Distributed precision point inspection and early warning method and system supporting mobile web access

Abstract

The application relates to a distributed precision point inspection early warning method and system supporting mobile Web access, which comprises the following steps: constructing a parameter space correlation matrix and a failure space correlation matrix based on equipment mechanism and historical data as prior knowledge; processing strong fluctuation signals by using an integrated heterogeneous width network to output parameter prediction values; calculating prediction data reconstruction errors by using a convolutional autoencoder, and obtaining health indexes by weighted fusion according to the failure space correlation matrix; after the latent representation of the encoder is normalized, the latent representation is fused with core statistical features to construct composite features, the Mahalanobis distance is calculated as an update discrimination index, data states are marked, and effective data is stored through an adaptive buffer mechanism; when the judgment model is to be updated, an integrated incremental mechanism is used to add nodes to the prediction network, and an adversarial regularization mechanism is used to fine-tune parameters of the encoder, so that online updating is completed; an edge unit is provided with a Web server, early warning information, prediction data and model states are visually displayed through a mobile terminal.

Description

Technical Field

[0001] This application relates to the field of equipment monitoring technology, and in particular to a distributed precision inspection and early warning method and system that supports mobile web access. Background Technology

[0002] Industrial moving equipment such as compressors, pumps, and fans are key basic equipment in fields such as petrochemicals, energy and power, and intelligent manufacturing. Their operating status directly affects the continuity and safety of production systems. With the rapid development of the Industrial Internet and intelligent manufacturing, equipment operating status monitoring technology has gradually evolved from traditional manual periodic inspections to online monitoring and intelligent early warning.

[0003] Traditional industrial equipment condition monitoring relies heavily on manual periodic inspections or offline vibration measurements, which suffers from discontinuous data acquisition, delayed response, and reliance on expert experience for analysis. While online monitoring systems can provide continuous data, their high deployment costs, system complexity, and high data analysis barriers hinder widespread adoption. Existing systems are mostly focused on data acquisition, lacking edge analytics and intelligent early warning capabilities, and are difficult to integrate efficiently with mobile or distributed control systems.

[0004] Based on this, this application provides a distributed precision inspection early warning method and system that supports mobile web access. Summary of the Invention

[0005] To address the shortcomings of traditional industrial equipment condition monitoring, which relies heavily on manual periodic inspections or offline vibration measurements, resulting in discontinuous data acquisition, delayed response, and reliance on expert experience for analysis, online monitoring systems, while providing continuous data, suffer from high deployment costs, system complexity, and high data analysis barriers, hindering widespread adoption. Existing systems primarily focus on data acquisition, lacking edge computing and intelligent early warning capabilities, and struggle to integrate efficiently with mobile devices or distributed control systems. This application provides a distributed precision inspection and early warning method and system that supports mobile web access.

[0006] Firstly, this application provides a distributed precision inspection and early warning method supporting mobile web access, which adopts the following technical solution: including: Multiple edge monitoring units are deployed at the industrial dynamic equipment site to collect operating parameters in real time; a parameter space correlation matrix and a fault space correlation matrix are constructed based on the equipment mechanism and historical data, and are solidified into the edge units as prior knowledge; Based on the operating parameters, the input features of the prediction model are determined according to the parameter space correlation matrix; empirical mode decomposition is performed on strong fluctuation signals, and one-dimensional convolution processing is performed on slow-changing signals. The extracted features are input into the recurrent width learning system to construct an integrated heterogeneous width network and output the future predicted values ​​of each parameter. A convolutional autoencoder is trained based on historical normal data. The future predicted value is input into the trained encoder, the reconstruction error of each parameter is calculated, and the values ​​are weighted and fused according to the fault space correlation matrix to obtain the health indicators of each fault type. When the health indicators exceed the preset threshold, an early warning is generated and stored. The latent representation of the convolutional autoencoder is normalized, and statistical features are extracted and core features are selected from the sliding window of real-time data. The two are then fused to construct a composite feature. The Mahalanobis distance from the composite feature to the center of the training sample is calculated as the update discrimination index. The data status is marked as normal, warning, or pending update according to the preset threshold, and continuous data of the warning status is stored through an adaptive buffer mechanism. When the model's data status is determined to be pending update, an integrated incremental mechanism is adopted for the integrated heterogeneous width network. This mechanism adds key nodes and updates the weights of the added nodes to ensure compatibility with both old and new data. An adversarial regularization mechanism is adopted for the convolutional autoencoder. The loss function, composed of the reconstruction error of the new data, the encoding cosine distance, and the adaptive regularization term, is used to fine-tune the parameters and complete the online update. The edge unit has a built-in web server that visualizes early warning information, prediction data, and model status through a mobile terminal browser, providing real-time monitoring, historical tracking, and alarm setting functions.

[0007] Preferably, the deployment of multiple edge monitoring units at the industrial equipment site to collect operating parameters in real time; and the construction of parameter space correlation matrices and fault space correlation matrices based on equipment mechanisms and historical data, which are then embedded in the edge units as prior knowledge, including: Based on the physical mechanism of the equipment and historical maintenance reports, the core relevant parameters of typical faults are extracted, and a parameter correlation matrix and a fault correlation matrix at the knowledge level are constructed. Spearman's rank correlation coefficient is calculated using historical normal operation data to obtain the parameter correlation matrix at the data level. The parameter correlation matrix at the knowledge level is then fused with the parameter correlation matrix at the data level to obtain the parameter space correlation matrix. Each parameter type is then labeled according to the parameter change characteristics. By using limited fault data and employing a multidimensional dimensionality reduction method, a quantitative correlation matrix between parameters and faults is obtained. The knowledge-level fault correlation matrix is ​​then fused with the quantitative correlation matrix to obtain a fault space correlation matrix.

[0008] Preferably, the step of determining the input features of the prediction model based on the parameter space correlation matrix according to the operating parameters; performing empirical mode decomposition on strong fluctuation signals, performing one-dimensional convolution processing on slowly changing signals, inputting the extracted features into a recurrent width learning system, constructing an integrated heterogeneous width network, and outputting the future predicted values ​​of each parameter includes: Based on the correlation between parameters recorded in the parameter space correlation matrix, the corresponding input feature set is determined for each running parameter to be predicted. The input feature set includes other running parameters that are strongly correlated with the corresponding parameter. The real-time acquired operating parameters are divided into two categories: strong fluctuation signals and slow change signals. Strong fluctuation signals include vibration signals, while slow change signals include temperature and pressure signals. Empirical mode decomposition is performed on strong fluctuation signals to extract several intrinsic mode components and their instantaneous frequencies as fluctuation features; Slowly changing signals are processed by a one-dimensional convolutional neural network. The convolution operation includes moving average filtering and second-order differential filtering to extract trend change features and abrupt change features. After concatenating the fluctuation feature and the trend change feature, the concatenation is input into the cascaded structure of the loop width learning system. Temporal features are extracted through recursive propagation, and at the same time, randomly generated weights are used to map the fluctuation feature and the trend change feature into enhanced features. The temporal features and the enhanced features are concatenated, and the output weights are solved by pseudo-inverse operation to construct an integrated heterogeneous width network model. The integrated heterogeneous width network model is used to predict the key operating parameters and output the future predicted values ​​of each parameter.

[0009] Preferably, the step of training a convolutional autoencoder based on historical normal data, inputting the future predicted value into the trained encoder, calculating the reconstruction error of each parameter, and weighting and fusing them according to the fault space correlation matrix to obtain health indicators for each fault type; when the health indicator exceeds a preset threshold, an early warning is generated and stored, including: Historical normal operation data is selected as training samples to train the convolutional autoencoder. By minimizing the reconstruction error, the encoder learns the potential representational patterns of normal data. The predicted future values ​​of each parameter are input into the trained convolutional autoencoder. After encoding and decoding operations, the reconstructed data is obtained. The reconstruction error of each running parameter is calculated based on the symmetric mean absolute percentage error to eliminate the difference in dimensions and improve stability. Read the fault space correlation matrix, and sum the correlation weights of each fault type and each parameter in the matrix with the reconstruction error of the corresponding parameter to obtain the health index of each fault type. Using the Three Sigma principle, the mean and standard deviation of the reconstruction error of historical normal data are used as the basis to set the warning thresholds for health indicators of each fault type. When the health indicator of any fault type exceeds the corresponding warning threshold, the device is determined to be in a sub-healthy or faulty state, and a warning message containing the fault type, abnormal parameters and current health score is generated and stored in the local database of the edge monitoring unit.

