Digital mechanical equipment data management system based on cloud computing

Abstract

The invention provides a digital mechanical equipment data management system based on cloud computing. The digital mechanical equipment data management system comprises a data acquisition module for acquiring equipment state and environment information; the feature coding module is composed of a gating circulation unit and a multi-layer perceptron and is used for extracting high-dimensional features; the feature fusion module fuses the high-dimensional features and outputs sparse joint codes; the cloud computing model construction module constructs a model and introduces a trainable enhancement layer; the state interaction module enables the code to obtain an advanced interaction state based on a model; the state synchronization module synchronizes a device state and environment information through MechLayer; the implicit state generation module obtains an implicit state by combining normalization and enhanced connection; operating an environment perception modeling module to improve perception estimation modeling; the perception learning vector generation module obtains a learning vector; the state prediction module predicts an operation state; and the track decoding module processes the input to obtain an equipment running track. The equipment management efficiency and the operation reliability can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data management, and more specifically, the present invention relates to a digital mechanical equipment data management system based on cloud computing. Background Art

[0002] In traditional mechanical equipment management, data collection and analysis usually rely on local servers and manual operations, which are inefficient and error-prone. With the advancement of industrial automation, the amount of data generated by mechanical equipment has increased sharply, posing higher requirements for data processing and analysis capabilities. Existing data management systems often face problems such as data islands, insufficient information fusion, and low prediction accuracy when dealing with large-scale and high-dimensional mechanical equipment data. For example, although some systems can collect equipment status data, they perform poorly in in-depth data mining and real-time prediction, and cannot provide accurate basis for preventive maintenance of equipment.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the prior art: First, the cooperation between the data collection module and the analysis module is insufficient, resulting in poor information transmission; Second, the feature extraction method is not advanced enough to effectively mine high-dimensional features in mechanical equipment status and operating environment information; Third, there is a lack of an effective feature fusion mechanism, and the in-depth interaction between mechanical equipment information and operating environment information cannot be achieved; Finally, the accuracy and real-time performance of the state prediction module need to be improved, and the requirements for accurate prediction of equipment operating status in industrial sites cannot be met. Summary of the Invention

[0004] The present invention provides a digital mechanical equipment data management system based on cloud computing, including: A data collection module for obtaining mechanical equipment status and operating environment information; A feature encoding module, including an embedding network composed of a gated recurrent unit and a multi-layer perceptron, for encoding the mechanical equipment status and operating environment information to extract high-dimensional features; A feature fusion module for fusing high-dimensional features to perform information exchange between mechanical equipment information and operating environment information, and outputting a sparse joint code; A cloud computing model construction module for constructing a cloud computing model, introducing a trainable enhancement layer, and updating the attention layer using a dynamic weight adjustment mechanism to form a cloud computing model; A state interaction module for enabling the sparse joint code to obtain a high-level interaction state based on the pre-trained cloud computing model; The state synchronization module synchronizes the operating state and operating environment information of the target mechanical equipment through MechLayer at each time step. The MechLayer includes Mech blocks, a normalization layer, a pointwise feedforward network, and enhanced connections. The MechLayer normalizes the first input formed by the high-level interaction state and partial sparse joint coding to obtain a first output, maps the first output through Mech blocks to obtain a second output. The implicit state generation module combines normalization and enhanced connections based on the first output and the second output to obtain an implicit state. The operating environment perception modeling module improves the operating environment perception estimation modeling in the hidden layer dimension through the pointwise feedforward network to obtain a third output. The perception learning vector generation module obtains the operating environment perception learning vector based on the implicit state and the third output through normalization and enhanced connections. The state prediction module feeds the operating environment perception learning vector into a multi-layer perceptron to obtain the prediction scores of multiple operating state segments within the corresponding time step. The prediction score of the operating state segment closest to the true operating state position is 1, and the prediction scores of the remaining operating state segments are 0. The trajectory decoding module processes the second input through a multi-modal Gaussian decoder. The second input consists of sparse joint coding, high-level interaction features, parameters formed by the candidate operating state segment with the highest evaluation score, and latent vectors sampled from a multivariate normal distribution, to obtain the operating trajectory of the mechanical equipment in continuous time steps.

[0005] Furthermore, the feature fusion module fuses high-dimensional features for information exchange between mechanical equipment information and operating environment information. The output sparse joint coding includes: performing feature fusion between mechanical equipment through a multi-head self-attention mechanism and an activation function. Updating the mechanical equipment and the operating environment through a multi-head attention mechanism and enhanced connections to achieve feature fusion between mechanical equipment and feature fusion between mechanical equipment and the operating environment. It is expressed as:

[0006] Where and are the high-dimensional features proposed after the mechanical equipment state and operating environment information pass through the embedding network. is the multi-head self-attention mechanism. is the activation function. Connect to form the sparse joint coding .

[0007] Furthermore, the cloud computing model construction module updates the attention layer using a dynamic weight adjustment mechanism, including: Let the parameter of the dynamic weight adjustment be , the An attention layer network has an input of whose dimension is and an output of whose dimension is and adjusts the network by adjusting the weight matrix in the following way:

[0008] where and the element values of are dynamically adjusted according to the performance feedback during the training process; During the training process, remains trainable and is updated according to the gradient of the loss function.

[0009] Further, the forward transfer function of the dynamic weight adjustment is expressed as: input the sparse joint coding into the cloud computing model and the output obtained is:

[0010] It is transformed through a multi-layer perceptron to match the dimension of to generate the final high-level interaction state .

[0011] Further, in the state synchronization module, MechLayer normalizes the first input to obtain a first output; the first output is mapped through the Mech block to obtain a second output, including:

[0012] where is the second output, is the normalization, represents the output after being processed by the Mech block, is the first input.

[0013] Further, the processing of the Mech block includes expanding the input dimension by using a linear projection, with the expansion coefficient being to generate different representations of two parallel branches, denoted as and respectively; capturing the running environment perception dependence through one-dimensional convolution and activation functions and linearly projecting into , and , for transformation and the structured state matrix to generate and ; the selective state model receives , and as inputs, and generates the corrected runtime environment perception features ; An activation function is introduced to generate a gating signal that can filter out irrelevant information ; is multiplied by and output after linear projection .

[0014] Furthermore, in the implicit state generation module, based on the first output and the second output, the implicit state is obtained by combining normalization and enhanced connection, including:

[0015] wherein, is the implicit state, is the weight parameter of the enhanced connection; The runtime environment perception modeling module improves the runtime environment perception estimation modeling in the hidden layer dimension through a pointwise feedforward network, and the third output includes:

[0016] wherein, is the third output, , , and are trainable parameters, is the activation function; The perception learning vector generation module obtains the runtime environment perception learning vector by combining normalization and enhanced connection based on the implicit state and the third output, including:

[0017] wherein, is the runtime environment perception learning vector, is the weight parameter of the enhanced connection.

