An industrial internet of things cloud platform data analysis method

Abstract

The application relates to the technical field of data processing, and discloses an industrial Internet of Things cloud platform data analysis method, which adopts a distributed federated learning model training architecture, realizes distributed model training of a multi-node joint model and industrial Internet of Things data mining, breaks the data resource island problem of a traditional centralized training mode, and aggregates more effective industrial Internet of Things data; meanwhile, the distributed federated learning architecture can prevent leakage of local industrial Internet of Things data of each node, the industrial Internet of Things data is used locally, the model is trained locally, and the industrial Internet of Things data security is effectively ensured.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a data analysis method for an industrial Internet of Things (IoT) cloud platform. Background Technology

[0002] With the rapid development of Industrial Internet of Things (IIoT) technology, an increasing number of industrial devices and sensors are being connected to the network, generating massive amounts of data. This data contains rich information and is of great value for optimizing production processes, improving equipment efficiency, and responding promptly to equipment failures. However, this data is often distributed across different geographical locations and organizations. Traditional centralized data processing methods not only face the risk of data privacy breaches but also struggle to achieve efficient data sharing and utilization due to resource and regulatory constraints. Furthermore, existing data augmentation technologies often fail to fully reflect the complexity of real-world operating environments, leading to insufficient generalization capabilities in model training.

[0003] Chinese invention patent CN202011083175.5 proposes an online monitoring system for the status of power equipment based on the Internet of Things (IoT). This system includes a data acquisition and aggregation platform, and connected sequentially to a data processing platform, a data analysis, deployment and application platform, and an equipment status alarm platform. The system employs technologies and methods such as IoT terminals, data acquisition, intelligent sensing, standard protocol conversion, information fusion, intelligent diagnosis, and monitoring and early warning to achieve data integration and analysis from IoT terminals within substations. It utilizes information interaction, data sharing, data analysis, and standardized interfaces to achieve information interaction with other external systems. Layered, hierarchical, and customized information acquisition methods are used to realize functions such as IoT terminal data acquisition, standard protocol conversion, and data analysis. This system can quickly realize the acquisition and analysis of IoT terminal data within the substation, as well as functions such as equipment status perception and monitoring, equipment status evaluation, and equipment status anomaly early warning.

[0004] Chinese invention patent CN202410439099.9 proposes an industrial IoT gateway and its device health analysis model based on topology data analysis, relating to the IoT field. This IoT gateway includes a high-performance edge computing gateway with communication protocols of 4G / 5G, Ethernet, and WiFi; a low-power data acquisition gateway with communication protocols of 4G / 5G, NB-IoT, and WiFi; a general-purpose wireless data transmission gateway with a communication protocol of 4G Cat.1; and a high-performance OPCUA acquisition gateway with communication protocols of 4G / 5G, WiFi, and Ethernet. The model's working logic includes the following steps: data acquisition and preprocessing; data acquisition, where the gateway actually collects the operating parameters of the devices. This invention is innovative, practical, and operable, and is of great significance for improving equipment operating efficiency, reducing failures, and lowering maintenance costs.

[0005] Chinese invention patent CN202410550168.3 proposes a data analysis method and system for handheld ultrasound devices based on the Internet of Things (IoT). The method includes: initializing device information on the ultrasound device; uploading operational data to the cloud; continuously monitoring the connection status between the ultrasound device and the cloud, as well as detecting the connection status between the ultrasound device and the control terminal, recording connection anomalies and reporting them to the cloud; analyzing the working status of the ultrasound device, predicting the probability of damage and remaining service life; and generating statistical reports after receiving operational data and anomaly information on the cloud. This invention utilizes IoT technology to achieve real-time data collection, transmission, and processing of the device, comprehensively statistically analyzing information such as scanning time, device status, damage, scanning anomalies, and connection anomalies. This helps equipment manufacturers understand the usage of devices scattered across different locations in real time, optimizing equipment management and maintenance processes.

[0006] While the above technical solutions have solved some technical challenges, the following issues still require further resolution:

[0007] 1. In traditional centralized model training, all data needs to be stored and processed centrally, which can easily lead to data privacy leaks. Furthermore, due to the centralized storage of data, the demand for resources is high, which limits the scalability of the model.

[0008] 2. Traditional data augmentation techniques often suffer from insufficient sample diversity, failing to fully simulate the complex data distribution of the real world, resulting in weak generalization ability of the model.

[0009] 3. When training neural networks using standard optimization algorithms, the problem of vanishing or exploding gradients is easily encountered, which affects the training stability and final performance of the model.

[0010] 4. Traditional graph neural networks typically aggregate information only in a single direction, failing to fully utilize global and local data relationships, thus reducing the model's ability to process complex data structures. Summary of the Invention

[0011] To address the shortcomings of existing technologies, this invention provides a data analysis method for an industrial IoT cloud platform. By employing a distributed federated learning model training architecture, it achieves distributed model training and industrial IoT data mining across multiple nodes, breaking down the data resource silos of traditional centralized training models and aggregating more effective industrial IoT data. Simultaneously, the distributed federated learning architecture prevents the leakage of local industrial IoT data on each node; industrial IoT data is used locally, and models are trained locally, effectively ensuring industrial IoT data security.

