Industrial data intelligent prediction method and system, electronic equipment and storage medium

Abstract

The invention provides an industrial data intelligent prediction method and system, electronic equipment and a storage medium, and the method comprises the steps: receiving a prediction request, and carrying out the analysis of the prediction request, and obtaining a prediction parameter; determining a target Pod matched with the prediction parameter from the Kubernetes cluster, and routing the prediction parameter to the target Pod; reading a plurality of target feature vectors corresponding to target equipment indicated by the prediction request through the target Pod; extracting a target time sequence feature vector and a target space feature vector from the plurality of target feature vectors through a bidirectional fusion model in the Docker mirror image, and performing prediction by using the target time sequence feature vector and the target space feature vector to obtain a prediction result; the bidirectional fusion model is obtained by training a to-be-trained bidirectional fusion model by using multi-source industrial data; and the prediction result is encrypted, and the encrypted prediction result is stored in a result database of the private cloud cluster by using a preset encryption industrial data transmission link.

Description

Technical Field

[0001] This invention relates to the field of industrial and artificial intelligence interdisciplinary technology, and more specifically, to an intelligent prediction method, system, electronic device and storage medium for industrial data. Background Technology

[0002] With the deepening of Industry 4.0, the digital transformation of industrial enterprises is accelerating, resulting in an exponential increase in industrial data, including real-time data from equipment sensors (such as temperature, pressure, and speed), production process data (such as production cycle time and process parameters), and environmental parameter data (such as workshop humidity and dust concentration). This industrial data contains key patterns in enterprise production and operation. Accurate prediction of core data is a crucial means for enterprises to optimize production, control costs, and provide early warnings of risks, directly impacting their market competitiveness.

[0003] Currently, there are two main technical solutions in the field of industrial data prediction: AI prediction solutions based on public clouds and traditional on-premises deployment solutions. AI prediction solutions based on public clouds upload industrial data to a public cloud platform and utilize the cloud's computing resources for AI model training and prediction. However, industrial data contains sensitive information such as enterprise production processes and equipment parameters. During the data transmission process to and storage in the public cloud, there are security risks of theft and leakage, seriously threatening enterprise data security and trade secrets. Traditional on-premises deployment solutions deploy data storage and computing resources on the enterprise's local servers. However, the storage capacity and computing power of local servers are fixed, making it difficult to elastically expand and scale according to the growth of industrial data volume and changes in prediction tasks, resulting in poor resource adaptability. Furthermore, when the prediction task volume is small, a large amount of computing resources are idle, resulting in low resource utilization. This makes it impossible to simultaneously handle the multi-source heterogeneity and spatiotemporal correlation of industrial data, leading to low prediction accuracy and failing to meet the enterprise's requirements for the reliability of prediction results. Summary of the Invention

[0004] In view of this, the present invention provides an intelligent prediction method, system, electronic device and storage medium for industrial data, in order to solve the security risks of theft and leakage in the storage of data in the prior art, and to improve resource utilization and prediction accuracy.

[0005] The first aspect of this application provides an intelligent prediction method for industrial data, the method comprising:

[0006] Receive a prediction request and parse the prediction request to obtain prediction parameters;

[0007] Determine the target Pod that matches the prediction parameters from the Kubernetes cluster, and route the prediction parameters to the target Pod;

[0008] The target Pod reads multiple target feature vectors corresponding to the target device indicated by the prediction request;

[0009] The target temporal feature vector and target spatial feature vector are extracted from multiple target feature vectors using a bidirectional fusion model in a Docker image, and prediction is performed using the target temporal feature vector and the target spatial feature vector to obtain the prediction result; the bidirectional fusion model is obtained by training a bidirectional fusion model to be trained using multi-source industrial data;

[0010] The prediction results are encrypted, and the encrypted prediction results are stored in the result database of a private cloud cluster using a preset encrypted industrial data transmission link.

[0011] Optionally, the bidirectional fusion model to be trained is trained using multi-source industrial data to obtain the bidirectional fusion model, including:

[0012] Collect multi-source industrial data using a data acquisition device;

[0013] The multi-source industrial data is preprocessed to generate a feature vector set, wherein the feature vector set includes multiple feature vectors;

[0014] The feature vector set is divided into a training dataset and a validation dataset;

[0015] The bidirectional fusion model is trained using the training dataset and validated using the validation dataset to obtain the bidirectional fusion model.

[0016] Optionally, after obtaining the bidirectional fusion model, the method further includes:

[0017] The bidirectional fusion model is encapsulated using automated scripts to obtain a Docker image, and a model Deployment is created in a Kubernetes cluster to achieve containerized deployment.

[0018] Optionally, the multi-source industrial data is preprocessed to generate a feature vector set, including:

[0019] Data cleaning is performed on the multi-source industrial data;

[0020] The cleaned multi-source industrial data is standardized to obtain a feature matrix.

[0021] A linear transformation is performed on the feature matrix, and attention is calculated based on the result of the linear transformation to obtain the fraction matrix of the feature matrix;

[0022] Calculate the weights of the fraction matrix, and perform weighted processing based on the weights of the fraction matrix and the result of the linear transformation to obtain the data matrix;

[0023] Extract multiple temporal feature vectors and multiple spatial feature vectors from the data matrix;

[0024] The time-series feature vectors and the spatial feature vectors are filtered to obtain a feature vector set.

[0025] Optionally, the bidirectional fusion model to be trained includes an input layer, a local ID-CNN feature extraction branch, a bidirectional LSTM temporal modeling branch, a branch global feature fusion branch, and an output layer;

[0026] The bidirectional fusion model is trained using the training dataset and validated using the validation dataset to obtain the bidirectional fusion model, which includes:

[0027] The training dataset is received through the input layer;

[0028] The ID-CNN local feature extraction branch captures local temporal features and short-term dependencies from the training dataset.