[0010] Preferably, the process involves normalizing the latent representation of the convolutional autoencoder, extracting statistical features from the real-time data sliding window and selecting core features, fusing the two to construct a composite feature; calculating the Mahalanobis distance from the composite feature to the training sample center as an update discrimination index; marking the data state as normal, warning, or pending update according to a preset threshold; and storing continuous data in the warning state through an adaptive buffering mechanism, including: The encoder part is extracted from the trained convolutional autoencoder, and the running parameters at the current time are input into the encoder to generate a low-dimensional latent representation. The low-dimensional latent representation is input into a two-layer fully connected network with a standard Gaussian distribution as the target distribution. The low-dimensional latent representation is mapped to a normalized representation that conforms to a Gaussian distribution by minimizing the maximum mean difference. Running parameters for consecutive time periods are extracted from the real-time running data stream using a sliding window approach. Various statistical features, including mean, variance, and autocorrelation coefficient, are calculated for the data within the window to form an initial set of statistical features. Using the Relief algorithm, guided by the parameter space correlation matrix and the fault space correlation matrix, core statistical features sensitive to changes in operating conditions are selected from the initial set of statistical features. The core statistical features are input into the embedding network, and mapped to embedding features of the same dimension as the normalized representation through a two-layer fully connected network. The embedding network is trained by minimizing the Euclidean distance so that the embedding features approximate the spatial distribution of the normalized representation. The normalized representation and the embedded features are respectively L2 normalized and then averaged to obtain the composite features of real-time samples that fuse model performance and data distribution. Using the historical normal data used by the convolutional autoencoder as a benchmark, the composite features of each historical sample are calculated and the mean vector and covariance matrix are obtained; the Mahalanobis distance from the composite features of the current real-time sample to the mean vector is calculated, and the corresponding Mahalanobis distance is defined as the update discriminant index. Based on the Mahalanobis distance distribution of the composite features of historical samples, the three Sigma rule is used to set the warning threshold, and the maximum Mahalanobis distance of the historical samples is set as the threshold to be updated; according to the comparison results of the update discrimination index with the warning threshold and the threshold to be updated, the current data status is marked as normal, warning, or pending update. When the data status first enters the warning state, an adaptive buffer is initialized and begins storing subsequent continuous data. The buffer is divided into three sub-areas according to its capacity. When the first sub-area is full, the system determines whether to continue storing or clear the buffer based on the sample ratio of warning status to pending update status within the sub-area. When the second sub-area is full, the system determines whether to immediately trigger an update or wait for the third sub-area to fill based on the status distribution. If a health indicator is detected to exceed a threshold warning during the buffering process, the buffer is immediately cleared to avoid using faulty data to update the model. Finally, the system outputs the filtered continuous warning status data through the buffer.

[0011] Preferably, the Relief algorithm is used, guided by the parameter space correlation matrix and the fault space correlation matrix, to screen out core statistical features sensitive to changes in operating conditions from the initial statistical feature set, including: Based on the parameter space correlation matrix and the fault space correlation matrix, historical operating data of the same equipment at different times are selected to construct a cross-condition reference dataset, and operating data of different equipment at the same time are selected to construct a cross-equipment reference dataset. Both reference datasets contain normal operation data and performance degradation data, and are labeled with binary classification labels. For the two reference datasets, various statistical features within a sliding window are extracted for each data sample to form an initial statistical feature vector including mean, variance, and autocorrelation coefficient. The Relief algorithm is used to calculate the weight of each statistical feature in the two reference datasets. The weight is calculated as follows: a sample is randomly sampled from the reference dataset, and the nearest neighbors of the same class and the opposite class of the target sample are found respectively. The Manhattan distance difference between the current statistical feature and the two nearest neighbors is calculated. After iterative accumulation, the weight value of the target feature in the corresponding reference dataset is obtained. After iterating through all samples, the feature weight vectors of the cross-condition reference dataset and the cross-device reference dataset are obtained respectively. Calculate the mean of the two feature weight vectors respectively, filter out the statistical features in each dataset whose weight is greater than the mean weight, and take the union of the two sets of filtering results to obtain the core statistical feature set that is applicable to both cross-working-condition scenarios and cross-device scenarios. The core statistical feature set is embedded in the edge monitoring unit for the construction of composite features in subsequent real-time data streams, ensuring the stability and generalization ability of the model update discrimination method in time-varying environments and device migration scenarios.

[0012] Preferably, when the data state of the model is determined to be pending update, an integrated incremental mechanism is adopted for the integrated heterogeneous width network, which adds key nodes and updates the weights of the added nodes to ensure compatibility with both old and new data; an adversarial regularization mechanism is adopted for the convolutional autoencoder, which uses a loss function composed of the new data reconstruction error, the encoding cosine distance, and the adaptive regularization term to fine-tune the parameters and complete the online update, including: When the current data status is determined to be pending update and the buffer mechanism has stored continuous data that meets the preset conditions, the online update process is triggered, and a differentiated update mechanism is adopted for the integrated heterogeneous width network and the convolutional autoencoder respectively. For the integrated incremental update of the heterogeneous width network: the network parameters of the temporal feature extraction part of the original integrated heterogeneous width network remain unchanged, and key feature nodes are added to the enhancement feature part. The initial weights of the new nodes are randomly generated. The new data stored in the buffer is input into the original network to extract its temporal features and the original enhancement features. At the same time, supplementary features of the new data are extracted through the new nodes. The original enhancement features and the new supplementary features are concatenated to form the updated enhancement feature matrix. Based on the updated enhancement feature matrix and the original temporal features, combined with the true value of the new data, the updated overall output weight is calculated through matrix pseudo-inverse operation, so that the updated integrated heterogeneous width network can be compatible with the prediction accuracy of the new and old data without retraining the original nodes. The number of new key feature nodes does not exceed the number of original key feature nodes. When the number of new nodes reaches the upper limit and still cannot meet the accuracy requirements, the new network is retrained and the original network is replaced. Adversarial regularization update for convolutional autoencoders: The target network to be updated is constructed using the trained convolutional autoencoder parameters as initial values. New data stored in a buffer is input into the target network, and the reconstruction error is calculated. Simultaneously, the new data is input into the original convolutional autoencoder to obtain the original latent representation, and then into the target network to obtain the updated latent representation. The cosine distance between the two is calculated as the encoding loss. The adversarial evolution between the reconstruction error and the encoding loss is used as the optimization objective, prompting the target network to learn the latent distribution characteristics of the new data while minimizing the reconstruction error. An L2 regularization term based on the original network parameters is added to the loss function, and a regularization direction guidance coefficient is introduced to adaptively adjust the strength of the regularization term. This regularization direction guidance coefficient is calculated based on the first-order difference of the reconstruction error of the fitted data from the old data, and is used to dynamically adjust the parameter constraint strength according to the degree of performance degradation of the old data during model updates. The loss function is minimized using the gradient descent algorithm to complete the fine-tuning of the convolutional autoencoder parameters. After the update is completed, the updated integrated heterogeneous width network and convolutional autoencoder are deployed on the edge detection unit, and the updated model parameters are synchronized to the baseline distribution calculation used to update the discrimination index to achieve model optimization.

[0013] Secondly, this application discloses a distributed precision inspection and early warning device that supports mobile web access, which adopts the following technical solution, including: The real-time data module is used to deploy multiple edge monitoring units at the industrial dynamic equipment site to collect operating parameters in real time; it constructs parameter space correlation matrix and fault space correlation matrix based on equipment mechanism and historical data, and embeds them into the edge units as prior knowledge; The parameter prediction module is used to determine the input features of the prediction model based on the running parameters and the parameter space correlation matrix; it performs empirical mode decomposition on strong fluctuation signals and one-dimensional convolution processing on slow-changing signals, inputs the extracted features into the recurrent width learning system, constructs an integrated heterogeneous width network, and outputs the future predicted values ​​of each parameter. The parameter index module is used to train a convolutional autoencoder based on historical normal data, input the future predicted value into the trained encoder, calculate the reconstruction error of each parameter, and fuse them according to the fault space correlation matrix to obtain the health index of each fault type; when the health index exceeds the preset threshold, an early warning is generated and stored. The data discrimination module is used to normalize the latent representation of the convolutional autoencoder, extract statistical features from the real-time data sliding window and filter the core features, and fuse the two to construct composite features; calculate the Mahalanobis distance from the composite feature to the center of the training sample as the update discrimination index, mark the data status as normal, warning or waiting to be updated according to the preset threshold, and store the continuous data of the warning status through an adaptive buffer mechanism. The model update module is used to perform an integrated incremental mechanism on the integrated heterogeneous width network when the model's data status is determined to be in need of updating. This mechanism adds key nodes and updates the weights of the added nodes to ensure compatibility with both old and new data. For the convolutional autoencoder, an adversarial regularization mechanism is used to fine-tune the parameters using a loss function composed of the reconstruction error of the new data, the encoding cosine distance, and the adaptive regularization term, thus completing the online update. The real-time access module is used to embed a web server in the edge unit to visualize early warning information, prediction data and model status through a mobile terminal browser, providing real-time monitoring, historical tracking and alarm setting functions.

[0014] Thirdly, this application also provides a control device, the device comprising: It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed, such as the distributed precision inspection and early warning method supporting mobile web access described above.

[0015] Fourthly, this application also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above for a distributed precision inspection and early warning method supporting mobile web access.