[0012] To achieve the above objectives, the present invention employs the following technical solution:

[0013] A data analysis method for an industrial IoT cloud platform includes the following steps:

[0014] Acquire industrial IoT data from each node;

[0015] The industrial IoT data from each node is input into a pre-built distributed federated learning model, and the trained model parameters are uploaded to the central server.

[0016] A generative adversarial network algorithm based on quantum state iteration is used to augment the acquired industrial IoT data;

[0017] The expanded industrial IoT data is then input into a neural network model for data mining.

[0018] The mined data is input into a pre-built classifier model for classification prediction to obtain classified data; the classifier model is a graph neural network model based on a bidirectional graph matching strategy.

[0019] Based on the categorized data, the data analysis results of the industrial IoT cloud platform are determined.

[0020] Furthermore, in the model training process of distributed federated learning, the federated averaging algorithm is used as the model aggregation algorithm. The federated averaging algorithm achieves collaborative training of local models through multiple global iterations.

[0021] Furthermore, the specific method of using a generative adversarial network algorithm based on quantum state iteration is as follows:

[0022] Randomly initialize the parameters of the generator and discriminator;

[0023] Update the generator, which takes random noise as input to generate new industrial IoT data samples;

[0024] The discriminator is updated. It receives generated samples from the generator and samples from the real industrial IoT dataset, evaluates the two types of samples, and outputs the probability that the sample is a real sample. When the probability that the generated sample is judged as a real sample by the discriminator is within the range of [45%, 55%], it indicates that the generated sample and the real sample are similar. The perturbation factor of the discriminator is increased to further optimize the quality of the generated sample.

[0025] During adversarial training, the generator and discriminator are updated alternately, each undergoing 100 training rounds.

[0026] The weight parameters of the generator and discriminator are updated using the momentum-based gradient descent method, and the bias parameters of the generator and discriminator are also updated using the gradient descent method.

[0027] Furthermore, the neural network model employs a neural network algorithm based on a self-regulating learning rate.

[0028] Furthermore, the specific method of the graph neural network model based on the bidirectional graph matching strategy is as follows:

[0029] An initial graph is constructed based on the extracted data features;

[0030] The feature vectors extracted from each sample are transformed into low-dimensional vectors through an embedding layer;

[0031] A bidirectional graph matching strategy is used to aggregate information in both directions, from global to local and from local to global.

[0032] The aggregated features are fused, and the fused node features are processed using a conditional random field layer.

[0033] Furthermore, the generator employs a dynamic adjustment mechanism for the probability amplitude of quantum states during the generation process.

[0034] Furthermore, the industrial IoT data at each node originates from the industrial IoT cloud platform, encompassing data from sensors, machinery, and industrial control systems on the production line.

[0035] Furthermore, the initial graph is constructed based on the similarity between devices, and the construction method is represented as follows:

[0036] ;

[0037] In the formula, Indicates the first in the figure The node and the first Edge weights between nodes; It is the bandwidth parameter of the Gaussian kernel, used to control the sensitivity of the edge weights; For the first feature in the dataset after feature extraction The nth sample, i.e., the nth sample in the figure One node; For the first feature in the dataset after feature extraction The nth sample, i.e., the nth sample in the figure One node; Represents an exponential function with the natural constant as its base; It is an L2 norm.

[0038] Furthermore, the conditional random field layer considers the dependencies between nodes and achieves the final classification of IoT devices by maximizing the posterior probability, thus optimizing the classification decision process.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] 1. In the data analysis task of the industrial Internet of Things cloud platform, the present invention adopts a distributed federated learning framework. By training the model locally on each node and summarizing and updating the model parameters to the central server, the problem of data privacy leakage and resource limitation in the traditional centralized training mode is solved.

[0041] 2. In the data analysis task of the industrial IoT cloud platform, a generative adversarial network with quantum state dynamic feature enhancement is adopted to optimize the feature distribution of data samples by adjusting the quantum probability amplitude without leaking the original data, thus solving the problems of insufficient number of samples and insufficient sample diversity in the dataset.

[0042] 3. In the data analysis task of the industrial IoT cloud platform, the self-adjusting learning rate optimization algorithm is adopted to optimize the parameter training of the neural network. The sensitivity of adaptive adjustment of the learning rate is used to enhance the model's response to the initial conditions of the parameters, which solves the problem of gradient vanishing or gradient explosion that may occur during model training.