[0029] The bidirectional LSTM temporal modeling branch captures global temporal features and long-term dependencies from the training dataset;

[0030] The local temporal features, short-term dependencies, global temporal features, and long-term dependencies are fused through the global feature fusion branch to obtain fused features;

[0031] The output layer is trained using the fused features and a pre-constructed joint loss function.

[0032] The trained bidirectional fusion model is validated using the validation dataset to obtain the bidirectional fusion model.

[0033] A second aspect of this application provides an intelligent prediction system for industrial data, the system comprising:

[0034] A receiving unit is used to receive a prediction request and parse the prediction request to obtain prediction parameters;

[0035] A routing unit is used to determine the target Pod that matches the prediction parameters from the Kubernetes cluster, and to route the prediction parameters to the target Pod;

[0036] The reading unit is used to read multiple target feature vectors corresponding to the target device indicated by the prediction request through the target Pod;

[0037] The prediction unit is used to extract target temporal feature vectors and target spatial feature vectors from multiple target feature vectors through a bidirectional fusion model in the Docker image, and to make predictions using the target temporal feature vectors and target spatial feature vectors to obtain prediction results; the bidirectional fusion model is obtained by the training unit training a bidirectional fusion model to be trained using multi-source industrial data;

[0038] The storage unit is used to encrypt the prediction results and store the encrypted prediction results in the result database of the private cloud cluster using a preset encrypted industrial data transmission link.

[0039] Optionally, the training unit includes:

[0040] The data acquisition unit is used to collect multi-source industrial data through a data acquisition device;

[0041] A preprocessing unit is used to preprocess the multi-source industrial data to generate a feature vector set, wherein the feature vector set includes multiple feature vectors;

[0042] The partitioning unit is used to divide the feature vector set into training and validation datasets.

[0043] The training and validation unit is used to train the bidirectional fusion model to be trained using the training dataset, and to validate the trained bidirectional fusion model using the validation dataset, thereby obtaining the bidirectional fusion model.

[0044] Optionally, the system further includes:

[0045] The packaging and deployment unit is used to package the bidirectional fusion model using automated scripts to obtain a Docker image and create a model Deployment in the Kubernetes cluster to achieve containerized deployment.

[0046] A third aspect of this application provides an electronic device, comprising: a processor and a memory, the processor and the memory being connected via a bus; wherein the processor is configured to call and execute a program stored in the memory; and the memory is configured to store the program, the program being configured to implement the industrial data intelligent prediction method provided in the first aspect of this application.

[0047] A fourth aspect of this application provides a computer-readable storage medium storing computer-executable instructions for performing the industrial data intelligent prediction method provided in the first aspect of this application.

[0048] This application provides an intelligent prediction method, system, electronic device, and storage medium for industrial data. The method involves receiving and parsing prediction requests to obtain prediction parameters; determining the target Pod matching the prediction parameters from a Kubernetes cluster and routing the prediction parameters to the target Pod; reading multiple target feature vectors corresponding to the target device indicated by the prediction request through the target Pod; extracting target temporal feature vectors and target spatial feature vectors from the target feature vectors using a bidirectional fusion model in a Docker image, and using these vectors to perform prediction to obtain the prediction result; the bidirectional fusion model is trained using multi-source industrial data; encrypting the prediction result and storing the encrypted prediction result in a result database of a private cloud cluster using a preset encrypted industrial data transmission link. Therefore, the technical means provided in this application, by calling heterogeneous computing cluster resources through a Kubernetes cluster, realizes on-demand allocation of CPU and GPU resources, thereby solving the problems of poor resource adaptability and resource utilization in the prior art. Furthermore, this application pre-trains a bidirectional fusion model using multi-source industrial data to be trained, so that after extracting the target feature vector, the bidirectional fusion model can be used to extract the target temporal feature vector and the target spatial feature vector from the target feature vector, and use the target temporal feature vector and the target spatial feature vector to make predictions, thereby obtaining high-precision prediction results. This solves the problem of low prediction accuracy caused by the inability of the prior art to handle multi-source heterogeneous data and spatiotemporal correlation. The prediction results are decrypted, and during the storage process, the encrypted prediction results are stored in the result database of the private cloud cluster using a preset encrypted industrial data transmission link, realizing full-process data security protection. This solves the problem of security risks of theft and leakage in the prior art, which seriously threatens the data security and trade secrets of enterprises. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0050] Figure 1 An example diagram of an industrial data intelligent prediction architecture provided in this application embodiment;

[0051] Figure 2 A flowchart illustrating an intelligent prediction method for industrial data provided in this application embodiment;

[0052] Figure 3This is a schematic diagram of the structure of an industrial data intelligent prediction system provided in an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0056] To better understand the intelligent prediction method for industrial data provided in this application, the technical terms used in this application are explained below:

[0057] POD: The smallest scheduling unit in a Kubernetes cluster, containing one or more containers with shared network and storage resources. It is the basic unit for deploying and managing applications in a Kubernetes cluster.

[0058] HPA: Horizontal Pod Autoscaler, is a resource auto-scaling component in Kubernetes clusters that can dynamically adjust the number of Pod replicas based on metrics such as CPU utilization and memory usage to cope with load changes and ensure stable service operation.

[0059] LSTM: Long Short-Term Memory, a special type of recurrent neural network (RNN) that solves the gradient vanishing or gradient explosion problems of traditional RNNs through gating mechanisms (input gate, forget gate, output gate). It is designed specifically for processing sequential data and can effectively capture the long-term temporal dependencies of data.

