is not inferior to individual H in all optimization objectives and is superior to individual H in at least one objective, then individual X is said to dominate individual H. Otherwise, the two individuals are in a non-dominated relationship. First, find all non-dominated individuals to form the first layer (Front1), then continue to find non-dominated individuals from the remaining individuals to form the second layer (Front2), and so on, until all individuals are assigned to the corresponding layer;

[0179] Calculation process of crowding distance: Assume that the fitness value of individual H in the optimization target is , then its crowding distance on the target ( represents the crowding distance) is calculated by the fitness difference of adjacent individuals, and the final crowding distance is the sum of the crowding distances of each target;

[0180] The meaning of using some excellent individuals as parent populations is that individuals in the earlier non-dominated layers are excellent individuals (if they are in the same layer, individuals with larger crowding distance are selected);

[0181] Select V (V=3) crossover points from the parent population and use the binary crossover algorithm to exchange gene segments to generate offspring populations;

[0182] The three crossover points selected here mean that a genetic algorithm is used to select three positions in the gene sequence of an individual for crossover operations (simulating the genetic recombination process of an organism). Dynamic adjustments can also be made, namely single-point crossover (selecting a random position as the crossover point, truncating the gene sequences of the two parent individuals at this position and exchanging the fragments) and multi-point crossover (selecting multiple random positions as crossover points, truncating and exchanging the gene fragments respectively);

[0183] Assume there are two parent individual populations (e and f), first represent their gene segments as binary strings: e=101010, f=110011;

[0184] If the crossover point is selected at position 3, the offspring population after exchange will be e1=101011, f1=110010;

[0185] Merge the parent population and the offspring population, select a set number of individual populations (depending on the convergence speed of the individual population; when the convergence speed is slow, the number of individual populations can be set between 50 and 100; when the convergence speed is fast, it can be set between 100 and 300) and iterate until the population evolution completes convergence to form the final population;

[0186] The final population formed in this step can ensure that the quality of each generation of population gradually improves, and quickly promotes the genetic algorithm to converge to the optimal solution;

[0187] Extract the Pareto frontier optimal solution from the final population as the resource scheduling strategy;

[0188] The Pareto front optimal solution refers to the set of solutions in a multi-objective optimization problem where no objective can be further improved without compromising other objectives, and these solutions constitute the optimal solution set;

[0189] Assume there is a smart factory with five production machines (Z1, Z2, Z3, Z4, Z5) and ten production tasks (M1, M2, ..., M10, where the proportion of production resources occupied by these ten production tasks is the same).

[0190] Case 1: First, production tasks are classified according to the estimated execution time. Then, tasks are assigned based on the current load of the production equipment. For example, if equipment Z1 is currently lightly loaded, tasks M1, M4, and M5 are assigned; if equipment Z2 is currently moderately loaded, tasks M2, M6, and M9 are assigned; if equipment Z3 is currently heavily loaded, tasks M3 is assigned; and tasks M7, M8, and M10 are assigned to equipment Z4 and Z5, respectively. At the same time, the equipment load is monitored in real time. If a certain equipment is found to be overloaded (such as Z3), some tasks are reallocated to other equipment with lower loads (such as Z1 or Z2).

[0191] Case 2: A task dependency graph is constructed using the task dependency relationships of node features and edge features (each node represents a device, and each edge represents the dependency relationship between two tasks).

[0192] After the task dependency graph is established, the GNN model is used to aggregate the dependency relationships between devices and tasks, extract the embedding vectors of each node and each edge, and sort them (generating the initial population - evaluating the fitness value - selecting some excellent individuals - exchanging gene segments - generating offspring populations - forming the final population - finding the Pareto frontier optimal solution);

[0193] The purpose of sorting is to make the dependencies between devices and tasks clearer in the location of the task dependency graph;

[0194] Assume that the dependency relationships between production tasks in the task dependency graph are: M1->M4, M2->M7, M5->M8, M6->M9 (that is, task M9 depends on the result of task M6, and the same applies to the previous ones). Then the specific situation of the production task allocation equipment is: M1 is assigned to equipment Z1, M2 is assigned to equipment Z2, M3 is assigned to equipment Z3, M4 is assigned to equipment Z1 (because M4 depends on Z1), M5 is assigned to equipment Z4, M6 is assigned to equipment Z5, M7 is assigned to equipment Z2 (because M7 depends on Z2), M8 is assigned to equipment Z4 (because M8 depends on Z5), M9 is assigned to equipment Z5 (because M9 depends on Z6), and M10 is assigned to equipment Z3.

[0195] This embodiment also provides a cloud-based smart factory management system, including:

[0196] Preprocessing module, collects production equipment data and environmental data, and performs preprocessing;

[0197] Feature extraction module, which extracts the features of pre-processed production equipment data and environmental data to form a feature data set;

[0198] The fault prediction module builds an equipment fault prediction model based on the feature data set and obtains the fault prediction results;

[0199] The policy generation module generates resource scheduling policies based on fault prediction results.

[0200] This embodiment also provides a computer device suitable for the case of a smart factory management method based on a cloud platform, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the smart factory management method based on a cloud platform proposed in the above embodiment.

[0201] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0202] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the cloud platform-based smart factory management method proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0203] In summary, the present invention extracts features from pre-processed production equipment data and environmental data to form a feature dataset, constructs an equipment failure prediction model, obtains failure prediction results, and generates a resource scheduling strategy based on the failure prediction results. The constructed equipment failure prediction model can detect potential equipment failure risks at an early stage, reduce downtime, and improve the overall operational efficiency and safety of the factory. The resource scheduling strategy realizes the dynamic optimization and scheduling of resources within the factory, which not only improves production efficiency, but also significantly reduces energy consumption and failure risks. It can also flexibly respond to changes in demand for different production tasks, thereby improving overall economic benefits.