[0016] In summary, the system in this application constructs parameter space correlation matrices and fault space correlation matrices based on equipment mechanisms and historical data, which are then embedded in edge units as prior knowledge to guide subsequent modeling and feature selection. Secondly, using real-time parameters as input, an integrated heterogeneous width network is employed to perform empirical mode decomposition on strong fluctuation signals and one-dimensional convolution processing on slowly changing signals. Multi-scale features are extracted and input into a recurrent width learning system, outputting future predicted values ​​for each parameter. Then, a convolutional autoencoder is used to calculate the reconstruction error of the predicted data, and these errors are weighted and fused according to the fault space correlation matrix to obtain health indicators for each fault type, achieving unsupervised health status assessment. Based on this, composite features are constructed by normalizing the encoder's latent representation and fusing core statistical features. Mahalanobis distance is calculated as an update discrimination index, and data states are marked as normal, warning, or pending update according to a preset threshold. Valid data is stored through an adaptive buffer mechanism. When the model is determined to be pending update, an integrated incremental mechanism is used to add key nodes and update weights in the integrated heterogeneous width network. An adversarial regularization mechanism is used in the convolutional autoencoder to fine-tune parameters using the new data reconstruction error, encoding cosine distance, and adaptive regularization terms, completing the online update. Finally, the edge unit's built-in web server visualizes early warning information, predicted data, and model status through a mobile browser, enabling remote mobile access. This combination of edge computing and lightweight models significantly reduces system deployment costs and network bandwidth pressure; the fusion of prior knowledge and incremental update mechanisms endow the model with cross-device and cross-operating-condition adaptive capabilities; mobile web access improves operational response speed and convenience; and the construction of health indicators through the fusion of mechanisms and data enhances the interpretability and credibility of fault diagnosis, ultimately achieving intelligent and universal industrial equipment condition monitoring. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a distributed precision inspection and early warning method that supports mobile web access.

[0018] Figure 2 This is a structural block diagram of a distributed precision inspection and early warning device that supports mobile web access. Detailed Implementation

[0019] The following combination Figures 1-2 This application will be described in further detail.

[0020] This application provides a distributed precision inspection and early warning method that supports mobile web access. The execution entity is a control system, and it mainly includes the following: 1. Distributed edge computing architecture Deploying AMS Asset Monitor units at the equipment site supports up to 12 CHARM channels and is compatible with vibration sensors (such as ICP, eddy current, and seismic sensors) and process quantity sensors (such as RTD, thermocouple, 4-20mA HART, etc.) to achieve local data acquisition and real-time analysis, reducing the burden of data transmission and central processing.

[0021] 2. Built-in intelligent analysis and early warning rule base The system incorporates various equipment health prediction rules, including those for imbalance, misalignment, looseness, blade overrun, turbulence, gear failure, bearing defects, lubrication problems, motor failure, and oil film oscillation. It supports ISO 20816 default alarm settings, and users can customize alarm thresholds to improve warning accuracy.

[0022] 3. Mobile Web Access and Visual Interface The unit has a built-in web server, allowing users to directly access the system interface via browsers on mobile devices such as phones and tablets to view asset health scores (0-100%), vibration waveforms, spectrum, historical trends, etc., enabling remote monitoring and diagnosis.

[0023] 4. Multi-protocol data integration and system collaboration It supports interfaces with OPC UA, Modbus TCP / IP, and Plantweb Optics platforms, enabling seamless integration with DCS (such as DeltaV), PLC, and other systems. It also supports MTP module export, facilitating collaboration with distributed control systems.

[0024] 5. Network security and lifecycle management The system design follows the principles of network security lifecycle management, and supports access authentication, communication encryption, and whitelisting mechanisms to ensure the security of data transmission and system access.

[0025] Reference Figure 1 The embodiments of this application include at least steps S10 to S60.

[0026] S10 deploys multiple edge monitoring units at the industrial dynamic equipment site to collect operating parameters in real time; constructs parameter space correlation matrix and fault space correlation matrix based on equipment mechanism and historical data, and solidifies them as prior knowledge in the edge units.

[0027] S20, based on the operating parameters, determines the input features of the prediction model according to the parameter space correlation matrix; performs empirical mode decomposition on strong fluctuation signals, performs one-dimensional convolution processing on slowly changing signals, inputs the extracted features into the recurrent width learning system, constructs an integrated heterogeneous width network, and outputs the future predicted values ​​of each parameter.

[0028] S30 trains a convolutional autoencoder based on historical normal data, inputs future predicted values ​​into the trained encoder, calculates the reconstruction error of each parameter, and weights and fuses them according to the fault space correlation matrix to obtain health indicators for each fault type; when the health indicators exceed the preset threshold, an early warning is generated and stored.

[0029] S40 normalizes the latent representation of the convolutional autoencoder, extracts statistical features from the real-time data sliding window and selects core features, and merges the two to construct composite features; calculates the Mahalanobis distance from the composite features to the center of the training samples as the update discrimination index, marks the data status as normal, warning or pending update according to the preset threshold, and stores continuous data of the warning status through an adaptive buffer mechanism.

[0030] S50: When the data state of the model is determined to be pending update, an integrated incremental mechanism is adopted for the integrated heterogeneous width network. This mechanism adds key nodes and updates the weights of the added nodes to ensure compatibility with both new and old data. An adversarial regularization mechanism is adopted for the convolutional autoencoder. The loss function, composed of the reconstruction error of the new data, the encoding cosine distance, and the adaptive regularization term, is used to fine-tune the parameters and complete the online update.

[0031] The S60 has a built-in web server in the edge unit, which visualizes early warning information, prediction data and model status through a mobile terminal browser, and provides real-time monitoring, historical tracking and alarm setting functions.

[0032] Specifically, edge monitoring units are deployed at the equipment site. The system constructs parameter space correlation matrices and fault space correlation matrices based on equipment mechanisms and historical data, which are then embedded in the edge units as prior knowledge. Then, using real-time operating parameters as input, features are selected according to the parameter space correlation matrices. Empirical mode decomposition is performed on strong fluctuation signals, and one-dimensional convolution processing is applied to slowly changing signals. The extracted features are input into a recurrent width learning system to construct an integrated heterogeneous width network, outputting future predicted values ​​for each parameter. Next, a convolutional autoencoder is trained using historical normal data. The predicted values ​​are input into the encoder to calculate the reconstruction error, and then weighted and fused according to the fault space correlation matrix to obtain health indicators for each fault type. Warnings are generated when thresholds are exceeded. Simultaneously… The encoder's latent representation is normalized and fused with the core statistical features filtered by the sliding window of real-time data to construct composite features. The Mahalanobis distance from the Mahalanobis distance to the training sample center is calculated as the update discrimination index, and the data status is marked as normal, warning, or pending update. The warning status continuous data is stored through an adaptive buffer mechanism. When the model is determined to be pending update, the integrated heterogeneous width network is added with an integrated incremental mechanism to add key nodes and update the weights. The convolutional autoencoder is used with an adversarial regularization mechanism to fine-tune the parameters using the new data reconstruction error, the encoding cosine distance, and the adaptive regularization term to complete the online update. Finally, the edge unit's built-in web server visualizes the warning information, prediction data, and model status through a mobile terminal browser.

[0033] In some embodiments, step S10 specifically includes the following steps: based on the physical mechanism of the equipment and historical maintenance reports, extract the core relevant parameters of typical faults, and construct a knowledge-level parameter correlation matrix and a fault correlation matrix; calculate the Spearman rank correlation coefficient using historical normal operation data to obtain a data-level parameter correlation matrix, merge the knowledge-level parameter correlation matrix and the data-level parameter correlation matrix to obtain a parameter space correlation matrix, and label each parameter type according to the parameter change characteristics; use limited fault data to obtain a quantitative correlation matrix between parameters and faults through a multidimensional dimensionality reduction method, and merge the knowledge-level fault correlation matrix and the quantitative correlation matrix to obtain a fault space correlation matrix.

[0034] Specifically, based on the physical mechanisms of the equipment and historical maintenance reports, core parameters of typical faults are extracted to construct a knowledge-level parameter correlation matrix and a fault correlation matrix. Spearman's rank correlation coefficient is calculated using historical normal operation data to construct a data-level parameter correlation matrix. These two matrices are then fused to obtain a parameter space correlation matrix, and the change type of each parameter is labeled. A quantitative correlation matrix between parameters and faults is obtained using limited fault data through multidimensional dimensionality reduction methods, and this matrix is ​​fused with the knowledge-level fault correlation matrix to obtain a fault space correlation matrix. This deep integration of physical mechanisms and data features provides a basis for feature selection in subsequent prediction models, reduces the feature extraction burden on lightweight networks, and enhances the interpretability of the condition monitoring model.

[0035] In practice, based on the physical mechanism of the equipment and historical maintenance reports, core parameters of typical faults are extracted, and a parameter correlation matrix at the knowledge level is constructed. Fault Correlation Matrix Where m is the number of parameters and n is the number of fault types. The Spearman rank correlation coefficient is calculated using historical normal operation data to obtain the parameter correlation matrix at the data level. The calculation formula is as follows: ; in, For parameters rank, The standard deviation is given. The parameter correlation matrices from the knowledge level and the data level are fused to obtain the parameter space correlation matrix: ; ⊙ Operation means when The value in is 1 and The correlation coefficient is set to 1 when it is not less than the threshold θ, and 0 otherwise. Simultaneously, each parameter type is labeled based on its variation characteristics. Finally, the parameter space correlation matrix is ​​obtained. .

[0036] Quantitative correlation matrix between parameters and faults is obtained using limited fault data and multidimensional dimensionality reduction methods. The calculation method is as follows: ; in, Obtained through linear discriminant analysis Obtained through isometric mapping. This involves creating a fault correlation matrix at the knowledge level. With quantitative correlation matrix The fusion yields the fault space correlation matrix. , that is, use Numerical replacement in The first point, 1, demonstrates how physical mechanisms and data characteristics can be deeply integrated, providing a basis for feature selection in subsequent prediction models and enhancing the interpretability of state monitoring models.