[0060] 1D-CNN: 1D-Convolutional Neural Network, a variant of convolutional neural networks. Convolution operations are performed along the time dimension or sequence dimension. It is specifically designed to process sequential data and is widely used in time series analysis, signal processing, and biomedicine. It can effectively extract local features and spatial correlation features from sequential data.

[0061] See Figure 1 This paper illustrates an industrial data intelligent prediction architecture provided by an embodiment of the present application. The industrial data intelligent architecture includes a presentation layer, a logic processing layer, a data calculation layer, and a data storage layer.

[0062] It should be noted that the presentation layer is used for configuration distribution, parameter updates, model validation, prediction result display, and prediction result feedback. Among them, model validation is to use a validation dataset to validate the trained bidirectional fusion model (the specific validation process can be found in the bidirectional fusion model validation process shown below), prediction result display is to show the prediction results obtained by using the bidirectional fusion model, and prediction result feedback is for users to provide feedback on the actual effect based on the displayed prediction results.

[0063] The logic processing layer is used for data access (receiving and collecting multi-source industrial data), data processing (processing multi-source industrial data to obtain feature vector sets), model training (training and validating the bidirectional fusion model to be trained), model image creation (encapsulating the bidirectional fusion model into a Docker image), containerized deployment, data prediction (receiving prediction requests to perform a series of subsequent operations to obtain prediction results), and pushing prediction results (pushing prediction results to the data computing layer, so that the data computing layer can distribute the prediction results in the data storage layer).

[0064] Data computation layer: This includes a Kubernetes cluster and HPA rules (used to automatically scale up / down the number of Pod replicas). Specifically, a Deployment model is created in the Kubernetes cluster, its YAML file is written, and resource requests and limits for CPU, memory, and storage are set. An HPA YAML file is also written, configuring the maximum and minimum number of replicas in the HPA rules, configuring CPU, memory, and QPS metrics, configuring the scaling-up / scaling-down cooling window, and configuring scaling-up / scaling-down strategies to achieve automatic scaling up and down of pods.

[0065] Data storage layer: The private cloud infrastructure layer adopts a distributed storage cluster + heterogeneous computing cluster architecture (Ceph distributed storage system). Specifically, the distributed storage cluster consists of more than 3 storage nodes and adopts the Ceph distributed storage system, which supports elastic expansion of storage capacity.

[0066] based on Figure 1 The illustrated industrial data intelligent prediction architecture corresponds to an industrial data intelligent prediction method provided in this application embodiment, such as... Figure 2 As shown, this intelligent prediction method for industrial data specifically includes the following steps:

[0067] S201: Receive the prediction request and parse the prediction request to obtain the prediction parameters.

[0068] During the specific execution of step S201, the user can send a prediction request through the operation interface on the presentation layer or other backend services; the prediction service receives the prediction request through the RESTful API and parses the prediction request to obtain the corresponding prediction parameters.

[0069] It should be noted that after receiving a prediction request through the RESTful API, the prediction service can process the prediction request into an HTTP request that conforms to the RESTful specification, and extract the corresponding prediction parameters from the body of the HTTP request. The prediction parameters may include the Podid and the result push method.

[0070] In some embodiments, user identity and device authentication can be performed first at the network layer. Specifically, a zero-trust network core design can be used to design the network layer to authenticate the user's password and dynamic token. If the user's password and dynamic token authentication is successful, the user can log in and send a prediction request through the presentation layer. After receiving the prediction request, attribute-based access control can be used to verify permissions (e.g., determine whether the user's operation belongs to the execution operation level; if so, verify the user's dynamic token again; otherwise, verify the user using the least privilege rule). If the permission verification passes, the user's prediction request is parsed; if the permission verification fails, the prediction request is rejected.

[0071] It should be noted that using the least privilege rule to verify users means granting the necessary predictive requests to complete the current task, or enabling a whitelist. If the user is on the whitelist, the user's permission authentication is confirmed to be successful.

[0072] S202: Determine the target Pod that matches the prediction parameters from the Kubernetes cluster and route the prediction parameters to the target Pod.

[0073] During the specific execution step S202, after the prediction service is called to parse and obtain the prediction parameters, it can determine the target Pod that matches the Podid from each Pod in the data computing layer, and route the prediction parameters to the target Pod through the result push method.

[0074] S203: Read multiple target feature vectors corresponding to the target device indicated by the prediction request through the target Pod.

[0075] In the specific execution step S203, after the target Pod is determined, the latest historical data of the target device bound to the Podid can be read from the feature data of the private cloud cluster through the target Pod, and the latest historical data can be preprocessed to obtain multiple target feature vectors.

[0076] It should be noted that the preprocessing method for the latest historical data is the same as the subsequent preprocessing process for multi-source industrial data. Please refer to the subsequent preprocessing process for multi-source industrial data, which will not be repeated here.

[0077] S204: Extract the target temporal feature vector and the target spatial feature vector from the target feature vector using the bidirectional fusion model in the Docker image, and use the target temporal feature vector and the target spatial feature vector to make predictions to obtain the prediction results; the bidirectional fusion model is obtained by training the bidirectional fusion model to be trained using multi-source industrial data.

[0078] In this embodiment of the application, multi-source industrial data can be collected in advance and divided into training datasets and validation datasets; the bidirectional fusion model to be trained is trained using the validation dataset, and the trained bidirectional fusion model is validated using the validation dataset to obtain the bidirectional fusion model.

[0079] It should be noted that multi-source industrial data can include equipment data (such as data from PLCs, sensors, etc.), system data (such as data from enterprise ERP, MES, etc.), and so on.