[0204] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A cloud platform-based smart factory management method, characterized by: include, Collect production equipment data and environmental data and perform pre-processing; Extract the features of pre-processed production equipment data and environmental data to form a feature data set; Based on the feature data set, an equipment fault prediction model is constructed to obtain the fault prediction results; Generate resource scheduling strategies based on fault prediction results; The device failure prediction model is constructed based on the feature data set, including the following steps: Divide the feature dataset into training set, test set and validation set; Selecting convolutional neural networks as the architecture for building equipment failure prediction models; Use the training set and Bayesian optimization method to randomly perturb the equipment failure prediction model to generate training samples; Use the sliding window technique to adjust the subsequence labels in the training samples, generate fault category labels and combine them into training sample feature vectors; Use the boosting ensemble learning framework to assign initial weights to each training sample feature vector and calculate the initial weight value and error rate of each training sample feature vector; Use the weak learner to iterate the initial weight value of each training sample feature vector to obtain the updated weight value of the training sample feature vector, which is expressed as: ； ； in, represents the initial weight value of each training sample feature vector, Indicates the The error rate of the feature vector of the training samples, Indicates the updated The feature vector of the training sample is The weight value in the round iteration, Indicates the updated The feature vector of the training sample is The weight value in the round iteration, Indicates the The fault category label value of the training sample feature vector, represents the exponential function, Indicates the Is the fault category label value of the training sample feature vector The exponential function value of the misclassification of the weak learner; Use the test set to perform confusion matrix analysis on the weight values ​​of the updated training sample feature vectors to obtain the feature matrix of the training sample and calculate its anomaly score; Draw the loss curve based on the feature matrix of the training sample, use the local outlier factor to identify the wrong training sample type, and make improvements; Retrain the equipment failure prediction model based on the improved training sample type until the loss curve stabilizes, thereby obtaining a trained equipment failure prediction model. The method of obtaining the fault prediction result includes the following steps: The validation set is input into the equipment failure prediction model and deployed on the factory's production equipment. The isolation forest algorithm is introduced to define the path length of the training sample feature matrix and calculate the comprehensive failure probability value of the factory's production equipment. The expression is: ； ； ； in, Represents the standardized failure probability value of factory production equipment, represents the activation function, represents the weight of the training sample feature matrix, represents the training sample feature matrix, represents the bias term, represents the abnormal score of the training sample feature matrix, represents the path length of the training sample feature matrix, Represents the number of training sample feature matrices The expected value of the path length is Represents the comprehensive failure probability value of factory production equipment, The weight coefficient representing the preliminary failure probability value of the factory production equipment, The weight coefficient representing the abnormality score of the training sample feature matrix; Set alarm thresholds and determine the risk of equipment failure based on whether the comprehensive failure probability of factory equipment falls within the threshold. The judgment result is a fault prediction result.

2. The cloud platform-based smart factory management method according to claim 1, characterized in that: The production equipment data includes temperature, pressure, vibration, rotation speed, load and production rate; The environmental data includes air quality and indoor temperature and humidity.

3. The cloud platform-based smart factory management method according to claim 2, characterized in that: The pretreatment comprises the following steps: Use a mean filter to remove noise from production equipment data and perform normalization; Outliers in the environmental data were identified and removed using the interquartile range, and missing data points were filled using linear interpolation.

4. The cloud platform-based smart factory management method according to claim 3, characterized in that: Extract the features of pre-processed production equipment data and environmental data to form a feature data set. The following steps are included: Select the multilayer perceptron network as the feature extraction method; The pre-processed production equipment data and environmental data are converted into a columnar storage file format and input into a multi-layer perceptron network for forward propagation to obtain features representing the status of the production equipment; Use principal component analysis to map the high-dimensional features of the production equipment status into a low-dimensional space to obtain the features of the production equipment status after dimensionality reduction; The features of the production equipment status after dimensionality reduction are stored in a fast read and write format to form a structured feature data set.

5. The cloud platform-based smart factory management method according to claim 4, characterized in that: Based on the fault prediction results, a resource scheduling strategy is generated, which includes the following steps: Select industrial IoT devices as edge computing nodes; Industrial IoT devices and actual production tasks of the current factory are listed as node features and edge features respectively; According to the fault detection results, update the node features and edge features and build the graph structure; Use a graph neural network model to encode the graph structure and extract the embedding vectors of each node feature and edge feature; Based on the embedding vectors of each node feature and edge feature, define the optimization objective; The optimization objectives include maximizing load balancing, minimizing response time, and minimizing failure risk; According to the defined optimization goal, set the population size and generate the initial population; Evaluate each initial population and introduce a fitness function to calculate the fitness value of each initial population on the optimization target; According to the fitness value of each initial population on the optimization target, each population is non-dominated and sorted, and the crowding distance is calculated to obtain some excellent individuals as the parent population; Select V crossover points from the parent population and use the binary crossover algorithm to exchange gene segments to generate offspring populations; Merge the parent population and the offspring population, select a set number of individual populations and iterate until the population evolution completes convergence to form the final population; The Pareto front optimal solution is extracted from the final population as the resource scheduling strategy.

6. A cloud-based smart factory management system, based on the cloud-based smart factory management method according to any one of claims 1 to 5, characterized in that: include, Preprocessing module, collects production equipment data and environmental data, and performs preprocessing; Feature extraction module, which extracts the features of pre-processed production equipment data and environmental data to form a feature data set; The fault prediction module builds an equipment fault prediction model based on the feature data set and obtains the fault prediction results; The policy generation module generates resource scheduling policies based on fault prediction results.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the cloud platform-based smart factory management method described in any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the cloud platform-based smart factory management method according to any one of claims 1 to 5 are implemented.

Images