[0018] Furthermore, in the state prediction module, the runtime environment perception learning vector enters a multi-layer perceptron to obtain the prediction scores of multiple runtime state segments within the corresponding time step, including:

[0019] wherein, is at time the The predicted score of a running status segment; For the running environment perception learning vector of the running status segment, For the running environment perception learning vector of the running status segment, Is the exponential function with the natural constant as the base.

[0020] Furthermore, the running environment perception estimation model is optimized for probability through the binary cross-entropy loss function, expressed as:

[0021] Among them, is the binary cross-entropy loss function, is the predicted score, is the actual score label, for each time step .

[0022] Furthermore, the second input is processed by a multimodal Gaussian decoder , the second input consists of sparse joint coding, high-level interaction features, parameters formed by the candidate running status segment with the highest evaluation score, and latent vectors sampled from a multivariate normal distribution, to obtain the operating trajectory of the mechanical equipment in continuous time steps is the trajectory, specifically expressed as:

[0023] Among them, is the mean, is the covariance matrix, is the latent vector sampled from the multivariate normal distribution ; The second input consists of the following parts:

[0024] Among them, is the sparse joint coding, is the high-level interaction feature guided by the running environment, is the latent vector sampled from the multivariate normal distribution; and perform cross-attention generation , formed by connecting the candidate running status segments with the highest evaluation scores.

[0025] The above embodiments of the present invention have at least the following beneficial effects: The digital mechanical equipment data management system of the present invention can efficiently collect and process the status and operating environment information of mechanical equipment. Through the collaborative work of the feature encoding module and the feature fusion module, high-dimensional features can be extracted and fused to form a sparse joint code, thereby realizing the deep interaction between mechanical equipment information and operating environment information, and providing a richer information basis for subsequent state prediction and trajectory decoding. The trainable enhancement layer and dynamic weight adjustment mechanism introduced by the cloud computing model construction module can further optimize the performance of the model and improve the sensitivity and adaptability to changes in the status of mechanical equipment.

[0026] In addition, the design of the state synchronization module and the implicit state generation module can ensure the accurate synchronization and update of the operating state and operating environment information of mechanical equipment at each time step. Combining the operating environment perception modeling module and the perception learning vector generation module can more accurately estimate and model the operating environment perception, and further improve the prediction accuracy of the state prediction module. The trajectory decoding module uses a multi-modal Gaussian decoder to generate the operating trajectory of mechanical equipment in consecutive time steps based on the predicted learning vector, providing an intuitive reference for equipment operation monitoring and fault diagnosis, and effectively improving the intelligent level of equipment management and the reliability of operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the drawings, several embodiments of the present invention are shown by way of illustration and not limitation, wherein: Figure 1 FIG. is a schematic structural diagram of a cloud computing-based digital mechanical equipment data management system provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] The principles and spirit of the present invention will now be described with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present invention, and not to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to be able to fully convey the scope of the present invention to those skilled in the art.

[0029] Those skilled in the art know that the embodiments of the present invention can be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention can be specifically implemented in the following forms: completely hardware, completely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0030] It should be noted that the number of any elements in the drawings is for illustration rather than limitation, and any naming is only for distinction and does not have any limiting meaning.

[0031] Reference is made below Figure 1 , Figure 1 which is a schematic structural diagram of a cloud computing-based digital mechanical equipment data management system provided by an embodiment of the present invention. As Figure 1 shown, a cloud computing-based digital mechanical equipment data management system 100 includes: A data acquisition module 101 for obtaining mechanical equipment status and operating environment information; A feature encoding module 102 includes an embedding network, which is composed of a gated recurrent unit and a multi-layer perceptron, for encoding the mechanical equipment status and operating environment information to extract high-dimensional features; a feature fusion module for fusing high-dimensional features to perform information exchange between mechanical equipment information and operating environment information, and outputting a sparse joint code; A cloud computing model construction module 103 for constructing a cloud computing model, introducing a trainable enhancement layer, and updating the attention layer using a dynamic weight adjustment mechanism to form a cloud computing model; a state interaction module for enabling the sparse joint code to obtain a high-level interaction state based on the pre-trained cloud computing model; A state synchronization module 104 synchronizes the operating status and operating environment information of the target mechanical equipment at each time step through MechLayer, and the MechLayer includes a Mech block, a normalization layer, a pointwise feed-forward network, and enhanced connections; the MechLayer normalizes the first input formed by the high-level interaction state and part of the sparse joint code to obtain a first output; maps the first output through the Mech block to obtain a second output; an implicit state generation module, based on the first output and the second output, combines normalization and enhanced connections to obtain an implicit state; an operating environment perception modeling module improves the operating environment perception estimation modeling in the hidden layer dimension through a pointwise feed-forward network to obtain a third output; a perception learning vector generation module, based on the implicit state and the third output, obtains an operating environment perception learning vector through normalization and enhanced connections; a state prediction module enables the operating environment perception learning vector to enter a multi-layer perceptron to obtain prediction scores for multiple operating status segments within the corresponding time step, and the prediction score of the operating status segment closest to the true operating status position is 1, and the prediction scores of the remaining operating status segments are 0; A trajectory decoding module 105 processes a second input through a multi-modal Gaussian decoder, and the second input is composed of a sparse joint code, high-level interaction features, parameters formed by the candidate operating status segment with the highest evaluation score, and a latent vector sampled from a multivariate normal distribution, to obtain the operating trajectory of the mechanical equipment in continuous time steps.

[0032] It should be noted that the data acquisition module is used to obtain the status and operating environment information of mechanical equipment. The status of mechanical equipment here refers to various parameters during the operation of the equipment, such as temperature, pressure, vibration frequency, etc. These parameters can reflect the health status and operating efficiency of the equipment. The operating environment information includes the external conditions where the equipment is located, such as temperature, humidity, dust concentration, etc. These conditions have an important impact on the performance and lifespan of the equipment. The data acquisition module realizes the real-time acquisition of data through a sensor network. The sensor network consists of multiple different types of sensors and can cover key parts of the equipment and key areas of the operating environment.

[0033] Specifically, the sensors in the data acquisition module can include temperature sensors, pressure sensors, vibration sensors, etc. The sampling frequency of these sensors can be set according to the operating characteristics and maintenance requirements of the equipment. For example, for high-speed rotating equipment, the sampling frequency of the vibration sensor can be set to 1000 times per second to ensure that small vibration changes of the equipment can be captured. The acquisition of operating environment information can be achieved through environmental monitoring sensors, such as temperature and humidity sensors, dust sensors, etc. The sampling frequency of these sensors can be set to once per minute to meet the basic requirements of environmental monitoring.