[0080] Optionally, the process of training the bidirectional fusion model to be trained using multi-source industrial data can be as follows: collect multi-source industrial data through a data collector; preprocess the multi-source industrial data to generate a feature vector set, wherein the feature vector set includes multiple feature vectors; divide the feature vector set to obtain a training dataset and a validation dataset; train the bidirectional fusion model to be trained using the training dataset, and validate the trained bidirectional fusion model using the validation dataset to obtain the bidirectional fusion model; encapsulate the bidirectional fusion model using an automated script to obtain a Docker image, and create a model Deployment in a Kubernetes cluster to achieve containerized deployment.

[0081] In some embodiments, configuration information can be configured to collect multi-source industrial data using a data acquisition device (which supports edge deployment and can be directly installed in the industrial field) based on the TLS 1.3 protocol.

[0082] It should be noted that collecting multi-source industrial data via a data acquisition device can effectively maintain compatibility with mainstream industrial protocols, covering the data acquisition needs of the vast majority of industrial equipment. These mainstream industrial protocols specifically include the following:

[0083] Industrial control protocols: OPC UA protocol (supports real-time data transmission and device information interaction), Modbus TCP protocol (suitable for data acquisition from PLCs, sensors, and other devices).

[0084] IoT Protocol: MQTT Protocol (lightweight, suitable for device data transmission in low-bandwidth, high-latency scenarios);

[0085] Common network protocols: HTTP / HTTPS protocols (suitable for data interaction with enterprise ERP, MES and other systems).

[0086] In this embodiment of the application, after collecting multi-source industrial data, the multi-source industrial data can be encrypted using the SM4 over-encryption algorithm, and the encrypted multi-source industrial data can be stored in the time-series database partition corresponding to the private cloud cluster according to the pre-configured configuration rules.

[0087] It should be noted that this application uses a private cloud architecture for data storage, which can effectively avoid the risk of leakage of industrial data when uploaded to a public cloud. Furthermore, this application uses the TLS 1.3 protocol and the SM4 over-encryption algorithm to effectively prevent data from being stolen during collection and transmission.

[0088] It should also be noted that the private cloud cluster provided in this application adopts a distributed storage cluster + heterogeneous computing cluster architecture; the distributed storage cluster consists of more than three storage nodes, uses the Ceph distributed storage system, and supports elastic expansion of storage capacity. The heterogeneous computing cluster architecture consists of CPU computing nodes and GPU computing nodes, and uses Kubernetes to achieve unified scheduling and containerized management of computing resources.

[0089] As one implementation of this application, the process of preprocessing multi-source industrial data to generate a feature vector set can be as follows: cleaning the multi-source industrial data; standardizing the cleaned multi-source industrial data to obtain a feature matrix; performing a linear transformation on the feature matrix and calculating attention based on the linear transformation result to obtain a fractional matrix of the feature matrix; calculating the weights of the fractional matrix and performing weighted processing based on the weights of the fractional matrix and the linear transformation result to obtain a data matrix; extracting multiple temporal feature vectors and multiple spatial feature vectors from the data matrix; and filtering each temporal feature vector and multiple spatial feature vectors to obtain a feature vector set.

[0090] In some embodiments, a combination of "rule filtering + isolated forest algorithm" can be used to remove outliers from multi-source industrial data, and a combination of "temporal interpolation + domain knowledge completion" can be used to complete missing values ​​in multi-source industrial data, thereby completing the data cleaning of multi-source industrial data.

[0091] It's important to note that rule-based filtering involves filtering based on physical and business logic rules. Specifically, filtering criteria include value range / physical limits, rate of change / gradient, and consistency of associated parameter states. Custom rules can also be added based on the process flow, allowing for filtering of multi-source industrial data according to these criteria and added rules. The Isolation Forest algorithm, on the other hand, is used to discover potentially complex and anomalous data. It extracts values ​​from correct original numerical features, statistical features, and derived features for model training. The predicted results are compared with actual data, and anomalous data is marked and output. Process engineers then confirm whether the anomalous data should be removed; if so, it is removed from the multi-source industrial data.

[0092] It should also be noted that since most multi-source industrial data is time-series data, a "time-series interpolation + domain knowledge completion" method can be used to complete missing values. Specifically, firstly, time-series interpolation is used to handle short-term missing values. Time-series interpolation methods include linear interpolation, forward imputation, and backward imputation. Then, for parts that time-series interpolation cannot handle (such as long-term missing values) or parts where the interpolation results do not conform to domain knowledge, domain knowledge completion is used. This involves using data correlations, process mechanisms, operating states, and historical data from the same period to complete the data. Finally, the relevant data from the time-series interpolation method and domain knowledge completion are output, confirmed by process engineers, and the confirmed missing value points are recorded, thus completing the missing value completion for multi-source industrial data.

[0093] In some embodiments, after data cleaning is completed, the Min-Max standardization method can be used to standardize the cleaned multi-source industrial data to convert data of different magnitudes (such as temperature ℃, rotation speed r / min) in the cleaned multi-source industrial data into a unified dimension, eliminate the influence of dimensions, and distribute them in the [0,1] interval to obtain a feature matrix; the attention-based fusion algorithm assigns weights according to the contribution of the feature matrix to the prediction results to generate the same data matrix of the related data.

[0094] It should be noted that the feature matrix obtained by standardizing the multi-source industrial data after data cleaning using the Min-Max standardization method is shown in formula (1).

[0095] (1)

[0096] Where n is the number of samples and m is the number of features.

[0097] It should be noted that a sample refers to the values ​​of multiple related devices at the same time.

[0098] A linear transformation is performed on the feature matrix, and the result of the linear transformation can include queries, keys and values, as shown in formula (2).