[0043] 4. In the data analysis task of the industrial IoT cloud platform, by adopting a bidirectional graph matching strategy in the graph neural network, information is aggregated in both directions from global to local and from local to global, which solves the problem that unidirectional information aggregation cannot fully capture the relationships between complex nodes. Attached Figure Description

[0044] Appendix Figure 1 This is a flowchart of the present invention;

[0045] Appendix Figure 2 This invention provides a model training architecture for distributed federated learning.

[0046] Appendix Figure 3 These are the ablation experiment results of the bidirectional aggregation mechanism of the present invention, where (a) is a comparison experiment result of the five models in terms of classification accuracy, and (b) is a comparison experiment result of the five models in terms of training time.

[0047] Appendix Figure 4 This is a diagram illustrating the impact of industrial scenario complexity on performance.

[0048] Appendix Figure 5 Noise robustness comparison experiment diagram. Detailed Implementation

[0049] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0050] This invention proposes a data analysis method for an industrial IoT cloud platform. Industrial IoT data analysis tasks require a large amount of effective training data for feature extraction, analysis, and learning. This invention employs a distributed federated learning model training architecture to achieve distributed model training and industrial IoT data mining across multiple nodes. This breaks down the data resource silos of traditional centralized training models, aggregating and acquiring more effective industrial IoT data. Simultaneously, the distributed federated learning architecture prevents the leakage of local industrial IoT data on each node; industrial IoT data is used locally, and models are trained locally, effectively ensuring industrial IoT data security.

[0051] To facilitate understanding of this embodiment, a method for constructing a device fault diagnosis model based on a small sample size, as disclosed in this embodiment of the invention, will first be described in detail. Figure 1 A flowchart of a data analysis method for an industrial Internet of Things (IoT) cloud platform disclosed in an embodiment of the present invention is shown, such as... Figure 1 As shown, the data analysis method of this industrial IoT cloud platform includes the following steps:

[0052] S1. Acquire industrial IoT data from each node;

[0053] The industrial IoT data collected at each node of this invention originates from an industrial IoT cloud platform, encompassing data from sensors, machinery, and industrial control systems on the production line. The collected industrial IoT data reflects the machine's operating status, environmental parameters, and key indicators during the production process. All industrial IoT data is stored in structured JSON format. In one embodiment, the attributes of the industrial IoT data include: temperature (reflecting the temperature value of the equipment or environment); humidity (the humidity level of the environment); pressure (the pressure index during equipment operation); vibration level (the degree of vibration generated during machine operation); energy consumption (the energy consumption of the equipment during operation); running time (the cumulative running time of the equipment); production speed (the working speed of the production line); failure rate (the frequency of equipment failures); maintenance records (the historical record of equipment maintenance); and safety status (the safe operating status of the machine).

[0054] It should be noted that this embodiment is only to illustrate one industrial IoT data format and type of the present invention. In practical applications, industrial IoT data usually has more than 10 attributes, and the number of attributes of industrial IoT data may reach dozens or even hundreds.

[0055] S2. Input the industrial IoT data from each node into the pre-built distributed federated learning model, and upload the trained model parameters to the central server.

[0056] In a federated learning framework, nodes do not need to share the industrial IoT data of their locally trained models. Instead, they train their local models and send the updated models to a centralized learning unit for aggregation. The model training architecture of distributed federated learning is as follows: Figure 2 As shown:

[0057] In the distributed federated learning model training architecture, each node trains the model locally in each iteration and uploads the trained model parameters to the central server. The central server aggregates and updates the parameters and then distributes the updated parameters to each node to start a new iteration until the training converges.

[0058] In the distributed federated learning model training architecture, the federated averaging algorithm is used as the model aggregation algorithm. The federated averaging algorithm achieves collaborative training of local models through multiple global iterations. Specifically, for each global iteration, let the number of nodes be... The total number of samples possessed is And the first The number of samples per node is The objective function of federated learning is defined as follows:

[0059] ;

[0060] ;

[0061] In the formula, The objective function for federated learning is... For model parameters, For model parameters For the Loss prediction for a single industrial IoT data sample For training loss, For the first Sample characteristics of an industrial IoT data sample No. Labels for individual industrial IoT data samples. Preferably, the training loss... The cross-entropy loss function is used for calculation.

[0062] Furthermore, for the first For a given node, the objective function for that node is defined as:

[0063] ;

[0064] In the formula, For the first The number of samples per node For the first The objective function for each node, For the first Industrial IoT data distribution across nodes.

[0065] In one embodiment, with the first During the nth iteration, the 1st Taking the parameter update method of the nth node as an example, let the nth node be... The gradient of the parameters of each node is Then in the first The way the model updates parameters in the next iteration is represented as follows:

[0066] ;

[0067] In the formula, For the first Model parameters for the next iteration For the first Model parameters for the next iteration The learning rate is updated for the current parameters. For the first The number of samples per node.

[0068] Furthermore, the parameter update method for the global model of the central server is represented as follows:

[0069] ;

[0070] In the formula, For the first The parameters of the global model of the central server in the next iteration. For the first The iteration of the ... Model parameters for each node.