[0034] Preferably, the data acquisition module can adopt wireless sensor network (WSN) technology. The sensor nodes transmit the acquired data to the central data processing unit through wireless communication. The wireless sensor network has the advantages of flexible deployment and strong scalability and can adapt to complex industrial environments. In addition, in order to improve the accuracy and reliability of data, data fusion technology can be adopted to fuse and process the same type of data collected by multiple sensors to reduce data noise and improve data quality.

[0035] It should be noted that the feature encoding module includes an embedding network, which consists of a gated recurrent unit (GRU) and a multi-layer perceptron (MLP). The gated recurrent unit is a variant of the recurrent neural network (RNN) and can effectively handle the long-term dependence problem in sequence data. The multi-layer perceptron is a feedforward neural network and can perform non-linear transformation on high-dimensional data. The role of the feature encoding module is to encode the status and operating environment information of mechanical equipment and extract high-dimensional features, which can better represent the operating status and environmental conditions of the equipment.

[0036] Specifically, the gated recurrent unit (GRU) can be set with multiple hidden layers, and the number of neurons in each hidden layer can be adjusted according to the complexity of the data. For example, for complex industrial equipment, 3 hidden layers can be set, with 128 neurons in each hidden layer. The multi-layer perceptron (MLP) can contain multiple fully connected layers, and the number of neurons in each fully connected layer can also be set as needed. For example, 2 fully connected layers can be set, with 256 neurons in the first fully connected layer and 128 neurons in the second fully connected layer. The activation function can use ReLU (Rectified Linear Unit), and the ReLU function can effectively avoid the problem of gradient disappearance and improve the training efficiency of the network.

[0037] Preferably, the feature encoding module can be initialized with a pre-trained model. The pre-trained model can be trained on a large amount of device operation data to learn general feature representations. In practical applications, the parameters of the pre-trained model can be used as initial parameters, and then fine-tuned on a specific dataset to improve the adaptability and accuracy of the model.

[0038] Furthermore, in order to further improve the robustness of feature extraction, regularization techniques such as Dropout can be introduced to prevent the model from overfitting.

[0039] It should be noted that the feature fusion module is used to fuse high-dimensional features for information exchange between mechanical equipment information and operating environment information, and outputs a sparse joint encoding. Here, the high-dimensional features refer to the features extracted by the feature encoding module, and these features have high dimensions and rich information. The sparse joint encoding refers to the encoding obtained after fusing the features of mechanical equipment information and operating environment information, and this encoding has sparsity, which can effectively reduce data redundancy and improve computational efficiency.

[0040] Specifically, the feature fusion module can use the multi-head self-attention mechanism (Multi-Head Self-Attention) and an activation function for feature fusion. The multi-head self-attention mechanism can simultaneously focus on features in different positions, improving the comprehensiveness and accuracy of feature fusion. The activation function can use the Softmax function, and the Softmax function can normalize the output values into a probability distribution for subsequent processing. For example, assuming that the mechanical equipment state and operating environment information respectively extract 128-dimensional high-dimensional features, after fusion by the multi-head self-attention mechanism and the Softmax function, a 256-dimensional sparse joint encoding can be obtained.

[0041] Preferably, the feature fusion module can introduce residual connection and layer normalization techniques. Residual connection can alleviate the problem of vanishing gradients in the training of deep networks and improve the training effect of the network. Layer normalization can normalize the output of each layer, improving the stability and convergence speed of the model.

[0042] Furthermore, to further improve the effect of feature fusion, a multi-scale feature fusion strategy can be adopted to fuse features of different scales to capture multi-scale information of the device state and operating environment.

[0043] In some embodiments, the feature fusion module fuses high-dimensional features for information exchange between mechanical equipment information and operating environment information, and the output sparse joint coding includes: performing feature fusion between mechanical equipment through a multi-head self-attention mechanism and an activation function; updating the mechanical equipment and the operating environment through a multi-head attention mechanism and an enhanced connection to achieve feature fusion between mechanical equipment and feature fusion between mechanical equipment and the operating environment; It is expressed as:

[0044] where, and are the high-dimensional features extracted after the mechanical equipment state and operating environment information pass through the embedding network, is the multi-head self-attention mechanism, is the activation function; Connect to form a sparse joint coding .

[0045] It should be noted that the feature fusion module fuses high-dimensional features for information exchange between mechanical equipment information and operating environment information, and outputs sparse joint coding. Here, the high-dimensional features refer to the features extracted by the feature encoding module. These features have high dimensions and rich information content, and can better represent the operating state of the device and environmental conditions. Information exchange means that through the feature fusion process, the mechanical equipment information and the operating environment information influence each other, enhancing the expression ability of the features. Sparse joint coding means that the fused features are sparsified to reduce data redundancy and improve computational efficiency.

[0046] Specifically, the feature fusion module performs feature fusion between mechanical devices through the multi-head self-attention mechanism and activation function. The multi-head self-attention mechanism can simultaneously focus on features at different positions, improving the comprehensiveness and accuracy of feature fusion. The activation function can use the ReLU or Softmax function. The ReLU function can effectively avoid the problem of gradient disappearance and improve the training efficiency of the network, while the Softmax function can normalize the output values into a probability distribution for subsequent processing. For example, assuming that the mechanical device state and operating environment information extract 128-dimensional high-dimensional features respectively, after fusion through the multi-head self-attention mechanism and activation function, a 256-dimensional sparse joint encoding can be obtained.

[0047] Preferably, the feature fusion module can introduce residual connections and layer normalization techniques. Residual connections can alleviate the problem of gradient disappearance in the training of deep networks and improve the training effect of the network. Layer normalization can normalize the output of each layer, improving the stability and convergence speed of the model.

[0048] Furthermore, to further improve the effect of feature fusion, a multi-scale feature fusion strategy can be adopted to fuse features of different scales to capture multi-scale information of the device state and operating environment. For example, local features and global features can be weighted and fused, and the weights can be adjusted according to specific application scenarios to optimize the effect of feature fusion.

[0049] It should be noted that the feature fusion module updates the mechanical devices and the operating environment through the multi-head attention mechanism and enhanced connections, realizing feature fusion between mechanical devices and feature fusion between mechanical devices and the operating environment. Here, the multi-head attention mechanism is an extension of the attention mechanism. It divides the input into multiple heads, calculates the attention weights respectively, then concatenates the results, and performs a linear transformation again, thereby improving the model's ability to focus on different subspaces. Enhanced connection means that during the feature fusion process, additional connections are added to enhance the expression ability of features, such as residual connections.

[0050] Specifically, the multi-head attention mechanism can set multiple heads, and the dimension of each head can be adjusted according to the dimension of the features. For example, assuming the feature dimension is 512, 8 heads can be set, and the dimension of each head is 64. The enhanced connection can adopt a residual connection, that is, directly adding the input to the output to avoid the problem of gradient disappearance. For example, assuming the input feature is X and the output after the multi-head attention mechanism is Y, the final output can be expressed as Y + X.