[0099] (2)

[0100] Where Q represents the query, K represents the key, V represents the value, and W represents the corresponding weight coefficient.

[0101] The query and key of the feature matrix are multiplied by a dot product to obtain the fractional matrix of the feature matrix, as shown in formula (3).

[0102] (3)

[0103] in, Each element in the fraction matrix This represents the correlation between the query of the i-th sample and the key of the j-th sample.

[0104] After obtaining the fractional matrix, in order to prevent the softmax gradient from vanishing due to an excessively large dot product, a scaling factor can be applied to the fractional matrix to scale it (as shown in formula (4)). The scaling factor is the dimension of the key vector. The square root of.

[0105] (4)

[0106] in, The scaling result obtained by scaling the fractional matrix.

[0107] The scaling result is converted into a fractional distribution using the Softmax function, so that the sum of all weights is 1 and each weight is between 0 and 1, to obtain the weights of the fractional matrix. The specific normalization process is shown in formula (5).

[0108] (5)

[0109] Where A is the weight in the fraction matrix.

[0110] After obtaining the weights of the score matrix, i.e. the attention weight matrix A, the values ​​in the linear transformation result can be weighted and summed using the attention weight matrix A to obtain the data matrix, as shown in formula (6).

[0111] Z = AV (6)

[0112] in, It is a unified data matrix.

[0113] In some embodiments, after obtaining a unified data matrix, a bidirectional LSTM model can be used to extract the temporal correlation features of the data matrix and output multiple temporal feature vectors; an ID-CNN model can be used to extract the spatial correlation features of multi-device data within the data matrix and output multiple spatial feature vectors.

[0114] Furthermore, in some embodiments, XGBoost feature importance assessment can be used to evaluate the importance of each feature vector (temporal feature vector or spatial feature vector) so that redundant feature vectors can be removed from each feature vector according to the importance of each feature vector, and finally an optimized feature vector set is generated. The feature vector set includes multiple feature vectors, which are temporal feature vectors or spatial feature vectors.

[0115] Specifically, the process of using XGBoost feature importance evaluation to assess the importance of each feature vector is as follows:

[0116] (1) Initial training: Train a baseline XGBoost model using all feature vectors;

[0117] (2) Assess importance: Obtain the importance of each feature vector based on gain;

[0118] (3) Eliminate low-contribution features: Use the cumulative contribution method to calculate the contribution of each feature vector to the baseline XGBoost model based on the importance of each feature vector, and eliminate feature vectors with small contributions to the baseline XGBoost model.

[0119] (4) Eliminate highly correlated features: Calculate the correlation between each feature vector in the remaining feature vectors, and retain the most important (highest contribution) feature vector in each group of highly correlated feature vectors;

[0120] (5) Retraining and Validation: Using the filtered feature subset (the remaining feature vectors), retrain a baseline XGBoost model on a new training set, and re-evaluate the performance of the retrained baseline XGBoost model on the validation dataset. If the model performance does not decrease significantly, or even improves, the feature selection is successful. Otherwise, the cumulative contribution rate needs to be increased, and the feature vectors are removed again until the model performance does not decrease significantly, thus obtaining the final feature vector set.

[0121] Furthermore, after obtaining the feature vector set, the feature vector set can be associated with the IDs of multi-source industrial data, and the associated feature vector set of multi-source industrial data can be stored in the feature database of the private cloud cluster.

[0122] In this embodiment of the application, a bidirectional LSTM-1D-CNN fusion model can be pre-constructed to serve as a bidirectional fusion model to be trained; wherein, the bidirectional fusion model to be trained includes an input layer, a local ID-CNN feature extraction branch, a bidirectional LSTM temporal modeling branch, a branch global feature fusion branch, and an output layer.

[0123] In some embodiments, after obtaining the feature vector set, the feature vector set is divided into a training dataset and a validation dataset. The Adam optimizer is used to train and validate the model with mean squared error and cross-entropy loss as the joint loss function, resulting in a bidirectional fusion model.

[0124] It should be noted that the main process for constructing the bidirectional model to be trained is as follows:

[0125] (1) Constructing the input layer: Define the input layer structure of the bidirectional model to be trained. The input layer will directly read the optimized feature vector (i.e. the training dataset) from the feature database and ensure the consistency of the read data through the ID association mechanism.

[0126] (2) Constructing a 1D-CNN local feature extraction branch: Constructing a one-dimensional convolutional neural network branch specifically for capturing local temporal patterns and short-term dependencies in the training dataset. Through hierarchical feature extraction using three convolutional kernels, local correlations, abrupt changes, and periodic fluctuations in the feature vectors are identified.

[0127] (3) Constructing a bidirectional LSTM temporal modeling branch: Constructing a bidirectional long short-term memory network branch, focusing on learning and optimizing long-term dependencies and dynamic evolution patterns (global temporal features) in the feature sequence. The bidirectional structure ensures that the model can consider both historical context and future trends.

[0128] (4) Fusion of global features of CNN branch and global features of LSTM branch (branchovering branch): The local temporal features extracted by the CNN branch are fused with the global temporal features learned by the LSTM branch.

[0129] (5) Constructing the output layer: Generate prediction output according to the prediction problem type, with Adam as the optimizer and mean squared error + cross entropy as the joint loss function; wherein, the prediction problem type can be: time series prediction type or prediction category type, the time series prediction can be the prediction of the operating trend data of the device in the next hour, the prediction category can be the prediction of whether the device is operating normally, etc., which are not limited in this embodiment of the application.

[0130] (6) Create the model and compile the model to obtain the bidirectional fusion model to be trained;

[0131] (7) Conduct model training and validation. During validation, the prediction results need to be de-standardized and converted into business data.