[0071] Furthermore, repeating this iterative operation signifies that the training of the global model on the central server and the models on each node is complete. In one embodiment, the preset stopping condition for iteration is reaching a preset maximum number of iterations; preferably, the preset maximum number of iterations is set to 5000.

[0072] S3. The acquired industrial IoT data is augmented using a generative adversarial network algorithm based on quantum state iteration.

[0073] The acquisition, annotation, and preprocessing of training data for Industrial Internet of Things (IIoT) are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor model generalization ability and affect model accuracy. This invention employs a quantum state-based generative adversarial network (GAN) algorithm for sample generation, thereby expanding IIoT data. The GAN algorithm consists of two parts: a generator and a discriminator. Building upon traditional GANs, the generator, when generating IIoT data, uses a quantum transcendental dynamic feature enhancement mechanism. Based on the probability amplitude of quantum states, it dynamically adjusts the feature distribution of IIoT data samples, increasing the diversity and complexity of the generated samples.

[0074] Specifically, the generative adversarial network algorithm based on quantum state iteration is as follows:

[0075] S301. Randomly initialize the parameters of the generator and discriminator, assuming the generator's weight is... The generator's bias is The weights of the discriminator are The bias of the discriminator is The initialization method is expressed as:

[0076] ;

[0077] ;

[0078] In the formula, This means that the mean is 0. The variance follows a normal distribution. It follows a normal distribution; This indicates that it conforms to a specific distribution. Preferably, Set it to 0.1.

[0079] S302. In each iteration, the generator is first updated. The generator receives random noise as input and generates new industrial IoT data samples. The distribution of the random noise is as follows:

[0080] ;

[0081] In the formula, This represents the Gaussian noise vector input to the generator; It is the identity matrix; Represents an identity matrix with a mean of 0. The variance follows a normal distribution.

[0082] Furthermore, the generated new industrial IoT data samples have the same attributes as real industrial IoT data samples. These attributes include: temperature (reflecting the temperature value of the equipment or environment); humidity (the humidity level of the environment); pressure (pressure indicators during equipment operation); vibration level (the degree of vibration generated during machine operation); energy consumption (the energy consumption of the equipment during operation); running time (the cumulative running time of the equipment); production speed (the working rate of the production line); failure rate (the frequency of equipment failures); maintenance records (the historical record of equipment maintenance); and safety status (the safe operating status of the machine). The generated data is represented as follows:

[0083] ;

[0084] In the formula, It is the hyperbolic tangent function; This is a sample of the generated industrial IoT data; These are coherence control parameters; It is a quantum coherence function; Indicates that the input is The quantum coherence function. Preferably, Set it to 0.1.

[0085] Industrial IoT data is typically complex and diverse. To generate data samples with similar attributes, the generator employs a dynamic adjustment mechanism of quantum state probability amplitude during the generation process. This optimizes the feature distribution of the data samples, making the generated data closer to the statistical characteristics of real industrial data and improving the diversity and complexity of the samples. For example, by adaptively calculating the quantum coherence function, industrial data with diverse attributes such as temperature, humidity, and pressure can be generated, thereby improving the model's generalization ability. In one embodiment, the quantum coherence function is used to dynamically adjust the feature distribution of the generated samples, and its calculation method is expressed as follows:

[0086] ;

[0087] In the formula, It is the first The phase of each qubit is obtained through the interaction of qubits in the quantum computing process; It is the first noise vector One portion, It is the total number of qubits.

[0088] S303. Update the discriminator. The discriminator receives generated samples from the generator and samples from the real industrial IoT dataset, evaluates the two types of samples, and outputs the probability that it is a real sample. When the probability that the generated sample is judged as a real sample by the discriminator is within the range of [45%, 55%], it indicates that the generated sample and the real sample are similar. Increase the perturbation factor of the discriminator to further optimize the quality of the generated sample, thereby more accurately simulating the real characteristics of industrial IoT data, expressed as:

[0089] ;

[0090] ;

[0091] In the formula, It is the Sigmoid activation function. For real samples, To generate samples; This represents the discriminator function; This is the discriminator perturbation factor, used to enhance or reduce the perturbation effect on the generated samples; This is the disturbance adjustment function; Indicates that the input is The discriminator function; Indicates that the input is The discriminator function. Preferably, Set it to 0.2.

[0092] In one embodiment, the perturbation adjustment function employs dynamic adjustment based on the difference in discrimination results, enabling the discriminator to dynamically adjust its feedback strength according to the similarity between the generated sample and the real sample, thereby improving discrimination efficiency and accuracy. The calculation method is expressed as follows:

[0093] ;

[0094] In the formula, It is the strength of the disturbed foundation. This represents the difference in output between the generated sample and the real sample in the discriminator. Preferably, Set to 0.005.