[0051] Preferably, the feature fusion module can experiment with different activation functions to find the activation function that best suits the current data. For example, the LeakyReLU or Tanh activation functions can be tried, which may perform better than the ReLU or Softmax functions in some cases.

[0052] Furthermore, to further improve the effect of feature fusion, a dynamic adjustment mechanism for attention weights can be introduced to dynamically adjust the attention weights according to the importance of the input features, so as to enhance the attention to important features.

[0053] It should be noted that the output of the feature fusion module is expressed as:

[0054] where and are the high-dimensional features extracted from the mechanical equipment status and operating environment information through the embedding network, represents the connection operation, which connects multiple high-dimensional features to form a sparse joint coding . The connection operation here refers to splicing multiple feature vectors into a longer feature vector to retain the information of all features.

[0055] Specifically, assuming that the dimension of each high-dimensional feature is 128 and there are 10 such features, after the connection operation, a 1280-dimensional sparse joint coding can be obtained. The connection operation can adopt a simple splicing method or a weighted splicing method, and the weights can be set according to the importance of the features. For example, if some features are considered more important, higher weights can be assigned.

[0056] Preferably, the feature fusion module can use dimensionality reduction techniques to process the sparse joint coding to reduce the computational complexity. For example, dimensionality reduction methods such as principal component analysis (PCA) or linear discriminant analysis (LDA) can be used to reduce the 1280-dimensional sparse joint coding to 256 dimensions or fewer dimensions while retaining most of the important information.

[0057] Furthermore, to further improve the effect of feature fusion, regularization techniques such as L1 or L2 regularization can be introduced to prevent the model from overfitting and improve the generalization ability of the model.

[0058] In some embodiments, the cloud computing model construction module uses the dynamic weight adjustment mechanism to update the attention layer, including: Let the parameter of the dynamic weight adjustment be , the th attention layer network has an input of , and its dimension is , the output is Its dimension is , adjust the network in the following way the weight matrix in :

[0059] wherein, , and the element values of are dynamically adjusted according to the performance feedback during the training process; During the training process, remains trainable, and is updated according to the gradient of the loss function.

[0060] It should be noted that the cloud computing model construction module uses a dynamic weight adjustment mechanism to update the attention layer. This mechanism aims to optimize the model's attention to different features by dynamically adjusting the weights, thereby improving the model's adaptability and prediction accuracy. The dynamic weight adjustment mechanism here refers to dynamically adjusting the values of the weight matrix in the attention layer according to the model's performance feedback during the training process to enhance the model's learning ability for important features.

[0061] Specifically, the parameters for dynamic weight adjustment are , the th attention layer network has an input of , its dimension is , and the output is , with a dimension of . Adjust the weight matrix in the network in the following way : , wherein, , and the element values of are dynamically adjusted according to the performance feedback during the training process. During the training process, remains trainable, and is updated according to the gradient of the loss function. Preferably, different performance feedback metrics can be used to adjust the weight matrix , for example, metrics such as accuracy, recall, or F1-score can be used.

[0062] Furthermore, in order to further improve the robustness of the model, a weight decay technique can be introduced. By adding an L2 norm term of the weights to the loss function, the size of the weights can be restricted to prevent overfitting.

[0063] More specifically, the parameter settings and concepts in the dynamic weight adjustment mechanism include: the initial value of the weight matrix can be randomly initialized or initialized with the weights of a pre-trained model. The input and the output dimension and can be set according to specific application scenarios and data characteristics. For example, if the input feature dimension is high, the output dimension can be appropriately increased to retain more useful information. The performance feedback metric can be the loss value of the model on the validation set or the convergence speed of the model during training. Specific parameters such as the learning rate and batch size can be adjusted according to the experimental results. For example, the learning rate can be set to 0.001 and the batch size can be set to 32 or 64.

[0064] Preferably, an adaptive learning rate algorithm such as Adam or RMSprop can be adopted. These algorithms can automatically adjust the learning rate according to the magnitude of the gradient, improving the training efficiency.

[0065] Furthermore, to further improve the generalization ability of the model, cross-validation techniques can be adopted. The dataset is divided into multiple subsets for multiple training and validations to evaluate the performance of the model.

[0066] Preferably, the dynamic weight adjustment mechanism can be further refined as follows: at the end of each training epoch, calculate the update value of the weight adjustment matrix according to the performance of the model on the validation set. For example, the update value of can be calculated using the gradient descent method, and the update formula is:

[0067] where is the learning rate, is the loss function with respect to gradient.

[0068] Furthermore, a momentum term can be introduced to accelerate the weight update process, and the momentum term can be set to 0.9. Alternative solutions can be to adopt different optimization algorithms such as Stochastic Gradient Descent (SGD) or Adagrad, which may perform better in different application scenarios.

[0069] Even further, to further improve the robustness of the model, data augmentation techniques can be adopted. By randomly transforming the input data, such as rotation, scaling, or noise addition, the diversity of the data is increased, and the model's adaptability to different inputs is improved.

[0070] In some embodiments, the forward propagation function of the dynamic weight adjustment is expressed as: input the sparse joint coding into the cloud computing model and the output obtained is:

[0071] Transform through a multi - layer perceptron to match the dimensions, thus generating the final high - level interaction state .

[0072] It should be noted that the forward propagation function of dynamic weight adjustment describes how to input the sparse joint coding into the cloud computing model and transform it through a multi - layer perceptron to generate the final high - level interaction state. Here, the forward propagation function refers to the process in a neural network where the input data passes through the network layers for calculation and finally obtains the output.

[0073] Specifically, the sparse joint coding is input into the cloud computing model and the output obtained is . Then, it is transformed through a multi - layer perceptron (MLP) to match the dimensions, thus generating the final high - level interaction state . This process ensures that the model can effectively process and fuse feature information from different sources to improve the prediction ability of the mechanical equipment state.

[0074] More specifically, the cloud computing model can be a deep neural network with multiple hidden layers. The number of neurons in each hidden layer can be set according to the complexity of the input features. For example, if the dimension of the sparse joint coding is 256, an MLP with two hidden layers can be set. The first hidden layer has 512 neurons and the second hidden layer has 256 neurons. The activation function can use ReLU because it performs well in dealing with non - linear problems and can effectively avoid the vanishing gradient problem.

[0075] Furthermore, to ensure that the output can match the input in dimension, a linear activation function can be used in the last layer of the MLP, or the dimension matching can be achieved by adjusting the number of neurons. Specific parameters such as the learning rate can be set to 0.001 and the batch size can be set to 64, and these parameters can be adjusted according to the actual training effect.