[0132] Optionally, the bidirectional fusion model is trained using a training dataset and validated using a validation dataset. The process of obtaining the bidirectional fusion model can be as follows: the input layer receives the training dataset; the ID-CNN local feature extraction branch captures local temporal features and short-term dependencies from the training dataset; the bidirectional LSTM temporal modeling branch captures global temporal features and long-term dependencies from the training dataset; the global feature fusion branch fuses the local temporal features, short-term dependencies, global temporal features, and long-term dependencies to obtain fused features; the output layer is trained using the fused features and a pre-built joint loss function; and the validation dataset is used to validate the trained bidirectional fusion model to obtain the bidirectional fusion model.

[0133] In some embodiments, the process of validating the trained bidirectional fusion model using a validation dataset can be as follows: the trained bidirectional fusion model is used to predict the corresponding results based on the validation dataset, and the results are de-standardized to obtain business data. The business data is then compared with the actual business data. If they are consistent, the trained bidirectional fusion model is determined to have passed validation, and the bidirectional fusion model is obtained. Otherwise, the corresponding training continues until the validation is passed, and the bidirectional fusion model is obtained.

[0134] Furthermore, in this embodiment of the application, after obtaining the bidirectional fusion model, the bidirectional fusion model can be packaged into a Docker image using an automated script, which includes the trained bidirectional LSTM-1D-CNN fusion model file (bidirectional fusion model), dependency libraries, and inference scripts.

[0135] It should be noted that the dependency library refers to the basic libraries and tools that the bidirectional fusion model depends on when loading, that is, the runtime environment of the bidirectional fusion model; the inference script (Python code) defines the entire prediction process of the bidirectional fusion model. Specifically, it defines the complete logic of how to load the bidirectional fusion model, receive input data, perform prediction and return results after the service starts.

[0136] In the specific execution step S204, after extracting multiple target feature vectors, the bidirectional fusion model in the pre-packaged Docker image can be used to extract target temporal feature vectors and target spatial feature vectors from the multiple target feature vectors, and the target temporal feature vectors and target spatial feature vectors can be used for prediction to obtain the prediction result.

[0137] In this embodiment, the bidirectional fusion model also includes: an input layer, a local ID-CNN feature extraction branch, a bidirectional LSTM temporal modeling branch, a branch-global feature fusion branch, and an output layer. The process of extracting target temporal feature vectors and target spatial feature vectors from multiple target feature vectors using the bidirectional fusion model, and then using these vectors for prediction to obtain the prediction result, can be as follows: the input layer of the bidirectional fusion model receives multiple target feature vectors; the local feature extraction branch of the ID-CNN captures the target's local temporal features and short-term dependencies from the multiple target feature vectors.

[0138] The bidirectional LSTM temporal modeling branch captures global temporal features and long-term dependencies of the target from multiple target feature vectors; the global feature fusion branch fuses the target local temporal features, target short-term dependencies, target global temporal features, and target long-term dependencies to obtain the target fused features; the output layer uses the joint loss function and target fusion to output the prediction result for the prediction request.

[0139] It should be noted that if the prediction request is for equipment health status, the prediction result can be the equipment health status (e.g., equipment failure, remaining lifespan); if the preset request is for production energy consumption prediction, the prediction result can be production energy consumption (e.g., workshop energy consumption trend, equipment energy consumption optimization direction); if the prediction request is for product quality compliance rate prediction, the prediction result can be the product quality compliance rate (e.g., probability of meeting key product quality indicators, quality anomaly warning).

[0140] S205: Encrypt the prediction results and store the encrypted prediction results in the result database of the private cloud cluster using a preset encrypted industrial data transmission link.

[0141] In the specific execution step S205, after obtaining the prediction result, the industrial data transmission link is encrypted using the TLS1.3 protocol to obtain the preset encrypted industrial data transmission link. At the same time, the prediction result is encrypted using the SM4 national cryptographic algorithm. Finally, the encrypted prediction result is stored in the result database of the private cloud cluster using the preset encrypted industrial data transmission link.

[0142] It should be noted that after obtaining the prediction results, the prediction results can be encrypted using the SM4 national cryptographic algorithm, and the decrypted prediction results can be pushed to the user according to the preset configuration. The user can then provide feedback on the actual effect through the feedback entry set on the prediction results page. After detecting the actual effect, the actual effect can be included in the model's training dataset. The prediction request carries a preset configuration, which can be a specified MQTT / WSS / Webhook result push configuration. This configuration can be tailored to the specific application and is not limited in this embodiment.

[0143] It should also be noted that the corresponding prediction results can be extracted from the results database at regular intervals (such as every week), and the prediction results can be compared with the actual results to obtain comparison data. The comparison data can be used to evaluate the performance of the bidirectional fusion model. If the model performance is lower than the preset threshold, the iteration process can be automatically triggered to retrain and deploy the bidirectional fusion model using incremental training.

[0144] As can be seen, this application pushes prediction results through multiple protocols (such as Webhook, MQTT, and WSS protocols) and sets up feedback entry points to collect actual effect data, so as to form a closed-loop service of prediction-analysis-suggestion-feedback, thereby improving the practical application value of prediction results.