[0037] In some embodiments, step S20 specifically includes the following steps: Based on the correlation between parameters recorded in the parameter space correlation matrix, determine the corresponding input feature set for each operating parameter to be predicted. The input feature set includes other operating parameters that are strongly correlated with the corresponding parameter. Divide the real-time acquired operating parameters into two categories: strong fluctuation signals and slowly changing signals. Strong fluctuation signals include vibration signals, and slowly changing signals include temperature and pressure signals. Perform empirical mode decomposition on the strong fluctuation signals to extract several intrinsic mode components and their instantaneous frequencies as fluctuation features. Perform convolution operations on the slowly changing signals using a one-dimensional convolutional neural network. The convolution operations include moving average filtering and second-order differential filtering to extract trend change features and abrupt change features. After concatenating the fluctuation features and trend change features, input them into the cascaded structure of the recurrent width learning system. Extract time-series features through recursive propagation, and simultaneously map the fluctuation features and trend change features to enhanced features using randomly generated weights. Concatenate the time-series features and enhanced features, solve the output weights through pseudo-inverse operation, and construct an integrated heterogeneous width network model. Use the integrated heterogeneous width network model to predict subsequent key operating parameters and output the future predicted values ​​of each parameter.

[0038] Specifically, firstly, based on the parameter space correlation matrix, strongly correlated features are selected as the input set for each parameter to be predicted. Real-time parameters are divided into highly fluctuating signals such as vibration and slowly changing signals such as temperature and pressure. For highly fluctuating signals, empirical mode decomposition is performed to extract intrinsic mode components as fluctuation features. For slowly changing signals, a one-dimensional convolutional neural network is used to extract trend change and abrupt change features through moving average and second-order differential filtering. The two types of features are concatenated and input into the cascaded structure of the recurrent width learning system to recursively extract time-series features while randomly generating enhanced features. The time-series features and enhanced features are concatenated, and the output weights are solved through pseudo-inverse operations to construct an integrated heterogeneous width network model, ultimately outputting the future predicted values ​​of each parameter. By processing multi-source heterogeneous data separately, the prediction accuracy is improved without increasing model complexity, providing an accurate data foundation for the subsequent construction of health indicators.

[0039] In some embodiments, step S30 specifically includes the following steps: selecting historical normal operation data as training samples to train the convolutional autoencoder, and minimizing the reconstruction error to enable the encoder to learn the potential representation rules of normal data; inputting the future predicted values ​​of each parameter into the trained convolutional autoencoder, and obtaining reconstructed data after encoding and decoding operations; calculating the reconstruction error of each operating parameter based on the symmetric mean absolute percentage error to eliminate dimensional differences and improve stability; reading the fault space correlation matrix, and weighting the correlation weight between each fault type and each parameter in the matrix with the reconstruction error of the corresponding parameter to obtain the health index of each fault type; using the three Sigma rule, based on the mean and standard deviation of the reconstruction error of historical normal data, setting the warning threshold of the health index of each fault type; when the health index of any fault type exceeds the corresponding warning threshold, determining that the device is in a sub-healthy or faulty state, generating warning information containing the fault type, abnormal parameters and current health score, and storing the warning information in the local database of the edge monitoring unit.

[0040] In practice, a convolutional autoencoder is trained using historical normal operating data. By minimizing the reconstruction error, the encoder learns the latent representational patterns of normal data. The future predicted values ​​of each parameter are then used to... The pre-trained encoder is input, and the reconstructed data is obtained after encoding and decoding. The reconstruction error of each parameter is calculated using the symmetrical mean absolute percentage error. : ; Read the fault space correlation matrix The association weights between each fault type and the parameters are determined. The health index for each fault type is obtained by weighted summation of the reconstruction error with the corresponding parameter. : .

[0041] Specifically, a convolutional autoencoder is trained using historical normal operation data. By minimizing the reconstruction error, the encoder learns the latent representational patterns of normal data. Then, the future predicted values ​​of each parameter are input into the trained encoder, and after encoding and decoding, reconstructed data is obtained. The reconstruction error of each parameter is calculated based on the symmetric mean absolute percentage error to eliminate the influence of dimensions. Next, the fault space correlation matrix is ​​read, and the correlation weights between each fault type and parameter are weighted and summed with the corresponding parameter's reconstruction error to obtain a health indicator for each fault type. Finally, the three-sigma rule is used to set warning thresholds for each health indicator based on the mean and standard deviation of the reconstruction error of historical normal data. When any indicator exceeds the threshold, a warning message containing the fault type, abnormal parameters, and health score is generated and stored. This approach addresses the scarcity of fault data through unsupervised learning, enhances the influence of key parameters through weighted fusion, and improves the interpretability and accuracy of condition monitoring.

[0042] In some embodiments, step S40 specifically includes the following steps: extracting the encoder part from the trained convolutional autoencoder, inputting the current running parameters into the encoder to generate a low-dimensional latent representation; inputting the low-dimensional latent representation into a two-layer fully connected network, using a standard Gaussian distribution as the target distribution, and mapping the low-dimensional latent representation to a normalized representation conforming to a Gaussian distribution by minimizing the maximum mean difference; extracting running parameters for consecutive time periods from the real-time running data stream using a sliding window method, calculating various statistical features including mean, variance, and autocorrelation coefficient for the data within the window, and constructing an initial statistical feature set; using the Relief algorithm, guided by the parameter space correlation matrix and the fault space correlation matrix, selecting core statistical features sensitive to changes in operating conditions from the initial statistical feature set; The core statistical features are input into the embedding network, which is then mapped to embedded features of the same dimension as the normalized representation through a two-layer fully connected network. The embedding network is trained by minimizing the Euclidean distance to make the embedded features approximate the spatial distribution of the normalized representation. The normalized representation and the embedded features are L2 normalized and then averaged to obtain the composite features of the real-time samples that fuse model performance and data distribution. Using the historical normal data used by the convolutional autoencoder as a benchmark, the composite features of each historical sample are calculated, and the mean vector and covariance matrix are obtained. The Mahalanobis distance from the composite features of the current real-time sample to the mean vector is calculated, and the corresponding Mahalanobis distance is defined as the update discriminant index. Based on the Mahalanobis distance distribution of the composite features of historical samples, a warning threshold is set using the three Sigma rule, and the maximum Mahalanobis distance of historical samples is set as the threshold to be updated. According to the comparison results of the update discriminant index with the warning threshold and the threshold to be updated, the current data status is marked as normal, warning, or pending update. When the data status first enters the warning state, an adaptive buffer is initialized and begins storing subsequent continuous data. The buffer is divided into three sub-areas according to its capacity. When the first sub-area is full, the system determines whether to continue storing or clear the buffer based on the sample ratio of warning status to pending update status within the sub-area. When the second sub-area is full, the system determines whether to immediately trigger an update or wait for the third sub-area to fill based on the status distribution. If a health indicator is detected to exceed a threshold warning during the buffering process, the buffer is immediately cleared to avoid using faulty data to update the model. Finally, the system outputs the filtered continuous warning status data through the buffer.

[0043] Specifically, the encoder is first extracted from the trained convolutional autoencoder, mapping the current running parameters to a low-dimensional latent representation. This representation is then normalized to a Gaussian distribution using a two-layer fully connected network with the maximum mean difference as a constraint, eliminating the distribution divergence problem of the original latent representation and enhancing feature robustness. Simultaneously, multiple statistical features such as mean, variance, and autocorrelation coefficient are extracted from the real-time data stream using a sliding window approach. The Relief algorithm is then used to select core statistical features guided by the parameter space correlation matrix and the fault space correlation matrix. The selected features are then mapped to high-dimensional features of the same dimension as the normalized representation through an embedding network, and the spatial distribution of the two is aligned by minimizing the Euclidean distance. Finally, the normalized representation and the high-dimensional features are L2 normalized and their mean values ​​are taken to obtain a real-time sample composite feature that integrates model performance and data distribution. Based on historical normal data, composite features of each historical sample are calculated, and the mean vector and covariance matrix are obtained. The Mahalanobis distance from the composite features of real-time samples to the mean vector is calculated as an update discrimination index. Based on the Mahalanobis distance distribution of historical samples, a warning threshold is set using the three-sigma rule, and the maximum Mahalanobis distance is used as the update threshold. Data status is marked as normal, warning, or pending update according to the discrimination index. When data first enters the warning state, an adaptive buffer is activated, dividing the buffer into three sub-regions. The decision to continue storing, clearing, or triggering an update is dynamically made based on the ratio of warning to pending update samples in each sub-region. If a health indicator is detected to exceed the threshold, the buffer is immediately cleared. Thus, through composite feature fusion and adaptive buffering mechanisms, the degree of model drift is accurately identified and effective update data is selected, providing a reliable basis for subsequent online updates.

[0044] In practice, the encoder part is extracted from the trained convolutional autoencoder, and the current runtime parameters are input into the encoder to generate a low-dimensional latent representation z. Since the original latent representation serves the reconstruction task, its distribution is unconstrained, and directly using it to update the discriminant may not be stable enough; therefore, it needs to be normalized. z is input into a two-layer fully connected network, with a standard Gaussian distribution as the target distribution. By minimizing the Maximum Mean Discrepancy (MMD), z is mapped to a normalized representation conforming to a Gaussian distribution, denoted as z. The formula for calculating the MMD loss function is: ; in, represents the normalized latent representation of the i-th sample; b represents the training batch size; Indicates the standard Gaussian distribution The j-th sample randomly selected from the data; The calculation method is as follows: ; here, This represents the Gaussian kernel function, used to measure the similarity between two samples in a high-dimensional space; and All samples are random samples drawn from a Gaussian distribution. By minimizing... , making The distribution approximates the standard Gaussian distribution, thereby enhancing its robustness as a representation of model performance.