[0076] Preferably, to further optimize the performance of the forward propagation function, the batch normalization technology can be introduced into the MLP to reduce the internal covariate shift, accelerate the training process and improve the stability of the model. In addition, the dropout technology can be used to prevent overfitting, and the dropout rate can be set to 0.5.

[0077] Furthermore, in an alternative solution, in addition to the MLP, a convolutional neural network (CNN) can also be considered to process the sparse joint coding. Especially when the input features have spatial correlation, the CNN can better capture local features. Additionally, to improve the generalization ability of the model, a cross-validation method can be adopted to evaluate the performance of the model on different data subsets, so as to select the optimal model structure and parameters.

[0078] In some embodiments, in the state synchronization module, MechLayer normalizes the first input to obtain a first output; mapping and obtaining a second output through the Mech block includes:

[0079] wherein, is the second output, is normalization, represents the output after being processed by the Mech block, is the first input.

[0080] It should be noted that MechLayer in the state synchronization module normalizes the first input to obtain a first output; mapping and obtaining a second output through the Mech block. Here, MechLayer is a specially designed network layer for synchronizing the operating states and operating environment information of mechanical equipment. The normalization process refers to adjusting the data to a common scale so that effective comparison and fusion can be carried out between different features. The Mech block is a processing unit for further mapping and transforming the normalized data to generate more useful outputs.

[0081] Specifically, the processing process of MechLayer can be expressed as:

[0082] wherein, is the first output, is the normalization operation, is the first input. Then, the second output is obtained through the processing of the Mech block:

[0083] wherein, represents the processing function of the Mech block. The normalization operation can adopt batch normalization (BatchNormalization) or layer normalization (Layer Normalization), and the specific selection depends on the characteristics of the data and the requirements of the model. The Mech block can include multiple neural network layers, such as fully connected layers or convolutional layers, for extracting and transforming features.

[0084] Specifically, the normalization operation in MechLayer can be set to layer normalization because it is insensitive to the distribution changes of input data and is more suitable for processing sequence data. The parameters of layer normalization include a scaling factor and an offset factor, and these parameters can be learned through the backpropagation algorithm during the training process. The Mech block can be designed to include two fully connected layers. The output dimension of the first fully connected layer can be set to twice the input dimension to increase the expressive power of features; the output dimension of the second fully connected layer is the same as the input dimension to maintain data consistency. The activation function can use ReLU to introduce non-linearity. For example, if the dimension of the first input is 128, the output dimension of the first fully connected layer of the Mech block can be set to 256, and the output dimension of the second fully connected layer is then set to 128.

[0085] Furthermore, to improve the robustness of the model, dropout technology can be introduced into the Mech block, and the dropout rate can be set to 0.2.

[0086] Preferably, the normalization operation in MechLayer can be further optimized. For example, adaptive normalization technology can be adopted, which can dynamically adjust the normalization parameters according to different inputs. In addition, the Mech block can adopt residual connections, that is, directly add the input to the output to avoid the problem of gradient disappearance and enhance the learning ability of the model. For example, the output of the Mech block can be expressed as .

[0087] Even further, in an alternative solution, the Mech block can adopt a more complex network structure, such as the self-attention mechanism in the Transformer architecture, to better capture long-range dependencies in the input data. In addition, to improve the generalization ability of the model, data augmentation techniques can be adopted during the training process, such as adding random noise or perturbing features to the input data.

[0088] In some embodiments, the processing of the Mech block includes expanding the input dimension using linear projection, with an expansion coefficient of , generating different representations of two parallel branches, denoted as and ; capturing the running environment perception dependencies through one-dimensional convolution and activation functions , and linearly projecting into , and respectively, for transforming and the structured state matrix to generate and ; The selective state model accepts , and are used as inputs to generate corrected runtime environment perception features ; An activation function is introduced to generate a gating signal that can filter out irrelevant information ; Multiply by and output after linear projection .

[0089] It should be noted that the processing process of the Mech block involves using linear projection to expand the input dimension and generating different representations of two parallel branches. This process aims to enhance the model's understanding of the input data through different feature transformation paths. Among them, linear projection is a mathematical transformation that maps data from one dimensional space to another through matrix multiplication to meet the different processing requirements of the model. Parallel branches mean that data will pass through multiple processing paths simultaneously, and each path focuses on extracting different aspects of features, thereby enriching the model's feature representation ability.

[0090] Specifically, the Mech block first expands the dimension of the input data through linear projection, and the expansion coefficient is , generating different representations of two parallel branches, denoted as and . Then, one-dimensional convolution and activation function are used to capture the runtime environment perception dependencies , and are linearly projected into , and , where is used to transform and the structured state matrix to generate and . The selective state model accepts , and as inputs to generate corrected runtime environment perception features . At the same time, an activation function is introduced to generate a gating signal that can filter out irrelevant information , Multiply by and output after linear projection .

[0091] Specifically, the expansion coefficient of the linear projection can be set according to the design requirements of the model and the characteristics of the input data. For example, if the original input dimension is 100, the expansion coefficient It can be set to 2, so that the extended dimension is 200, providing a richer feature space. The kernel size and stride parameters of the one-dimensional convolution should be adjusted according to the time series characteristics of the data to effectively capture local time dependencies. The activation function can be ReLU or LeakyReLU to introduce non-linearity and prevent gradient vanishing. For example, for processing, three one-dimensional convolutional layers can be used, with a kernel size of 3 and a stride of 1 for each convolutional layer, and the number of output channels being 64, 128, and 256 respectively, to gradually extract deeper features. For the selective state model, a long short-term memory network (LSTM) or a gated recurrent unit (GRU) can be used to process the input sequence data to model long-term dependencies in the time series.

[0092] Preferably, to further improve the performance of the Mech block, different regularization strategies can be adopted for the parallel branches after linear projection. For example, Dropout can be applied to the branch, and Batch Normalization can be applied to the branch to enhance the generalization ability and stability of the model. In addition, for the choice of activation function, the Swish activation function can be tried. It combines the advantages of the ReLU and Sigmoid functions, can provide a smoother gradient, and helps optimize the training process.

[0093] Furthermore, in an alternative, the one-dimensional convolution can be replaced by a causal convolution to ensure that no leakage of future information is introduced when processing time series data, which is particularly important for prediction tasks that need to strictly follow the time order. At the same time, the selective state model can adopt the self-attention mechanism in the Transformer architecture to more flexibly capture long-range dependencies in the sequence data and improve the model's ability to model complex time series.