[0145] This application provides an intelligent prediction method for industrial data. The method involves receiving and parsing prediction requests to obtain prediction parameters; identifying target Pods matching the prediction parameters from a Kubernetes cluster and routing the prediction parameters to the target Pods; reading multiple target feature vectors corresponding to the target device indicated by the prediction request through the target Pods; extracting target temporal feature vectors and target spatial feature vectors from the target feature vectors using a bidirectional fusion model in a Docker image; and using these target temporal and spatial feature vectors for prediction to obtain the prediction result. The bidirectional fusion model is trained using multi-source industrial data. The prediction result is encrypted and stored in a result database of a private cloud cluster using a pre-set encrypted industrial data transmission link. Therefore, the technical means provided in this application, by calling heterogeneous computing cluster resources through a Kubernetes cluster, realizes on-demand allocation of CPU and GPU resources, thereby solving the problems of poor resource adaptability and resource utilization in the prior art. Furthermore, this application pre-trains a bidirectional fusion model using multi-source industrial data to be trained, so that after extracting the target feature vector, the bidirectional fusion model can be used to extract the target temporal feature vector and the target spatial feature vector from the target feature vector, and use the target temporal feature vector and the target spatial feature vector to make predictions, thereby obtaining high-precision prediction results. This solves the problem of low prediction accuracy caused by the inability of the prior art to handle multi-source heterogeneous data and spatiotemporal correlation. The prediction results are decrypted, and during the storage process, the encrypted prediction results are stored in the result database of the private cloud cluster using a preset encrypted industrial data transmission link, realizing full-process data security protection. This solves the problem of security risks of theft and leakage in the prior art, which seriously threatens the data security and trade secrets of enterprises.

[0146] based on Figure 2 The present application provides an industrial data intelligent prediction system, as illustrated by the industrial data intelligent prediction method. Figure 3 As shown, the industrial data intelligent prediction system includes:

[0147] The receiving unit 31 is used to receive the prediction request and parse the prediction request to obtain the prediction parameters;

[0148] Routing unit 32 is used to determine the target Pod that matches the prediction parameters from the Kubernetes cluster and route the prediction parameters to the target Pod;

[0149] Reading unit 33 is used to read multiple target feature vectors corresponding to the target device indicated by the prediction request through the target Pod;

[0150] The prediction unit 34 is used to extract the target temporal feature vector and the target spatial feature vector from multiple target feature vectors through the bidirectional fusion model in the Docker image, and to make predictions using the target temporal feature vector and the target spatial feature vector to obtain the prediction results; the bidirectional fusion model is obtained by the training unit using multi-source industrial data to train the bidirectional fusion model to be trained;

[0151] Storage unit 35 is used to encrypt the prediction results and store the encrypted prediction results in the result database of the private cloud cluster using a preset encrypted industrial data transmission link.

[0152] This application provides an intelligent prediction system for industrial data. It pre-trains a bidirectional fusion model using multi-source industrial data to obtain the bidirectional fusion model. After extracting the target feature vector, the bidirectional fusion model extracts the target temporal feature vector and the target spatial feature vector from the target feature vector. The system then uses these target temporal and spatial feature vectors for prediction, resulting in high-precision prediction results. This solves the problem of low prediction accuracy in existing technologies due to the inability to handle multi-source heterogeneous data and spatiotemporal correlation. The system decrypts the prediction results and stores them in a private cloud cluster's result database using a pre-set encrypted industrial data transmission link, achieving end-to-end data security protection. This addresses the security risks of theft and leakage in existing technologies, which seriously threaten enterprise data security and trade secrets.

[0153] Optional training units include:

[0154] The data acquisition unit is used to collect multi-source industrial data through a data acquisition device;

[0155] The preprocessing unit is used to preprocess multi-source industrial data and generate a feature vector set, wherein the feature vector set includes multiple feature vectors.

[0156] The partitioning unit is used to divide the feature vector set into training and validation datasets.

[0157] The training and validation unit is used to train the bidirectional fusion model to be trained using the training dataset and to validate the trained bidirectional fusion model using the validation dataset, thus obtaining the bidirectional fusion model.

[0158] Optionally, the industrial data intelligent prediction system provided in this application embodiment further includes:

[0159] The system also includes:

[0160] The packaging and deployment unit is used to package the bidirectional fusion model using automated scripts to obtain a Docker image and create a model Deployment in the Kubernetes cluster to achieve containerized deployment.

[0161] Optional, the preprocessing unit includes:

[0162] The data cleaning unit is used to clean multi-source industrial data.

[0163] The data fusion unit is used to standardize the cleaned multi-source industrial data to obtain a feature matrix; perform a linear transformation on the feature matrix and calculate attention based on the result of the linear transformation to obtain a fractional matrix of the feature matrix; calculate the weights of the fractional matrix and perform weighted processing based on the weights of the fractional matrix and the result of the linear transformation to obtain a data matrix.

[0164] The feature extraction unit is used to extract multiple temporal feature vectors and multiple spatial feature vectors from the data matrix;

[0165] The feature filtering unit is used to filter each temporal feature vector and multiple spatial feature vectors to obtain a feature vector set.

[0166] Optionally, the bidirectional fusion model to be trained includes an input layer, a local ID-CNN feature extraction branch, a bidirectional LSTM temporal modeling branch, a branch-global feature fusion branch, and an output layer; the training and validation unit includes:

[0167] The input layer is used to receive the training dataset;

[0168] The ID-CNN local feature extraction branch is used to capture local temporal features and short-term dependencies from the training dataset;

[0169] A bidirectional LSTM temporal modeling branch is used to capture global temporal features and long-term dependencies from the training dataset;

[0170] The global feature fusion branch is used to fuse local temporal features, short-term dependencies, global temporal features, and long-term dependencies to obtain fused features.

[0171] The output layer is used for training using fused features and a pre-built joint loss function; the trained bidirectional fusion model is validated using a validation dataset to obtain the bidirectional fusion model.

[0172] This application also provides a storage medium storing program instructions, which, when loaded and executed by a processor, implement any of the above-described embodiments of the intelligent prediction method for industrial data.