[0045] Furthermore, using the Relief algorithm, guided by the parameter space correlation matrix and the fault space correlation matrix, core statistical features sensitive to changes in operating conditions are selected from the initial set of statistical features, including: Based on the parameter space correlation matrix and the fault space correlation matrix, historical operating data of the same equipment at different times are selected to construct a cross-condition reference dataset, and operating data of different equipment at the same time are selected to construct a cross-equipment reference dataset. Both reference datasets contain normal operation data and performance degradation data, and are labeled with binary classification labels. For each data sample in the two reference datasets, multiple statistical features within a sliding window are extracted to form an initial statistical feature vector including mean, variance, and autocorrelation coefficient. The Relief algorithm is used to calculate the weight of each statistical feature in the two reference datasets.

[0046] The weights are calculated as follows: a sample is randomly sampled from the reference dataset, and the nearest neighbors of the same class and the opposite class of the target sample are found respectively. The Manhattan distance difference between the current statistical feature and the two nearest neighbors is calculated, and the weight value of the target feature in the corresponding reference dataset is obtained after iterative accumulation. After iterating through all samples, feature weight vectors for the cross-condition reference dataset and the cross-device reference dataset are obtained respectively. The mean of the two feature weight vectors is calculated, and statistical features with weights greater than the mean weight in each dataset are selected. The union of the two selection results is taken to obtain the core statistical feature set applicable to both cross-condition and cross-device scenarios. The core statistical feature set is solidified in the edge monitoring unit for the construction of composite features in subsequent real-time data streams, ensuring the stability and generalization ability of the model update discrimination method in time-varying environments and device migration scenarios.

[0047] Specifically, the system constructs cross-operating condition reference datasets and cross-device reference datasets based on the parameter space correlation matrix and the fault space correlation matrix, respectively. Each dataset contains normal operation data and performance degradation data, labeled with binary classification tags. Multiple statistical features within a sliding window are extracted from both datasets to form initial statistical feature vectors. The Relief algorithm is used to randomly sample and calculate the Manhattan distance difference between the target sample and its nearest neighbors (both similar and dissimilar), iteratively accumulating the differences to obtain the weight of each feature in its corresponding dataset. The mean of the feature weights for both datasets is calculated, and features with weights greater than the mean are selected. The union of the two selection results yields a core statistical feature set applicable to both cross-operating condition and cross-device scenarios, which is then embedded in the edge monitoring unit. This mechanism-guided feature selection ensures that the core statistical features possess stable discriminative capabilities in time-varying environments and equipment migration scenarios, providing a reliable basis for subsequent composite feature construction and model update discrimination.

[0048] In some embodiments, step S50 specifically includes the following steps: when it is determined that the current data status is to be updated and the buffer mechanism has stored continuous data that meets the preset conditions, the online update process is triggered, and a differentiated update mechanism is adopted for the integrated heterogeneous width network and the convolutional autoencoder respectively. For the integrated incremental update of the heterogeneous width network: the network parameters of the temporal feature extraction part of the original integrated heterogeneous width network remain unchanged, and key feature nodes are added to the enhancement feature part. The initial weights of the new nodes are randomly generated. The new data stored in the buffer is input into the original network to extract its temporal features and the original enhancement features. At the same time, supplementary features of the new data are extracted through the new nodes. The original enhancement features and the newly added supplementary features are concatenated to form the updated enhancement feature matrix. Based on the updated enhancement feature matrix and the original temporal features, combined with the true values ​​of the new data, the updated overall output weights are calculated through matrix pseudo-inverse operation. This allows the updated integrated heterogeneous width network to be compatible with the prediction accuracy of both the old and new data without retraining the original nodes. The number of new key feature nodes does not exceed the number of original key feature nodes. When the number of new nodes reaches the upper limit and still cannot meet the accuracy requirements, the new network is retrained and the original network is replaced. Adversarial regularization update for convolutional autoencoders: The target network to be updated is constructed using the trained convolutional autoencoder parameters as initial values. New data stored in a buffer is input into the target network, and the reconstruction error is calculated. Simultaneously, the new data is input into the original convolutional autoencoder to obtain the original latent representation, and then into the target network to obtain the updated latent representation. The cosine distance between the two is calculated as the encoding loss. The adversarial evolution between the reconstruction error and the encoding loss is used as the optimization objective, prompting the target network to learn the latent distribution characteristics of the new data while minimizing the reconstruction error. An L2 regularization term based on the original network parameters is added to the loss function, and a regularization direction guidance coefficient is introduced to adaptively adjust the strength of the regularization term. This regularization direction guidance coefficient is calculated based on the first-order difference of the reconstruction error of the fitted data from the old data, and is used to dynamically adjust the parameter constraint strength according to the degree of performance degradation of the old data during model updates. The loss function is minimized using the gradient descent algorithm to complete the fine-tuning of the convolutional autoencoder parameters. After the update is completed, the updated integrated heterogeneous width network and convolutional autoencoder are deployed on the edge detection unit, and the updated model parameters are synchronized to the baseline distribution calculation used to update the discrimination index to achieve model optimization.

[0049] Specifically, when the data status is determined to be pending update and the buffer stores continuous data that meets the conditions, the online update process is triggered. For integrated heterogeneous width networks, integrated incremental updates are adopted: the parameters of the time-series extraction remain unchanged, key nodes are added to the enhancement part and randomly initialized, and after new data is input, time-series features, original enhancement features, and newly added supplementary features are extracted, concatenated, and the updated output weights are quickly calculated through matrix pseudo-inverse operations. This allows the network to be compatible with both old and new data without retraining, and the number of added nodes does not exceed the original number of nodes. For convolutional autoencoders, adversarial regularization updates are adopted: using the original network parameters as initial values, the reconstruction error of the new data and the cosine distance between the old and new encodings are calculated as adversarial targets. Simultaneously, a regularization direction guide coefficient based on the first-order difference of the old data fitting error is introduced to dynamically adjust the strength of the L2 regularization term, and parameter fine-tuning is completed through gradient descent. The updated model is deployed on edge units and the discrimination benchmark is updated synchronously. Thus, through a differentiated incremental mechanism, low-resource-consumption continuous learning is achieved, ensuring that the model maintains high accuracy and stability across devices and operating conditions.

[0050] The implementation principle of a distributed precision inspection and early warning method supporting mobile web access in this application embodiment is as follows: The system constructs a parameter space correlation matrix and a fault space correlation matrix based on equipment mechanism and historical data, which are solidified in the edge units as prior knowledge to guide subsequent modeling and feature selection. Secondly, using real-time parameters as input, an integrated heterogeneous width network is used to perform empirical mode decomposition on strong fluctuation signals and one-dimensional convolution processing on slowly changing signals. After extracting multi-scale features, these are input into a recurrent width learning system, which outputs the future predicted values ​​of each parameter. Then, a convolutional autoencoder is used to calculate the reconstruction error of the predicted data, and the health indicators for each fault type are obtained by weighted fusion according to the fault space correlation matrix, achieving unsupervised health status assessment. Based on this, by standardizing the encoder's latent representation and fusing core statistical features to construct composite features, Mahalanobis distance is calculated as the update discrimination index. Data status is marked as normal, warning, or pending update according to a preset threshold, and valid data is stored through an adaptive buffering mechanism. When the model is determined to be in need of updating, an integrated incremental mechanism is used to add key nodes and update weights for the integrated heterogeneous width network. For the convolutional autoencoder, an adversarial regularization mechanism is used to fine-tune parameters using new data to reconstruct errors, encoding cosine distance, and adaptive regularization terms, completing the online update. Finally, the edge unit's built-in web server visualizes early warning information, predicted data, and model status through a mobile terminal browser, enabling remote mobile access. Thus, by combining edge computing with a lightweight model, system deployment costs and network bandwidth pressure are significantly reduced; through prior knowledge fusion and incremental update mechanisms, the model possesses cross-device and cross-operating-condition adaptive capabilities; mobile web access improves the speed and convenience of operation and maintenance response; and through the construction of health indicators through the fusion of mechanisms and data, the interpretability and credibility of fault diagnosis are enhanced, ultimately achieving intelligent and universal industrial equipment condition monitoring.