[0095] S304. During adversarial training, the generator and discriminator are updated alternately, each undergoing 100 training rounds. Through continuous adversarial training and coordination, both models are driven towards optimization. Specifically, the generator strives to improve the quality of generated samples to deceive the discriminator, while the discriminator works to distinguish generated samples from real samples, improving its discrimination ability. The adversarial training of the generator and discriminator is achieved by minimizing the loss function, expressed as:

[0096] ;

[0097] ;

[0098] In the formula, This represents the generator's loss function, which updates the generator parameters by maximizing the probability that a generated sample is judged as real. The loss function of the discriminator is defined by maximizing the probability of correctly identifying real samples while minimizing the probability of incorrectly identifying generated samples. The information entropy of the generated sample is used to assess sample diversity; Cosine similarity represents the similarity between real and generated samples, used to assess how close they are in the feature space.

[0099] Furthermore, the information entropy of the generated samples, based on the measure of information entropy, uses the diversity of generated samples as a loss term for the generator to improve the model's exploration ability. The measurement of information entropy enables the generator to explore more diverse sample distributions, avoiding overly homogeneous generated samples. Cosine similarity ensures the closeness of generated samples to real samples in the feature space, thereby reducing overfitting. For example, the generated device runtime or energy consumption data is both diverse in distribution and similar to real data, further improving the data augmentation effect. The calculation method for the information entropy of the generated samples is expressed as follows:

[0100] ;

[0101] In the formula, These are the weights of the control terms, adjusting the proportion of the information entropy term in the loss function; This represents the probability distribution of the generated samples, estimated using the probabilities of the samples generated by the generator. Preferably, Set it to 0.1.

[0102] Furthermore, the cosine similarity between real samples and generated samples is a regularization term based on sample similarity to reduce overfitting and improve discrimination performance. The calculation method is expressed as follows:

[0103] ;

[0104] In the formula, It is the weight of the regularization term, used to control the impact of similarity measurement on the loss function; It is the L2 norm. Preferably, Set to 0.3;

[0105] S305. The generator and discriminator weight parameters are updated using a momentum-based gradient descent method, and the bias parameters of the generator and discriminator are also updated using gradient descent. This method addresses the high-dimensionality of industrial IoT data by updating the generator and discriminator parameters more efficiently. By optimizing the momentum term of the weights, the model can converge quickly even under complex data characteristics, improving training efficiency. The update method is expressed as follows:

[0106] ;

[0107] ;

[0108] ;

[0109] ;

[0110] In the formula, This indicates a parameter update operation. This represents the learning rate of the generative adversarial network; It is the momentum term of the generator weights. It is the momentum term of the discriminator weights; It is the gradient of the generator's loss function with respect to the generator's bias parameters. It is the gradient of the discriminator's loss function with respect to the discriminator's bias parameters. Preferably, Set to 0.03.

[0111] Furthermore, the update method for the momentum terms of the generator and discriminator weights is expressed as follows:

[0112] ;

[0113] ;

[0114] In the formula, It is the cumulative momentum term. It is the momentum decay parameter; It is the gradient of the generator's loss function with respect to the generator's weight parameters. This is the gradient of the discriminator's loss function with respect to the discriminator's weight parameters. Preferably, Set it to 0.95.

[0115] S306. Repeat the above steps until a preset stopping iteration condition is met, indicating that the model training is complete. Through adversarial training, the generated production speed samples are not only diverse but also have realistic industrial distribution characteristics, effectively improving the model's predictive performance. In one embodiment, the preset stopping iteration condition is reaching a preset maximum number of iterations; preferably, the preset maximum number of iterations is set to 1000.

[0116] S4. Input the expanded industrial IoT data into the neural network model for data mining;

[0117] The expanded industrial IoT data is used for industrial IoT data mining through a neural network model, specifically for feature extraction. This invention employs a 6-layer fully connected neural network for feature extraction. In existing technologies, some solutions using neural networks for feature extraction may encounter problems such as vanishing gradients, exploding gradients, or getting trapped in local optima in certain neural network structures, affecting training stability and model performance. This invention uses a self-regulating learning rate optimization algorithm to optimize the parameters of the neural network model to achieve optimal training. Specifically, this invention employs a self-regulating learning rate mechanism, allowing the network to dynamically adjust its weight update strategy based on the error of the previous training cycle, enhancing the model's responsiveness to changes in initial conditions and parameters.

[0118] Specifically, the neural network algorithm method based on self-adjusting learning rate is as follows:

[0119] S401. Initialize the weights and biases of the neural network. In one embodiment, the initialization is expressed as follows:

[0120] ;

[0121] ;

[0122] In the formula, These are the initial values ​​for the weights of the neural network; These are the initial values ​​for the biases of the neural network; This represents the mean offset of the weights in the neural network. This represents the mean offset of the weights in the neural network. It is the standard deviation of the weight initialization in the neural network; This represents a normal distribution with a mean of 0 and a standard deviation equal to the identity matrix. It is an identity matrix. Preferably, Set to 0.01, Set to 0.05, Set to 0.001.