[0094] In some embodiments, in the implicit state generation module, based on the first output and the second output, combining normalization and enhanced connection to obtain the implicit state includes:

[0095] where, is the implicit state, is the weight parameter of the enhanced connection; The running environment perception modeling module improves the running environment perception estimation modeling in the hidden layer dimension through a pointwise feed-forward network to obtain the third output, including:

[0096] where, is the third output, 、 , and are trainable parameters, is an activation function; The perception learning vector generation module obtains the running environment perception learning vector based on the implicit state and the third output through normalization and enhanced connection, including:

[0097] Among them, is the running environment perception learning vector, is the weight parameter of the enhanced connection.

[0098] It should be noted that the implicit state generation module obtains the implicit state based on the first output and the second output, combined with normalization and enhanced connection. Here, the implicit state refers to the device state implicitly represented through the internal mechanism of the model. It integrates various information and can more comprehensively reflect the actual operating conditions of the device. Normalization is a data processing method used to adjust the scale of data to conform to specific distribution characteristics, such as a mean of 0 and a variance of 1. Enhanced connection is a network structure design that enhances the information transfer efficiency and the learning ability of the model by adding direct connections between different layers.

[0099] Specifically, the generation of the implicit state can be expressed as:

[0100] Among them, is the implicit state, represents the normalization operation, is the weight parameter of the enhanced connection, used to adjust the relative importance between the first output and the second output .

[0101] Specifically, the normalization operation can adopt layer normalization, which normalizes each feature of each sample separately and is suitable for processing sequence data. The weight parameter of the enhanced connection can be automatically learned through the backpropagation algorithm during the training process, and the initial value can be set to 1.0, indicating that the same weight is given to the two outputs at the initial stage of training.

[0102] In addition, the running environment perception modeling module improves the running environment perception estimation modeling in the hidden layer dimension through a pointwise feedforward network to obtain the third output, which can be expressed as:

[0103] Among them, It represents a pointwise feed-forward network, which usually contains two linear transformation layers with a ReLU activation function in between. Specific parameters such as the output dimension of the linear transformation layer can be set according to the size of the hidden layer. For example, if the dimension of the hidden layer is 512, the output dimension of the linear transformation layer can be set to 2048 and then changed back to 512.

[0104] Preferably, in order to further optimize the generation of the implicit state, a gating mechanism can be added after the normalization operation, which can dynamically adjust the flow of information according to the current input. For example, a sigmoid activation function can be introduced to generate a gating signal for controlling the weight parameters of the enhanced connection. The actual impact.

[0105] Furthermore, in an alternative, the activation function in the pointwise feed-forward network can be replaced with GELU (Gaussian Error Linear Unit), which performs better than ReLU in some cases, especially when dealing with deep networks. In addition, to improve the generalization ability of the model, dropout technology can be introduced in the pointwise feed-forward network, and the dropout rate can be set to 0.1 to prevent overfitting. At the same time, residual connections can be considered to directly add the input of the pointwise feed-forward network to the output to maintain the integrity of information.

[0106] In some embodiments, the running environment perception learning vector in the state prediction module enters a multi-layer perceptron, and obtaining the prediction scores of multiple running state segments within the corresponding time step includes:

[0107] Where, is the prediction score of the th running state segment at time ; is the running environment perception learning vector of the th running state segment, is the running environment perception learning vector of the th running state segment, is the exponential function with the natural constant as the base.

[0108] It should be noted that the learning vector of the operating environment perception in the state prediction module enters the multi-layer perceptron, and the prediction scores of multiple operating state segments within the corresponding time step are obtained. Here, the learning vector of the operating environment perception refers to the vector representation obtained through model learning, which can reflect the state of the device under a specific operating environment. The multi-layer perceptron (MLP) is a feed-forward neural network composed of multiple fully connected layers, which is used to perform non-linear transformation on the input data to achieve complex function mapping. The prediction score refers to the value output by the model, representing the possibility of the device in different operating state segments, and these scores can be used to determine the most likely operating state of the device.

[0109] Specifically, the learning vector of the operating environment perception is input into the MLP. After a series of non-linear transformations, the prediction scores of different operating state segments within each time step are output , where the prediction score of each state segment is expressed as

[0110] is the learning vector of the operating environment perception at time step , is the weight vector of the th operating state segment, and

[0111] is the total number of operating state segments.

[0112] Specifically, the MLP can be designed to include two hidden layers, and the number of neurons in each hidden layer can be set according to the complexity of the actual problem. For example, if the dimension of the input vector is 128, the first hidden layer can be set to 256 neurons, and the second hidden layer can be set to 128 neurons. The activation function can adopt ReLU because it performs well in dealing with non-linear problems and can effectively avoid the problem of gradient disappearance. The calculation of the prediction score can adopt the softmax function, which can normalize the output value into a probability distribution for easy interpretation and use. For example, for a problem with 5 operating state segments, the softmax function will output a 5-dimensional vector, and the value of each dimension represents the probability that the device is in the corresponding state segment.

[0112] Preferably, in order to improve the prediction accuracy, a batch normalization layer can be added between the hidden layers of the MLP to reduce the internal covariate shift, accelerate the training process and improve the stability of the model. In addition, the dropout technique can be adopted to prevent overfitting, and the dropout rate can be set to 0.5.

[0113] Furthermore, in the alternative solution, in addition to the MLP, a Convolutional Neural Network (CNN) can also be considered to process the running environment perception learning vector. Especially when the input vector has spatial correlation, the CNN can better capture local features. Additionally, to further optimize the calculation of the prediction score, a temperature parameter can be introduced into the softmax function, that is

[0114] The temperature parameter can control the sharpness of the probability distribution. A smaller value can make the probability distribution sharper and improve the confidence of the prediction.

[0115] In some embodiments, the running environment perception estimation modeling is optimized for probability through the binary cross-entropy loss function, expressed as:

[0116] where, is the binary cross-entropy loss function, is the prediction score, is the actual score label, and each time step .

[0117] It should be noted that the running environment perception estimation modeling optimizes the probability through the binary cross-entropy loss function. This process aims to adjust the model parameters by minimizing the loss function so that the prediction score output by the model is as close as possible to the true label. The binary cross-entropy loss function here is a loss function commonly used in binary classification problems, which measures the difference between the model's predicted probability and the true label. Optimizing the probability means adjusting the model parameters to make the model's prediction of the device's running state more accurate.

[0118] Specifically, the binary cross-entropy loss function can be expressed as:

[0119] where, is the number of samples, is the true score label of the th sample, is the score predicted by the model.

[0120] More specifically, and in the binary cross-entropy loss function represent the true label and the predicted label respectively. The true label is a binary value indicating whether the device is in a certain specific running state segment. For example, 1 indicates that the device is in this state segment, and 0 indicates not. The predicted label It is the predicted probability output by the model, indicating the likelihood that the device is in this state segment.

[0121] Furthermore, in practical applications, at each time step there is a corresponding and , and the loss function will evaluate the prediction results for all time steps. For example, if a device is in a specific state segment at 3 out of 10 time steps, then the for these 3 time steps is 1, and the rest are 0. The goal of the model is to minimize , that is, by adjusting the model parameters to make as close as possible to .