[0051] Figure 1 This is a flowchart illustrating a distributed precision inspection and early warning method supporting mobile web access in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows; unless explicitly stated otherwise, there is no strict order requirement for the execution of these steps, and they can be executed in other orders; and Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0052] Based on the same technical concept, referring to Figure 2 This application also provides a distributed precision inspection and early warning device that supports mobile web access, employing the following technical solution: the device includes: The real-time data module is used to deploy multiple edge monitoring units at the industrial dynamic equipment site to collect operating parameters in real time; it constructs parameter space correlation matrix and fault space correlation matrix based on equipment mechanism and historical data, and embeds them into the edge units as prior knowledge; The parameter prediction module is used to determine the input features of the prediction model based on the running parameters and the parameter space correlation matrix; it performs empirical mode decomposition on strong fluctuation signals and one-dimensional convolution processing on slow-changing signals, inputs the extracted features into the recurrent width learning system, constructs an integrated heterogeneous width network, and outputs the future predicted values ​​of each parameter. The parameter index module is used to train a convolutional autoencoder based on historical normal data, input future predicted values ​​into the trained encoder, calculate the reconstruction error of each parameter, and fuse them according to the fault space correlation matrix to obtain the health index of each fault type; when the health index exceeds the preset threshold, an early warning is generated and stored. The data discrimination module is used to normalize the latent representation of the convolutional autoencoder, extract statistical features from the real-time data sliding window and filter the core features, and fuse the two to construct composite features; calculate the Mahalanobis distance from the composite feature to the center of the training sample as the update discrimination index, mark the data status as normal, warning or waiting to be updated according to the preset threshold, and store the continuous data of the warning status through an adaptive buffer mechanism. The model update module is used to perform an integrated incremental mechanism on the integrated heterogeneous width network when the model's data status is determined to be in need of updating. This mechanism adds key nodes and updates the weights of the added nodes to ensure compatibility with both old and new data. For the convolutional autoencoder, an adversarial regularization mechanism is used to fine-tune the parameters using a loss function composed of the reconstruction error of the new data, the encoding cosine distance, and the adaptive regularization term, thus completing the online update. The real-time access module is used to embed a web server in the edge unit to visualize early warning information, prediction data and model status through a mobile terminal browser, providing real-time monitoring, historical tracking and alarm setting functions.

[0053] In some embodiments, the real-time data module is specifically used to extract core relevant parameters of typical faults based on the physical mechanism of the equipment and historical maintenance reports, and to construct a knowledge-level parameter correlation matrix and fault correlation matrix. Spearman's rank correlation coefficient is calculated using historical normal operation data to obtain the parameter correlation matrix at the data level. The parameter correlation matrix at the knowledge level is then fused with the parameter correlation matrix at the data level to obtain the parameter space correlation matrix. Each parameter type is then labeled according to the parameter change characteristics. By using limited fault data and employing a multidimensional dimensionality reduction method, a quantitative correlation matrix between parameters and faults is obtained. The knowledge-level fault correlation matrix is ​​then fused with the quantitative correlation matrix to obtain a fault space correlation matrix.

[0054] In some embodiments, the parameter prediction module is specifically used to determine the corresponding input feature set for each running parameter to be predicted based on the correlation between parameters recorded in the parameter space correlation matrix. The input feature set includes other running parameters that are strongly correlated with the corresponding parameter. The real-time acquired operating parameters are divided into two categories: strong fluctuation signals and slow change signals. Strong fluctuation signals include vibration signals, while slow change signals include temperature and pressure signals. Empirical mode decomposition is performed on strong fluctuation signals to extract several intrinsic mode components and their instantaneous frequencies as fluctuation features; Slowly changing signals are processed by a one-dimensional convolutional neural network. The convolution operation includes moving average filtering and second-order differential filtering to extract trend change features and abrupt change features. After concatenating the fluctuation features and trend change features, the concatenation is input into the cascaded structure of the loop width learning system. Temporal features are extracted through recursive propagation, and at the same time, randomly generated weights are used to map the fluctuation features and trend change features into enhanced features. By concatenating temporal features and enhanced features, and solving the output weights through pseudo-inverse operations, an integrated heterogeneous width network model is constructed. The integrated heterogeneous width network model is used to predict the key operating parameters and output the future predicted values ​​of each parameter.

[0055] In some embodiments, the parameter index module is specifically used to select historical normal operation data as training samples to train the convolutional autoencoder, and to enable the encoder to learn the potential representation rules of normal data by minimizing the reconstruction error. The future predicted values ​​of each parameter are input into the trained convolutional autoencoder. After encoding and decoding operations, the reconstructed data is obtained. The reconstruction error of each running parameter is calculated based on the symmetric mean absolute percentage error to eliminate the difference in dimensions and improve stability. Read the fault space correlation matrix, and sum the correlation weights of each fault type and each parameter in the matrix with the reconstruction error of the corresponding parameter to obtain the health index of each fault type. Using the Three Sigma principle, the mean and standard deviation of the reconstruction error of historical normal data are used as the basis to set the warning thresholds for health indicators of each fault type. When the health indicator of any fault type exceeds the corresponding warning threshold, the device is determined to be in a sub-healthy or faulty state, and a warning message containing the fault type, abnormal parameters and current health score is generated and stored in the local database of the edge monitoring unit.

[0056] In some embodiments, the data discrimination module is specifically used to extract the encoder part from the trained convolutional autoencoder, input the running parameters at the current time into the encoder, generate a low-dimensional latent representation, input the low-dimensional latent representation into a two-layer fully connected network, with a standard Gaussian distribution as the target distribution, and map the low-dimensional latent representation into a normalized representation that conforms to a Gaussian distribution by minimizing the maximum mean difference. Running parameters for consecutive time periods are extracted from the real-time running data stream using a sliding window approach. Various statistical features, including mean, variance, and autocorrelation coefficient, are calculated for the data within the window to form an initial set of statistical features. The Relief algorithm is used, guided by the parameter space correlation matrix and the fault space correlation matrix, to select core statistical features that are sensitive to changes in operating conditions from the initial set of statistical features; The core statistical features are input into the embedding network, which is then mapped to embedded features of the same dimension as the normalized representation through a two-layer fully connected network. The embedding network is trained by minimizing the Euclidean distance so that the embedded features approximate the spatial distribution of the normalized representation. After L2 normalization of the normalized representation and the embedded features, the mean value is taken to obtain the composite features of the real-time samples that fuse the model performance and data distribution. Using the historical normal data used by the convolutional autoencoder as a benchmark, the composite features of each historical sample are calculated and the mean vector and covariance matrix are obtained; the Mahalanobis distance from the composite features of the current real-time sample to the mean vector is calculated, and the corresponding Mahalanobis distance is defined as the update discriminant index. Based on the Mahalanobis distance distribution of the composite features of historical samples, the three Sigma rule is used to set the warning threshold, and the maximum Mahalanobis distance of the historical samples is set as the threshold to be updated. According to the comparison results of the update discrimination index with the warning threshold and the threshold to be updated, the current data status is marked as normal, warning, or pending update. When the data status first enters the warning state, an adaptive buffer is initialized and begins storing subsequent continuous data. The buffer is divided into three sub-areas according to its capacity. When the first sub-area is full, the system determines whether to continue storing or clear the buffer based on the sample ratio of warning status to pending update status within the sub-area. When the second sub-area is full, the system determines whether to immediately trigger an update or wait for the third sub-area to fill based on the status distribution. If a health indicator is detected to exceed a threshold warning during the buffering process, the buffer is immediately cleared to avoid using faulty data to update the model. Finally, the system outputs the filtered continuous warning status data through the buffer.

[0057] In some embodiments, the data discrimination module is specifically used to construct a cross-condition reference dataset by selecting historical operating data of the same device at different times based on the parameter space correlation matrix and the fault space correlation matrix, and to construct a cross-device reference dataset by selecting operating data of different devices at the same time. Both reference datasets contain normal operating data and performance degradation data, and are labeled with binary classification labels. For the two reference datasets, various statistical features within a sliding window are extracted for each data sample to form an initial statistical feature vector including mean, variance, and autocorrelation coefficient. The Relief algorithm is used to calculate the weight of each statistical feature in the two reference datasets. The weight is calculated as follows: a sample is randomly sampled from the reference dataset, and the nearest neighbors of the same class and the opposite class of the target sample are found respectively. The Manhattan distance difference between the current statistical feature and the two nearest neighbors is calculated. After iterative accumulation, the weight value of the target feature in the corresponding reference dataset is obtained. After iterating through all samples, the feature weight vectors of the cross-condition reference dataset and the cross-device reference dataset are obtained respectively. Calculate the mean of the two feature weight vectors respectively, filter out the statistical features in each dataset whose weight is greater than the mean weight, and take the union of the two sets of filtering results to obtain the core statistical feature set that is applicable to both cross-working-condition scenarios and cross-device scenarios. The core statistical feature set is embedded in the edge monitoring unit and used to construct composite features in subsequent real-time data streams, ensuring the stability and generalization ability of the model update discrimination method in time-varying environments and device migration scenarios.

[0058] In some embodiments, the model update module is specifically used to trigger an online update process when it is determined that the current data state is to be updated and the buffer mechanism has stored continuous data that meets the preset conditions, and to adopt a differentiated update mechanism for the integrated heterogeneous width network and the convolutional autoencoder respectively. For the integrated incremental update of the heterogeneous width network: the network parameters of the temporal feature extraction part of the original integrated heterogeneous width network remain unchanged, and key feature nodes are added to the enhancement feature part. The initial weights of the new nodes are randomly generated. The new data stored in the buffer is input into the original network to extract its temporal features and the original enhancement features. At the same time, supplementary features of the new data are extracted through the new nodes. The original enhancement features and the newly added supplementary features are concatenated to form the updated enhancement feature matrix. Based on the updated enhancement feature matrix and the original temporal features, combined with the true values ​​of the new data, the updated overall output weights are calculated through matrix pseudo-inverse operation. This allows the updated integrated heterogeneous width network to be compatible with the prediction accuracy of both the old and new data without retraining the original nodes. The number of new key feature nodes does not exceed the number of original key feature nodes. When the number of new nodes reaches the upper limit and still cannot meet the accuracy requirements, the new network is retrained and the original network is replaced. Adversarial regularization update for convolutional autoencoders: The target network to be updated is constructed using the trained convolutional autoencoder parameters as initial values. New data stored in a buffer is input into the target network, and the reconstruction error is calculated. Simultaneously, the new data is input into the original convolutional autoencoder to obtain the original latent representation, and then into the target network to obtain the updated latent representation. The cosine distance between the two is calculated as the encoding loss. The adversarial evolution between the reconstruction error and the encoding loss is used as the optimization objective, prompting the target network to learn the latent distribution characteristics of the new data while minimizing the reconstruction error. An L2 regularization term based on the original network parameters is added to the loss function, and a regularization direction guidance coefficient is introduced to adaptively adjust the strength of the regularization term. This regularization direction guidance coefficient is calculated based on the first-order difference of the reconstruction error of the fitted data from the old data, and is used to dynamically adjust the parameter constraint strength according to the degree of performance degradation of the old data during model updates. The loss function is minimized using the gradient descent algorithm to complete the fine-tuning of the convolutional autoencoder parameters. After the update is completed, the updated integrated heterogeneous width network and convolutional autoencoder are deployed on the edge detection unit, and the updated model parameters are synchronized to the baseline distribution calculation used to update the discrimination index to achieve model optimization.