[0123] S402. During training, the output error of the neural network model is evaluated. To address the complex issues arising from the high dimensionality, nonlinearity, and diversity of industrial IoT data, a self-adjusting learning rate is used to avoid overfitting or underfitting, helping the neural network escape local optima. This also effectively alleviates the gradient vanishing and gradient exploding problems found in traditional methods. For example, through a mechanism of dynamically adjusting the learning rate, the model can automatically optimize the learning rate based on error changes, thus adapting more flexibly to the characteristics of industrial IoT data. The dynamic update method of the self-adjusting learning rate is expressed as follows:

[0124] ;

[0125] In the formula, For the first The learning rate of the neural network in each iteration; For the first The learning rate of the neural network in each iteration; It is a regulator of the neural network, controlling the sensitivity to changes in the learning rate; It represents the change in error of the neural network in the previous cycle.

[0126] Furthermore, the weights are updated based on the updated learning rate, as follows:

[0127] ;

[0128] In the formula, It is the first The weights of the neural network in the next iteration; It is the first The weights of the neural network in the next iteration; It is the error of the neural network, which is calculated by the preset Softmax function on the output features of the last layer of the neural network; The symbol represents the partial derivative.

[0129] S403. During training, a regularization strategy is employed to prevent overfitting and ensure the model's generalization ability. Specifically, regularization sparsity is achieved by dynamically adjusting the number of nodes in the hidden layers of the neural network, as follows:

[0130] ;

[0131] In the formula, It is the number of nodes in the hidden layer of the neural network in the i-th iteration; It is the first The number of nodes in the hidden layer of the neural network in each iteration; It refers to the step size adjustment of nodes in the hidden layer of a neural network; It is a sign function of the error change, used to determine whether to increase or decrease the number of nodes; that is, for ,when When greater than zero, When it is 1, When less than or equal to zero, It is -1.

[0132] In one embodiment, the node adjustment step size of the neural network hidden layer depends on a function of the network hierarchy, and the calculation method is expressed as follows:

[0133] ;

[0134] In the formula, This is a rounding operation; It is the influence coefficient of adjusting the step size; This is the depth of the current network layer. Preferably, The value is 6. Set it to 2.5.

[0135] S404. Repeat the above steps until a preset stopping iteration condition is met, indicating that the model training is complete. In one embodiment, the preset stopping iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000.

[0136] S5. Input the mined data into a pre-built classifier model for classification prediction to obtain classified data; the classifier model is a graph neural network model based on a bidirectional graph matching strategy.

[0137] Data is used to make decisions through a classifier model, that is, the predicted category of the data is obtained using the classifier model. This invention uses a graph neural network model based on a bidirectional graph matching strategy as the classification algorithm model. Traditional graph neural networks usually aggregate information of neighboring nodes only in a single direction. This invention uses bidirectional graph matching, that is, it aggregates information from global to local and from local to global simultaneously, in order to capture the complex relationships between nodes more comprehensively.

[0138] Specifically, the training process of the graph neural network algorithm based on the bidirectional graph matching strategy is as follows:

[0139] S501. Construct an initial graph based on the extracted data features. Each node represents a device, and edges represent connections or data similarities between devices. Let the extracted dataset be... The graph is constructed based on the similarity between devices, and the construction method is represented as follows:

[0140] ;

[0141] In the formula, Indicates the first in the figure The node and the first Edge weights between nodes; It is the bandwidth parameter of the Gaussian kernel, used to control the sensitivity of the edge weights; For the first feature in the dataset after feature extraction The nth sample, i.e., the nth sample in the figure One node; For the first feature in the dataset after feature extraction The nth sample, i.e., the nth sample in the figure One node; Represents an exponential function with the natural constant as its base; It is the L2 norm. Preferably, Set it to 0.1.

[0142] S502. Utilizing the features of the nodes, i.e., the feature vectors extracted from each sample, these features are transformed into low-dimensional vectors through an embedding layer, represented as:

[0143] ;

[0144] In the formula, It is the first The embedding vector of each node. and These are the weights and biases of the embedding layer, respectively. This is the ReLU activation function.

[0145] S503. Perform bidirectional information aggregation. Traditional graph neural network models only aggregate neighbor information in one direction. However, industrial IoT data often exhibits both global and local dependencies. This invention uses a bidirectional graph matching strategy to aggregate information in both global-to-local and local-to-global directions. This allows for a more comprehensive capture of the complex topological relationships between devices. For example, in an industrial production line, a device may be affected by both neighboring devices (local) and the overall state of the entire production line (global). Through bidirectional information aggregation, the model can effectively capture these hierarchical dependencies, improving classification accuracy. During the global-to-local aggregation process, information is propagated throughout the entire graph. Each node aggregates information from all its neighbors, emphasizing the influence of the global context, as shown below:

[0146] ;

[0147] In the formula, Indicates the first in the figure The node and the first Edge weights between nodes;

[0148] Furthermore, in the local-to-global aggregation process, each node only aggregates information from its direct neighbors to capture local features, as shown below:

[0149] ;

[0150] In the formula, and They represent the first The feature vectors after global-to-local and local-to-global aggregation in each iteration. and They represent the first The feature vectors after global-to-local and local-to-global aggregation in each iteration. It is the first The attention-based weights in the next iteration, Represents a set of nodes.