[0122] Preferably, in order to further improve the training effect of the model, class weights can be introduced into the binary cross-entropy loss function to handle the problem of class imbalance. For example, if the frequency of some running state segments is much higher than that of other state segments, the sensitivity of the loss function to different state segments can be balanced by assigning smaller weights to these state segments and larger weights to the state segments with lower frequencies.

[0123] Even further, the early stopping technique can be adopted to prevent overfitting, that is, during the training process, if the loss on the validation set no longer decreases after a certain number of iterations, the training is stopped. In an alternative solution, in addition to the binary cross-entropy loss function, focal loss can also be considered. It gives higher weights to difficult-to-classify samples, which helps the model pay more attention to these samples, thereby improving the overall classification performance. At the same time, a learning rate decay strategy can be adopted to gradually reduce the learning rate as the training progresses to help the model more finely adjust the parameters in the later stage of training.

[0124] In some embodiments, the second input is processed by a multi-modal Gaussian decoder , and the second input consists of a sparse joint encoding, high-level interaction features, parameters formed by the candidate running state segment with the highest evaluation score, and a latent vector sampled from a multivariate normal distribution, to obtain the operating trajectory of the mechanical equipment in continuous time steps as the trajectory, specifically expressed as:

[0125] where, is the mean, is the covariance matrix, is the latent vector sampled from the multivariate normal distribution ; The second input consists of the following parts:

[0126] Among them, is sparse joint coding, is an advanced interaction feature guided by the operating environment, is a latent vector sampled from a multivariate normal distribution; and perform cross-attention generation , which is formed by connecting the candidate operating state segments with the highest evaluation scores.

[0127] It should be noted that the trajectory decoding module processes the second input through a multimodal Gaussian decoder to obtain the operating trajectory of the mechanical equipment at continuous time steps. Here, the multimodal Gaussian decoder is a decoder that can process various types of data and output parameters conforming to the Gaussian distribution, such as the mean and covariance matrix, so as to generate continuous trajectory data. The second input refers to the input data that combines sparse joint coding, advanced interaction features, the parameters formed by the candidate operating state segments with the highest evaluation scores, and the latent vector sampled from a multivariate normal distribution.

[0128] Specifically, the multimodal Gaussian decoder receives the second input , which is composed of sparse joint coding , advanced interaction features , the parameters formed by the candidate operating state segments with the highest evaluation scores and the latent vector sampled from a multivariate normal distribution , that is , where is the advanced interaction feature guided by the operating environment. The decoder outputs the operating trajectory of the mechanical equipment at continuous time steps , and each point of the trajectory can be represented as , where is the mean, is the covariance matrix, is the latent vector sampled from the multivariate normal distribution .

[0129] More specifically, the multimodal Gaussian decoder can be designed to include multiple fully connected layers, and each fully connected layer is followed by an activation function, such as ReLU or Tanh, to introduce non-linearity. For example, three fully connected layers can be set. The output dimension of the first fully connected layer is 512, the second is 256, and the third outputs the mean and the covariance matrix with the same dimension as the trajectory points. The covariance matrix can be designed as a diagonal matrix to simplify the calculation, where the elements on the diagonal represent the variances of each dimension. For the latent vector For sampling, the reparameterization trick can be used, that is , where is a noise vector sampled from the standard normal distribution, which can ensure that the gradient can be backpropagated through the sampling process to update the parameters of the decoder during training.

[0130] Preferably, in order to improve the accuracy and smoothness of trajectory decoding, an attention mechanism can be introduced into the multi-modal Gaussian decoder to enable the decoder to pay more attention to important features in the input. For example, the self-attention mechanism can be used to make the decoding at each time step consider the information of the entire input sequence, rather than just the input at the current time step.

[0131] Furthermore, a recurrent neural network (RNN) or a long short-term memory network (LSTM) can be used to replace the fully connected layer to better capture the temporal dependencies in time series data. In the alternative solution, the covariance matrix can be designed as a full matrix instead of a diagonal matrix. Although this will increase the computational complexity, it can more accurately represent the correlation between trajectory points. At the same time, in order to further optimize the sampling process of the latent vector, a complex prior distribution under the variational autoencoder (VAE) framework, such as a mixture of Gaussian distributions, can be adopted to generate more diverse trajectory samples.

[0132] The above embodiments of the present invention have the following beneficial effects: First, the data acquisition module can accurately obtain the status and operating environment information of the mechanical equipment, providing a reliable data source for subsequent processing. The feature encoding module combines a gated recurrent unit and a multi-layer perceptron to effectively extract high-dimensional features and enhance the data representation ability. The feature fusion module can achieve the deep fusion of mechanical equipment information and operating environment information through the multi-head self-attention mechanism and activation function, forming a sparse joint encoding, providing strong support for equipment status analysis in complex scenarios. The cloud computing model construction module introduces a trainable enhancement layer and a dynamic weight adjustment mechanism to optimize the model structure in real time and improve the adaptability and prediction accuracy for different operating states. The state interaction module enables the sparse joint encoding to obtain a high-level interaction state based on the pre-trained cloud computing model, further improving the interaction quality and depth of information. The design of the MechLayer in the state synchronization module can accurately synchronize the operating status and operating environment information of the target mechanical equipment to ensure the consistency and timeliness of the data. The implicit state generation module can generate a stable implicit state by combining normalization and enhanced connection, providing a stable input for state prediction. The operating environment perception modeling module can more accurately reflect the impact of the operating environment on the equipment status by improving the operating environment perception estimation modeling in the hidden layer dimension through a point-by-point feedforward network. The perception learning vector generation module obtains an operating environment perception learning vector based on the implicit state and the third output through normalization and enhanced connection, providing richer feature information for state prediction.

[0133] Secondly, the state prediction module enables the operation environment perception learning vector to enter a multi-layer perceptron, and can obtain the prediction scores of multiple operation state segments within the corresponding time step, realizing the accurate prediction of the device operation state and providing a scientific basis for preventive maintenance. The trajectory decoding module processes the second input through a multi-modal Gaussian decoder, and can obtain the operation trajectory of the mechanical device in continuous time steps, providing an intuitive visualization result for the operation monitoring and fault diagnosis of the device. The collaborative work of these modules can not only improve the efficiency and accuracy of device management, but also reduce the maintenance cost and fault risk of the device, and enhance the stability and reliability of the entire production system.

[0134] Furthermore, the storage medium of the embodiment of the present application stores program instructions capable of implementing all the above methods. Among them, the program instructions can be stored in the above storage medium in the form of a software product, including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in various embodiments of the present application. The foregoing storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs, or terminal devices such as computers, servers, mobile phones, and tablets.