[0059] This application also discloses a control device.

[0060] Specifically, the control device includes a memory and a processor, the memory storing a computer program that can be loaded by the processor and executed to support the aforementioned distributed precision inspection and early warning method for mobile web access.

[0061] This application also discloses a computer-readable storage medium.

[0062] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed, such as the distributed precision inspection and early warning method supporting mobile web access described above. The computer-readable storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0063] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A distributed precision inspection and early warning method supporting mobile web access, characterized in that, include: Multiple edge monitoring units are deployed at industrial equipment sites to collect operating parameters in real time. Based on equipment mechanisms and historical data, parameter space correlation matrices and fault space correlation matrices are constructed and solidified as prior knowledge in edge units; Based on the operating parameters, the input features of the prediction model are determined according to the parameter space correlation matrix; Operating parameters are divided into two categories: strong fluctuation signals and slow change signals. Strong fluctuation signals include vibration signals, while slow change signals include temperature and pressure signals. Empirical mode decomposition is performed on strong fluctuation signals, and one-dimensional convolution is performed on slowly changing signals. The extracted features are input into the cascaded structure of the recurrent width learning system. Temporal features are extracted through recursive propagation. At the same time, randomly generated weights are used to map the extracted features into enhanced features. The temporal features and the enhanced features are concatenated and the weights are solved through pseudo-inverse operation. An integrated heterogeneous width network is constructed to output the future predicted values ​​of each parameter. A convolutional autoencoder is trained based on historical normal data. The future predicted value is input into the trained encoder, the reconstruction error of each parameter is calculated, and the values ​​are weighted and fused according to the fault space correlation matrix to obtain the health indicators of each fault type. When the health indicators exceed the preset threshold, an early warning is generated and stored. Using the standard Gaussian distribution as the target distribution, the latent representation of the convolutional autoencoder is mapped to a normalized representation that conforms to the Gaussian distribution by minimizing the maximum mean difference. At the same time, statistical features are extracted from the real-time data sliding window and core features are selected. The two are then fused to construct composite features. The Mahalanobis distance from the composite feature to the center of the training sample is calculated as the update discrimination index. The data status is marked as normal, warning or pending update according to the preset threshold. The continuous data of the warning status is stored through an adaptive buffer mechanism. The continuous data of the warning status is stored through a buffer that is divided into three sub-regions according to the capacity. The continuous data of the warning status is stored, and the continuous data of the warning status is stored, cleared or triggered for update is dynamically decided according to the sample ratio of the warning status and the pending update status in each sub-region. When a health indicator exceeds the threshold warning, the buffer is cleared. When the model's data state is determined to be in need of updating, an integrated incremental mechanism is adopted for the integrated heterogeneous width network. By adding key nodes and updating the weights of the new nodes, the new and old data are compatible. An adversarial regularization mechanism is adopted for the convolutional autoencoder. The adversarial evolution between the reconstruction error of the new data and the cosine distance between the encoding of the new and old potential representations is used as the optimization objective, so that the target network learns the potential distribution characteristics of the new data while minimizing the reconstruction error. The parameters are fine-tuned by combining the loss function composed of the L2 regularization term based on the original network parameters. The L2 regularization term is adaptively adjusted by the regularization direction guidance coefficient calculated based on the first difference of the reconstruction error of the old data, thus completing the online update. The edge unit has a built-in web server that visualizes early warning information, prediction data, and model status through a mobile terminal browser, providing real-time monitoring, historical tracking, and alarm setting functions.

2. The distributed precision inspection and early warning method supporting mobile Web access according to claim 1, characterized in that, The construction of parameter space correlation matrices and fault space correlation matrices based on equipment mechanisms and historical data, and their solidification as prior knowledge in edge units, includes: Based on the physical mechanism of the equipment and historical maintenance reports, the core relevant parameters of typical faults are extracted, and a parameter correlation matrix and a fault correlation matrix at the knowledge level are constructed. Spearman's rank correlation coefficient is calculated using historical normal operation data to obtain the parameter correlation matrix at the data level. The parameter correlation matrix at the knowledge level is then fused with the parameter correlation matrix at the data level to obtain the parameter space correlation matrix. Each parameter type is then labeled according to the parameter change characteristics. By using limited fault data and employing a multidimensional dimensionality reduction method, a quantitative correlation matrix between parameters and faults is obtained. The knowledge-level fault correlation matrix is ​​then fused with the quantitative correlation matrix to obtain a fault space correlation matrix.

3. The distributed precision inspection and early warning method supporting mobile Web access according to claim 1, characterized in that, The input features of the prediction model are determined based on the operating parameters and the parameter space correlation matrix. Empirical mode decomposition is performed on strong fluctuation signals, and one-dimensional convolution processing is applied to slowly changing signals. The extracted features are input into a recurrent width learning system to construct an integrated heterogeneous width network, which outputs future predicted values ​​for each parameter, including: Based on the correlation between parameters recorded in the parameter space correlation matrix, the corresponding input feature set is determined for each running parameter to be predicted. The input feature set includes other running parameters that are strongly correlated with the corresponding parameter. Empirical mode decomposition is performed on strong fluctuation signals to extract several intrinsic mode components and their instantaneous frequencies as fluctuation features; Slowly changing signals are processed by a one-dimensional convolutional neural network. The convolution operation includes moving average filtering and second-order differential filtering to extract trend change features and abrupt change features. After concatenating the fluctuation feature and the trend change feature, the concatenation structure of the loop width learning system is input into the cascade structure of the loop width learning system. Temporal features are extracted through recursive propagation, and at the same time, the fluctuation feature and the trend change feature are mapped into enhanced features using randomly generated weights. The temporal features and the enhanced features are concatenated, and the output weights are solved by pseudo-inverse operation to construct an integrated heterogeneous width network model. The integrated heterogeneous width network model is used to predict the key operating parameters and output the future predicted values ​​of each parameter.

4. The distributed precision inspection and early warning method supporting mobile Web access according to claim 3, characterized in that, The convolutional autoencoder trained based on historical normal data is input into the trained encoder with the future predicted value, the reconstruction error of each parameter is calculated, and the weighted fusion is performed according to the fault space correlation matrix to obtain the health index of each fault type. When health indicators exceed preset thresholds, an alert is generated and stored, including: Historical normal operation data is selected as training samples to train the convolutional autoencoder. By minimizing the reconstruction error, the encoder learns the potential representational patterns of normal data. The predicted future values ​​of each parameter are input into the trained convolutional autoencoder. After encoding and decoding operations, the reconstructed data is obtained. The reconstruction error of each running parameter is calculated based on the symmetric mean absolute percentage error to eliminate the difference in dimensions and improve stability. Read the fault space correlation matrix, and sum the correlation weights of each fault type and each parameter in the matrix with the reconstruction error of the corresponding parameter to obtain the health index of each fault type. Using the Three Sigma principle, the mean and standard deviation of the reconstruction error of historical normal data are used as the basis to set the warning thresholds for health indicators of each fault type. When the health indicator of any fault type exceeds the corresponding warning threshold, the device is determined to be in a sub-healthy or faulty state, and a warning message containing the fault type, abnormal parameters and current health score is generated and stored in the local database of the edge monitoring unit.