[0173] This application also provides an electronic device, such as Figure 4 As shown, the device includes a processor 401 and a memory 402, which are connected via a bus; the memory stores program instructions; the processor calls the program instructions in the memory to execute any of the above-described embodiments of the intelligent prediction method for industrial data.

[0174] The processor mentioned in this article can be the terminal's CPU, an integrated MCU within the terminal, or a combination of a CPU and an MCU. Furthermore, the processor contains a kernel that retrieves the corresponding program from memory; one or more kernels can be configured.

[0175] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0176] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0177] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0178] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0179] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An intelligent prediction method for industrial data, characterized in that, The method includes: Receive a prediction request and parse the prediction request to obtain prediction parameters; Determine the target Pod that matches the prediction parameters from the Kubernetes cluster, and route the prediction parameters to the target Pod; The target Pod reads multiple target feature vectors corresponding to the target device indicated by the prediction request; The target temporal feature vector and target spatial feature vector are extracted from multiple target feature vectors using a bidirectional fusion model in a Docker image, and prediction is performed using the target temporal feature vector and the target spatial feature vector to obtain the prediction result; the bidirectional fusion model is obtained by training a bidirectional fusion model to be trained using multi-source industrial data; The prediction results are encrypted, and the encrypted prediction results are stored in the result database of a private cloud cluster using a preset encrypted industrial data transmission link.

2. The method according to claim 1, characterized in that, The bidirectional fusion model is trained using multi-source industrial data to obtain the bidirectional fusion model, which includes: Collect multi-source industrial data using a data acquisition device; The multi-source industrial data is preprocessed to generate a feature vector set, wherein the feature vector set includes multiple feature vectors; The feature vector set is divided into a training dataset and a validation dataset; The bidirectional fusion model is trained using the training dataset and validated using the validation dataset to obtain the bidirectional fusion model.

3. The method according to claim 2, characterized in that, After obtaining the bidirectional fusion model, the method further includes: The bidirectional fusion model is encapsulated using automated scripts to obtain a Docker image, and a model Deployment is created in a Kubernetes cluster to achieve containerized deployment.

4. The method according to claim 2, characterized in that, The multi-source industrial data is preprocessed to generate a feature vector set, including: Data cleaning is performed on the multi-source industrial data; The cleaned multi-source industrial data is standardized to obtain a feature matrix. A linear transformation is performed on the feature matrix, and attention is calculated based on the result of the linear transformation to obtain the fraction matrix of the feature matrix; Calculate the weights of the fraction matrix, and perform weighted processing based on the weights of the fraction matrix and the result of the linear transformation to obtain the data matrix; Extract multiple temporal feature vectors and multiple spatial feature vectors from the data matrix; The time-series feature vectors and the spatial feature vectors are filtered to obtain a feature vector set.

5. The method according to claim 2, characterized in that, The bidirectional fusion model to be trained includes an input layer, a local ID-CNN feature extraction branch, a bidirectional LSTM temporal modeling branch, a branch global feature fusion branch, and an output layer; The bidirectional fusion model is trained using the training dataset and validated using the validation dataset to obtain the bidirectional fusion model, which includes: The training dataset is received through the input layer; The ID-CNN local feature extraction branch captures local temporal features and short-term dependencies from the training dataset. The bidirectional LSTM temporal modeling branch captures global temporal features and long-term dependencies from the training dataset; The local temporal features, short-term dependencies, global temporal features, and long-term dependencies are fused through the global feature fusion branch to obtain fused features; The output layer is trained using the fused features and a pre-constructed joint loss function. The trained bidirectional fusion model is validated using the validation dataset to obtain the bidirectional fusion model.

6. An intelligent prediction system for industrial data, characterized in that, The system includes: A receiving unit is used to receive a prediction request and parse the prediction request to obtain prediction parameters; A routing unit is used to determine the target Pod that matches the prediction parameters from the Kubernetes cluster, and to route the prediction parameters to the target Pod; The reading unit is used to read multiple target feature vectors corresponding to the target device indicated by the prediction request through the target Pod; The prediction unit is used to extract target temporal feature vectors and target spatial feature vectors from multiple target feature vectors through a bidirectional fusion model in the Docker image, and to make predictions using the target temporal feature vectors and target spatial feature vectors to obtain prediction results; the bidirectional fusion model is obtained by the training unit training a bidirectional fusion model to be trained using multi-source industrial data; The storage unit is used to encrypt the prediction results and store the encrypted prediction results in the result database of the private cloud cluster using a preset encrypted industrial data transmission link.

7. The system according to claim 6, characterized in that, The training unit includes: The data acquisition unit is used to collect multi-source industrial data through a data acquisition device; A preprocessing unit is used to preprocess the multi-source industrial data to generate a feature vector set, wherein the feature vector set includes multiple feature vectors; The partitioning unit is used to divide the feature vector set into training and validation datasets. The training and validation unit is used to train the bidirectional fusion model to be trained using the training dataset, and to validate the trained bidirectional fusion model using the validation dataset, thereby obtaining the bidirectional fusion model.

8. The system according to claim 7, characterized in that, The system also includes: The packaging and deployment unit is used to package the bidirectional fusion model using automated scripts to obtain a Docker image and create a model Deployment in the Kubernetes cluster to achieve containerized deployment.

9. An electronic device, characterized in that, include: A processor and a memory are connected via a bus; wherein the processor is used to call and execute a program stored in the memory; The memory is used to store a program for implementing the industrial data intelligent prediction method as described in any one of claims 1-5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for performing the industrial data intelligent prediction method as described in any one of claims 1-5.

Images