[0135] The above description is only some preferred embodiments of the present invention and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the (but not limited to) technical features having similar functions disclosed in the embodiments of the present invention.

Claims

1. A digital mechanical equipment data management system based on cloud computing, characterized in that: include: Data acquisition module, used to obtain information about the status of mechanical equipment and operating environment; A feature encoding module, including an embedding network, wherein the embedding network is composed of a gated recurrent unit and a multi-layer perceptron, and is used to encode the state of the mechanical equipment and the operating environment information and extract high-dimensional features; feature A fusion module is used to fuse high-dimensional features to exchange mechanical equipment information and operating environment information, and output sparse joint coding; The cloud computing model construction module is used to construct the cloud computing model, introduce a trainable enhancement layer, and use a dynamic weight adjustment mechanism to update the attention layer to form a cloud computing model; the state interaction module is used to enable sparse joint coding to obtain advanced interaction states based on the pre-trained cloud computing model; A state synchronization module synchronizes the operating state and operating environment information of the target mechanical equipment through MechLayer at each time step, wherein the MechLayer includes a Mech block, a normalization layer, a point-by-point feedforward network, and an enhanced connection; the MechLayer normalizes the first input formed by the high-level interactive state and the partial sparse joint coding to obtain a first output; the first output is mapped through the Mech block to obtain a second output; an implicit state generation module obtains an implicit state based on the first output and the second output in combination with normalization and enhanced connection; An operating environment perception modeling module improves the operating environment perception estimation modeling in the hidden layer dimension through a point-by-point feedforward network to obtain a third output; The perception learning vector generation module obtains the operating environment perception learning vector through normalization and enhanced connection based on the implicit state and the third output; the state prediction module allows the operating environment perception learning vector to enter the multi-layer perceptron to obtain the prediction scores of multiple operating state segments within the corresponding time step, the prediction score of the operating state segment closest to the actual operating state position is 1, and the prediction scores of the remaining operating state segments are 0; The trajectory decoding module processes the second input through a multimodal Gaussian decoder, where the second input consists of sparse joint coding, high-level interactive features, parameters formed by candidate operating state segments with the highest measurement scores, and latent vectors sampled from a multivariate normal distribution to obtain the operating trajectory of the mechanical equipment in continuous time steps.

2. The digital mechanical equipment data management system based on cloud computing according to claim 1 is characterized in that: The feature fusion module fuses high-dimensional features to exchange mechanical equipment information and operating environment information, and outputs sparse joint coding including: fusing features between mechanical equipment through a multi-head self-attention mechanism and an activation function; Through the multi-head attention mechanism and enhanced connection, the mechanical equipment and the operating environment are updated to achieve feature fusion between mechanical equipment and feature fusion between mechanical equipment and the operating environment; It is expressed as: ； in, and The high-dimensional features of the mechanical equipment status and operating environment information are proposed after embedding the network. is a multi-head self-attention mechanism, is the activation function; Will Connect them to form a sparse joint code .

3. The digital mechanical equipment data management system based on cloud computing according to claim 1 is characterized in that: The cloud computing model building module uses a dynamic weight adjustment mechanism to update the attention layer, including: Assume the parameter of dynamic weight adjustment is , No. Attention layer network The input is , whose dimensions are , the output is Its dimension is , adjust the network by The weight matrix in : ； in, ,and The element values ​​of are dynamically adjusted based on performance feedback during training; During the training process, Stay trainable, Updates are made based on the gradient of the loss function.

4. The digital mechanical equipment data management system based on cloud computing according to claim 3 is characterized in that: The forward transfer function of dynamic weight adjustment is expressed as: Sparse joint coding Input to cloud computing model The output is: ； Transformed by multi-layer perceptron to match , thereby generating the final high-level interaction state .

5. The digital mechanical equipment data management system based on cloud computing according to claim 1 is characterized in that: The MechLayer in the state synchronization module normalizes the first input to obtain a first output; and maps the first output through the Mech block to obtain a second output, including: ； in, For the second output, For normalization, Represents the output after processing by the Mech block. is the first input.

6. The digital mechanical equipment data management system based on cloud computing according to claim 5 is characterized in that: The processing of the Mech block includes expanding the input dimension using linear projection with a factor of , generating two different representations of parallel branches, respectively denoted as and ; Capturing runtime environment awareness dependencies through one-dimensional convolution and activation functions ,Will The linear projections are , and , For transformation and the structured state matrix generate and ; Selective state model acceptance , and As input, generate modified runtime environment awareness features ; Introduce activation function to generate gating signal that can filter irrelevant information ; and Output after multiplication and linear projection .

7. The digital mechanical equipment data management system based on cloud computing according to claim 6 is characterized in that: The implicit state generation module obtains the implicit state based on the first output and the second output by combining normalization and enhanced connection, including: ； in, is an implicit state, is the weight parameter for enhancing the connection; The operating environment perception modeling module improves the operating environment perception estimation modeling in the hidden layer dimension through a point-by-point feedforward network, and obtains a third output including: ； in, For the third output, , , and is a trainable parameter, is the activation function; The perception learning vector generation module obtains the operating environment perception learning vector based on the implicit state and the third output through normalization and enhanced connection, including: ； in, To learn vectors for environment awareness, is the weight parameter for enhancing the connection.

8. The digital mechanical equipment data management system based on cloud computing according to claim 7 is characterized in that: The operating environment perception learning vector in the state prediction module enters the multi-layer perceptron to obtain the prediction scores of multiple operating state segments in the corresponding time step, including: ； in, For in time No. The prediction score of each operating state segment; For the The operating environment perception learning vector of the operating state segment, For the The operating environment perception learning vector of the operating state segment, The natural constant An exponential function with base .

9. The digital mechanical equipment data management system based on cloud computing according to claim 8 is characterized in that: The operating environment perception estimation model optimizes the probability through the binary cross entropy loss function, which is expressed as: ； in, is the binary cross entropy loss function, is the prediction score, is the actual score label, each time step .

10. The digital mechanical equipment data management system based on cloud computing according to claim 1 is characterized in that: Process the second input through the multimodal Gaussian decoder The second input is composed of sparse joint coding, high-level interactive features, parameters formed by the candidate running state segments with the highest test scores, and latent vectors sampled from multivariate normal distribution, and the running trajectory of the mechanical equipment in continuous time steps is obtained. is the trajectory, which is specifically expressed as: ； in, is the mean, is the covariance matrix, For a multivariate normal distribution Sampled latent vectors; Second input It consists of the following parts: ； in, is sparse joint coding, Advanced interactive features for runtime guidance, is the latent vector sampled from a multivariate normal distribution; and Performing Cross-Attention Generation , formed by connecting the candidate operating state segments with the highest test scores.

Images