[0151] Furthermore, the first The calculation method for the attention-based weights in the next iteration is expressed as follows:

[0152] ;

[0153] In the formula, The LeakyReLU activation function is used. This is the parameter vector in the attention mechanism, used to calculate the attention weights; for transpose; This is the weight matrix used for node feature transformation in the attention mechanism; Indicates the first The iteration of the ... Embedding vectors of each node; Indicates the first The iteration of the ... Embedding vectors of each node; Indicates the first The iteration of the ... Embedding vectors of each node; This indicates a vector concatenation operation.

[0154] S504. Further, the features obtained from the two aggregation strategies are fused. A Conditional Random Field (CRF) layer is then used to process the fused node features. The CRF layer considers the dependencies between nodes and achieves the final classification of IoT devices by maximizing the posterior probability, thus optimizing the classification decision process. For example, in an industrial cold chain logistics monitoring scenario, the temperature control status of equipment often exhibits strong dependencies (such as a group of connected refrigeration units). The CRF layer models this dependency relationship, making the classification more accurate, as shown below:

[0155] ;

[0156] ;

[0157] In the formula, It is the fused feature vector. and These are the weights and biases of the fusion layer. Given a graph structure and node features, the conditional probability of each node category is represented by the loss function, which minimizes this conditional probability. and represent the feature vectors after aggregation from global to local and from local to global, respectively; where, This represents the feature vector after global-to-local aggregation. This represents the feature vector after local-to-global aggregation; The true label matrix of the samples; This represents the layer function of a conditional random field.

[0158] Furthermore, the weights of the fusion layer are calculated as follows:

[0159] ;

[0160] In the formula, It is an orthogonal matrix, representing the basis of the characteristic space; It is a diagonal matrix, whose diagonal elements represent scaling factors in each direction, used to adjust the sensitivity of feature fusion; yes The transpose of .

[0161] S505. Repeat the above steps until a preset stopping iteration condition is met, indicating that the model training is complete. In one embodiment, the preset stopping iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000.

[0162] S6. Based on the classification data, determine the data analysis results of the industrial IoT cloud platform. Specifically, the industrial IoT data acquired from each node is input into the trained distributed federated learning model, followed by data augmentation and data mining. Furthermore, the processed features are input into the classifier model for classifier training, thereby obtaining the classification results. In this embodiment, the classification categories include the labeled categories, which include different operating states of production line equipment, specifically: "normal operation," "initial fault stage," "mid-fault stage," and "final fault stage."

[0163] To verify the effect shown by this method, as follows: Figure 3As shown, to verify the necessity of the bidirectional information aggregation strategy and the conditional random field fusion layer, the experiment compared five model configurations horizontally: the complete model, the basic model without bidirectional aggregation, the simplified model with only global aggregation, the simplified model with only local aggregation, and the model without conditional random field fusion. Figure (a) shows the comparison experiment of the five models in terms of classification accuracy, with the horizontal axis representing the five models (complete model, basic model without bidirectional aggregation, simplified model with only global aggregation, simplified model with only local aggregation, and model without conditional random field fusion), and the vertical axis representing classification accuracy; Figure (b) shows the comparison experiment of the five models in terms of training time, with the horizontal axis representing the five models (complete model, basic model without bidirectional aggregation, simplified model with only global aggregation, simplified model with only local aggregation, and model without conditional random field fusion), and the vertical axis representing training time (in hours). Experimental results show that the red histogram of the complete model significantly outperforms other blue histogram configurations in terms of accuracy. The performance degradation is most pronounced after removing bidirectional aggregation, demonstrating that the complementarity of global and local information plays a crucial role in modeling topological relationships in the Industrial Internet of Things (IIoT). Although the training time of the complete model in the right-hand graph increases slightly, the improvement in accuracy far outweighs the time cost, highlighting the advantage of bidirectional aggregation in feature learning efficiency.

[0164] like Figure 4 As shown, by simulating the dynamic changes in device connection density and signal strength in a real industrial environment, the adaptability of different algorithms in complex scenarios is evaluated. The horizontal axis uses the product of device density and connection strength as a quantification of scenario complexity, and the vertical axis uses the comprehensive evaluation index F1 score. The red curve of this technique maintains a gradual downward trend as complexity increases, while the dashed curve of the traditional method shows an accelerated downward trend, especially forming a significant performance gap in the high complexity region. The gradient-filled area between the curves represents the stability range of the algorithm's performance. The narrow filled band of this technique indicates its strong robustness to environmental changes. When the horizontal axis exceeds the critical value, the fluctuation amplitude of the traditional method increases significantly, while this technique effectively suppresses the performance oscillations caused by complex topology by adaptively adjusting the information weight through a bidirectional attention mechanism.