5. A distributed precision inspection and early warning method supporting mobile Web access according to claim 1, characterized in that, The process involves normalizing the latent representation of the convolutional autoencoder, extracting statistical features from the real-time data sliding window and selecting core features, then fusing the two to construct a composite feature. The Mahalanobis distance from the composite feature to the training sample center is calculated as an update criterion. Data states are labeled as normal, warning, or pending update based on a preset threshold, and continuous warning data is stored using an adaptive buffering mechanism. The encoder part is extracted from the trained convolutional autoencoder, and the running parameters at the current time are input into the encoder to generate a low-dimensional latent representation. The low-dimensional latent representation is input into a two-layer fully connected network with a standard Gaussian distribution as the target distribution. The low-dimensional latent representation is mapped to a normalized representation that conforms to a Gaussian distribution by minimizing the maximum mean difference. Running parameters for consecutive time periods are extracted from the real-time running data stream using a sliding window method. Various statistical features, including mean, variance, and autocorrelation coefficient, are calculated for the data within the window to form an initial set of statistical features. Using the Relief algorithm, guided by the parameter space correlation matrix and the fault space correlation matrix, core statistical features sensitive to changes in operating conditions are selected from the initial set of statistical features. The core statistical features are input into the embedding network, and mapped to embedded features of the same dimension as the normalized representation through a two-layer fully connected network. The embedding network is trained by minimizing the Euclidean distance so that the embedded features approximate the spatial distribution of the normalized representation. The normalized representation and the embedded features are respectively L2 normalized and then averaged to obtain the composite features of real-time samples that fuse model performance and data distribution. Using the historical normal data used by the convolutional autoencoder as a benchmark, the composite features of each historical sample are calculated and the mean vector and covariance matrix are obtained; the Mahalanobis distance from the composite features of the current real-time sample to the mean vector is calculated, and the corresponding Mahalanobis distance is defined as the update discriminant index. Based on the Mahalanobis distance distribution of the composite features of historical samples, the three Sigma rule is used to set the warning threshold, and the maximum Mahalanobis distance of the historical samples is set as the threshold to be updated; according to the comparison results of the update discrimination index with the warning threshold and the threshold to be updated, the current data status is marked as normal, warning, or pending update. When the data status first enters the warning state, an adaptive buffer is initialized and begins storing subsequent continuous data. The buffer is divided into three sub-areas according to its capacity. When the first sub-area is full, the system determines whether to continue storing or clear the buffer based on the sample ratio of warning status to pending update status within the sub-area. When the second sub-area is full, the system determines whether to immediately trigger an update or wait for the third sub-area to fill based on the status distribution. If a health indicator is detected to exceed a threshold warning during the buffering process, the buffer is immediately cleared to avoid using faulty data to update the model. Finally, the filtered continuous warning status data is output through the buffer.

6. A distributed precision inspection and early warning method supporting mobile Web access according to claim 5, characterized in that, The Relief algorithm, guided by the parameter space correlation matrix and the fault space correlation matrix, is used to screen out core statistical features sensitive to changes in operating conditions from the initial statistical feature set, including: Based on the parameter space correlation matrix and the fault space correlation matrix, historical operating data of the same equipment at different times are selected to construct a cross-condition reference dataset, and operating data of different equipment at the same time are selected to construct a cross-equipment reference dataset. Both reference datasets contain normal operation data and performance degradation data, and are labeled with binary classification labels. For the two reference datasets, various statistical features within a sliding window are extracted for each data sample to form an initial statistical feature vector including mean, variance, and autocorrelation coefficient. The Relief algorithm is used to calculate the weight of each statistical feature in the two reference datasets. The weight is calculated as follows: a sample is randomly sampled from the reference dataset, and the nearest neighbors of the same class and the opposite class of the target sample are found respectively. The Manhattan distance difference between the current statistical feature and the two nearest neighbors is calculated. After iterative accumulation, the weight value of the target feature in the corresponding reference dataset is obtained. After iterating through all samples, the feature weight vectors of the cross-condition reference dataset and the cross-device reference dataset are obtained respectively. Calculate the mean of the two feature weight vectors respectively, filter out the statistical features in each dataset whose weight is greater than the mean weight, and take the union of the two sets of filtering results to obtain the core statistical feature set that is applicable to both cross-working-condition scenarios and cross-device scenarios. The core statistical feature set is embedded in the edge monitoring unit for the construction of composite features in subsequent real-time data streams, ensuring the stability and generalization ability of the model update discrimination method in time-varying environments and device migration scenarios.

7. A distributed precision inspection and early warning method supporting mobile Web access according to claim 1, characterized in that, When the model's data state is determined to be pending update, an integrated incremental mechanism is adopted for the integrated heterogeneous width network. This involves adding key nodes and updating the weights of the added nodes to ensure compatibility with both old and new data. An adversarial regularization mechanism is employed for the convolutional autoencoder, using a loss function composed of the new data reconstruction error, encoding cosine distance, and adaptive regularization term to fine-tune the parameters and complete the online update. This includes: When the current data status is determined to be pending update and the buffer mechanism has stored continuous data that meets the preset conditions, the online update process is triggered, and a differentiated update mechanism is adopted for the integrated heterogeneous width network and the convolutional autoencoder respectively. For the integrated incremental update of the heterogeneous width network: the network parameters of the temporal feature extraction part of the original integrated heterogeneous width network remain unchanged, and key feature nodes are added to the enhancement feature part. The initial weights of the new nodes are randomly generated. The new data stored in the buffer is input into the original network to extract its temporal features and the original enhancement features. At the same time, supplementary features of the new data are extracted through the new nodes. The original enhancement features and the new supplementary features are concatenated to form the updated enhancement feature matrix. Based on the updated enhancement feature matrix and the original temporal features, combined with the true value of the new data, the updated overall output weight is calculated through matrix pseudo-inverse operation, so that the updated integrated heterogeneous width network can be compatible with the prediction accuracy of the new and old data without retraining the original nodes. The number of new key feature nodes does not exceed the number of original key feature nodes. When the number of new nodes reaches the upper limit and still cannot meet the accuracy requirements, the new network is retrained and the original network is replaced. Adversarial regularization update for convolutional autoencoders: The target network to be updated is constructed using the trained convolutional autoencoder parameters as initial values. New data stored in a buffer is input into the target network, and the reconstruction error is calculated. Simultaneously, the new data is input into the original convolutional autoencoder to obtain the original latent representation, and then into the target network to obtain the updated latent representation. The cosine distance between the two is calculated as the encoding loss. The adversarial evolution between the reconstruction error and the encoding loss is used as the optimization objective, prompting the target network to learn the latent distribution characteristics of the new data while minimizing the reconstruction error. An L2 regularization term based on the original network parameters is added to the loss function, and a regularization direction guidance coefficient is introduced to adaptively adjust the strength of the regularization term. This regularization direction guidance coefficient is calculated based on the first-order difference of the reconstruction error of the fitted data from the old data, and is used to dynamically adjust the parameter constraint strength according to the degree of performance degradation of the old data during model updates. The loss function is minimized using the gradient descent algorithm to complete the fine-tuning of the convolutional autoencoder parameters. After the update is completed, the updated integrated heterogeneous width network and convolutional autoencoder are deployed on the edge detection unit, and the updated model parameters are synchronized to the baseline distribution calculation used to update the discrimination index to achieve model optimization.

8. A distributed precision inspection and early warning device supporting mobile web access, characterized in that, The device includes: The real-time data module is used to deploy multiple edge monitoring units at the industrial dynamic equipment site to collect operating parameters in real time; it constructs parameter space correlation matrix and fault space correlation matrix based on equipment mechanism and historical data, and embeds them into the edge units as prior knowledge; The parameter prediction module is used to determine the input features of the prediction model based on the operating parameters and the parameter space correlation matrix. The operating parameters are divided into two categories: strong fluctuation signals and slowly changing signals. Strong fluctuation signals include vibration signals, and slowly changing signals include temperature and pressure signals. Empirical mode decomposition is performed on strong fluctuation signals, and one-dimensional convolution processing is performed on slowly changing signals. The extracted features are input into the cascaded structure of the recurrent width learning system. Temporal features are extracted through recursive propagation. At the same time, randomly generated weights are used to map the extracted features into enhanced features. The temporal features and the enhanced features are concatenated and the weights are solved through pseudo-inverse operation to construct an integrated heterogeneous width network and output the future predicted values ​​of each parameter. The parameter index module is used to train a convolutional autoencoder based on historical normal data, input the future predicted value into the trained encoder, calculate the reconstruction error of each parameter, and fuse them according to the fault space correlation matrix to obtain the health index of each fault type; when the health index exceeds the preset threshold, an early warning is generated and stored. The data discrimination module is used to map the latent representation of the convolutional autoencoder to a normalized representation conforming to the Gaussian distribution by minimizing the maximum mean difference, with the standard Gaussian distribution as the target distribution. At the same time, it extracts statistical features from the sliding window of real-time data and selects core features, and merges the two to construct composite features. The Mahalanobis distance from the composite feature to the center of the training samples is calculated as the update discrimination index. The data status is marked as normal, warning, or pending update according to the preset threshold. The continuous data of the warning status is stored through an adaptive buffer mechanism. The continuous data of the warning status is stored in a buffer divided into three sub-regions according to the capacity. The continuous data of the warning status is stored, and the continuous decision to continue storing, clearing, or triggering the update is dynamically decided according to the sample ratio of the warning status to the pending update status in each sub-region. When a health indicator exceeds the threshold warning, the buffer is cleared. The model update module is used to address the data state of the model when it is determined to be in need of updating. For the integrated heterogeneous width network, an integrated incremental mechanism is employed, adding key nodes and updating their weights to ensure compatibility with both old and new data. For the convolutional autoencoder, an adversarial regularization mechanism is used, with the adversarial evolution between the reconstruction error of the new data and the cosine distance between the encodings of the new and old latent representations as the optimization objective. This prompts the target network to learn the latent distribution characteristics of the new data while minimizing the reconstruction error. Furthermore, the module combines a loss function based on the L2 regularization term of the original network parameters for parameter fine-tuning. The L2 regularization term adaptively adjusts its strength using the regularization direction guidance coefficient calculated based on the first-order difference of the reconstruction error of the old data, thus completing the online update. The real-time access module is used to embed a web server in the edge unit to visualize early warning information, prediction data and model status through a mobile terminal browser, providing real-time monitoring, historical tracking and alarm setting functions.

9. A control device, characterized in that, The device includes: A memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 7.

Images