[0165] like Figure 5 As shown, to address the common noise interference problem in industrial sensor data, Gaussian noise of varying intensities was added to test the feature stability of the algorithm. The horizontal axis represents the noise standard deviation, and the vertical axis displays the classification accuracy. The four colored curves represent our technique and three mainstream graph neural network methods, respectively. As the noise increases, the red curve of our technique shows a slow, linear decline, while the curves of other methods exhibit a sharp drop in the medium-to-high noise range. The divergence between the curves visually demonstrates the dual advantages of our technique in feature cleaning and error correction.

Claims

1. A data analysis method for an industrial Internet of Things (IoT) cloud platform, characterized in that, Includes the following steps: Acquire industrial IoT data from each node, which originates from the industrial IoT cloud platform and includes data from sensors, machinery, and industrial control systems on the production line. The industrial IoT data from each node is input into a pre-built distributed federated learning model, and the trained model parameters are uploaded to the central server. A generative adversarial network algorithm based on quantum state iteration is used to augment the acquired industrial IoT data; The expanded industrial IoT data is then input into a neural network model for data mining. The mined data is input into a pre-built classifier model for classification prediction to obtain categorical data; The classifier model is a graph neural network model based on a bidirectional graph matching strategy; Based on the aforementioned classification data, the data analysis results of the industrial IoT cloud platform are determined; The specific method of using a generative adversarial network algorithm based on quantum state iteration is as follows: Randomly initialize the parameters of the generator and discriminator; Update the generator, which takes random noise as input to generate new industrial IoT data samples; The discriminator is updated. It receives generated samples from the generator and samples from the real industrial IoT dataset, evaluates the two types of samples, and outputs the probability that the sample is a real sample. When the probability that the generated sample is judged as a real sample by the discriminator is within the range of [45%, 55%], it indicates that the generated sample and the real sample are similar. The perturbation factor of the discriminator is increased to further optimize the quality of the generated sample. During adversarial training, the generator and discriminator are updated alternately, each undergoing 100 training rounds. The weight parameters of the generator and discriminator are updated using the momentum-based gradient descent method, and the bias parameters of the generator and discriminator are also updated using the gradient descent method. The specific steps involved in generating new industrial IoT data samples are as follows: ； In the formula, It is the hyperbolic tangent function; This is a sample of the generated industrial IoT data; These are coherence control parameters; The weights of the generator, This is the bias of the generator. This represents the Gaussian noise vector input to the generator. It is a quantum coherence function. It is the first The phase of each qubit is obtained through the interaction of qubits in the quantum computing process; It is the first noise vector One portion, It is the total number of qubits.

2. The industrial IoT cloud platform data analysis method according to claim 1, characterized in that, In the model training process of distributed federated learning, the federated averaging algorithm is used as the aggregation algorithm for the model. The federated averaging algorithm achieves the collaborative training of local models through multiple global iterations.

3. The industrial IoT cloud platform data analysis method according to claim 1, characterized in that, The neural network model employs a neural network algorithm based on a self-adjusting learning rate.

4. The industrial IoT cloud platform data analysis method according to claim 1, characterized in that, The specific method of the graph neural network model based on the bidirectional graph matching strategy is as follows: An initial graph is constructed based on the extracted data features; The feature vectors extracted from each sample are transformed into low-dimensional vectors through an embedding layer; A bidirectional graph matching strategy is used to aggregate information in both directions, from global to local and from local to global. The aggregated features are fused, and the fused node features are processed using a conditional random field layer.

5. The industrial IoT cloud platform data analysis method according to claim 1, characterized in that, The generator employs a dynamic adjustment mechanism for the probability amplitude of quantum states during the generation process.

6. The industrial IoT cloud platform data analysis method according to claim 4, characterized in that, The initial graph is constructed based on the similarity between devices, and the construction method is represented as follows: ； In the formula, Indicates the first in the figure The node and the first Edge weights between nodes; It is the bandwidth parameter of the Gaussian kernel, used to control the sensitivity of the edge weights; For the first feature in the dataset after feature extraction The nth sample, i.e., the nth sample in the figure One node; For the first feature in the dataset after feature extraction The nth sample, i.e., the nth sample in the figure One node; Represents an exponential function with the natural constant as its base; It is an L2 norm.

7. The industrial IoT cloud platform data analysis method according to claim 4, characterized in that, The conditional random field layer considers the dependencies between nodes and achieves the final classification of IoT devices by maximizing the posterior probability, thus optimizing the classification decision process.

Images