Multi-dimensional monitoring method for real-time operation state of intelligent fusion terminal

Abstract

The invention relates to the field of electric digital data processing, in particular to a multi-dimensional monitoring method for the real-time running state of an intelligent fusion terminal. Collecting real-time operation state data of the intelligent fusion terminal; calculating a fusion feature vector of each sample; calculating a feature enhancement vector of each sample; calculating a prediction state probability vector of each sample; calculating a composite loss function; after a sample is randomly sampled from the training data set repeatedly, the steps S2, S3, S4 and S5 are executed in sequence until a final target is achieved, and a trained state recognition model is obtained; and acquiring new real-time operation state data, inputting the new real-time operation state data into the trained state recognition model, outputting a prediction state probability vector of the new real-time operation state data, and taking the operation state with the highest probability value as the real-time operation state of the intelligent fusion terminal. An existing intelligent fusion terminal real-time operation state multi-dimensional monitoring method has the problem that the state recognition accuracy is low. The method provided by the invention is relatively high in state recognition accuracy.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing, and specifically to a method for multi-dimensional monitoring of the real-time operating status of an intelligent fusion terminal. Background Technology

[0002] A smart converged terminal is a highly integrated intelligent device with edge computing capabilities. Internally, it typically includes monitoring agents, sensor interfaces, and system performance counters, enabling simultaneous data acquisition, communication, computation, and control. With the further development of the Internet of Things (IoT) and edge computing, smart converged terminals are already being used in the power industry and other industrial sectors. For example, in the power industry, smart converged terminals in distribution transformer areas can collect comprehensive data on electricity, water, gas, and heat consumption, accurately monitor the load operation status and power quality of the distribution area in real time, quickly identify and locate faults, and significantly improve the efficiency of power grid operation and maintenance.

[0003] The stability and reliability of intelligent converged terminals directly affect the security and efficiency of the entire system. Therefore, it is necessary to monitor the real-time operating status of intelligent converged terminals. Existing multi-dimensional monitoring methods for the real-time operating status of intelligent converged terminals mainly analyze the collected real-time operating status data by constructing a status recognition model. However, due to the high complexity of the internal state of intelligent converged terminals and the dynamic changes of the external environment, the operating status data of intelligent converged terminals exhibits characteristics of being multi-source, heterogeneous, high-dimensional, and closely time-series correlated. This leads to the following problems with existing methods: (1) The real-time operating status data generated by the intelligent fusion terminal has the characteristics of being multi-source, heterogeneous, high-dimensional, time-series fluctuation and local anomaly. Conventional standardization methods are difficult to effectively deal with the scale differences and correlations between different data dimensions, as well as the nonlinear changes in time series, which can easily lead to the smoothing of local fluctuation features or the masking of abnormal information. (2) Faced with noise interference and feature aliasing in the fused high-dimensional data, conventional feature enhancement methods are difficult to effectively capture the nonlinear relationships and local structural features inside the data, which can easily lead to insufficient feature representation of minor anomalies or state changes. (3) The data contains time-series dependencies and complex nonlinear boundaries between states. Conventional classification networks are difficult to effectively model the dynamic characteristics of state transitions and the discrimination features of different state categories, and are prone to misjudgment at the state switching boundary. (4) In order to train the running state recognition network and ensure its recognition accuracy, temporal smoothness of state prediction and representation ability of learned features, conventional cross-entropy loss is difficult to fully constrain the behavior of the state recognition model. Common training objectives usually only focus on classification accuracy, ignore the temporal smoothness constraint that state prediction should have, and lack guidance on the physical meaning of features, which leads to unstable and difficult-to-understand decision-making of the state recognition model. Therefore, existing multi-dimensional monitoring methods for the real-time operating status of intelligent fusion terminals suffer from low accuracy in status recognition. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method for real-time multi-dimensional monitoring of the operating status of an intelligent fusion terminal with high accuracy in status recognition.

[0005] To address the aforementioned technical problems, this invention provides a method for multi-dimensional monitoring of the real-time operating status of an intelligent fusion terminal, comprising: S1. Collect multiple real-time operating status data with timestamps from the intelligent fusion terminal at fixed time intervals. Each real-time operating status data with timestamps is a sample. Each sample includes multiple dimensions of data from multiple data sources. Label the operating status of each sample and use all labeled samples as training dataset. S2. Based on the data of each sample in each dimension, obtain the normalized data vector of each sample, and calculate the fusion feature vector of each sample based on the normalized data vector of each sample. S3. Calculate the feature enhancement vector of each sample based on the fused feature vector of each sample, the adaptive weight vector of each sample, and the gradient enhancement vector of each sample; S4. Calculate the multi-branch fusion feature vector of each sample based on the feature enhancement vector of each sample, the temporal context vector of each sample, and the local statistical feature vector of each sample; calculate the predicted state probability vector of each sample based on the temporal context-aware feature vector of each sample. S5. Calculate the composite loss function based on state recognition loss, temporal consistency loss, and feature reconstruction loss; S6. Repeatedly sample a sample randomly from the training dataset and execute S2, S3, S4, and S5 in sequence until the final goal is achieved, and the trained state recognition model is obtained. S7. Collect new real-time operating status data, input the new real-time operating status data into the trained state recognition model, output the predicted state probability vector of the new real-time operating status data, and take the operating status with the highest probability value as the real-time operating status of the intelligent fusion terminal.

[0006] As a further improvement of the present invention: step S2 includes: S201. Perform independent normalization on the original observation values ​​of each sample in each dimension to obtain the normalized values ​​of each sample in each dimension, and calculate the normalized data vector of each sample based on the normalized values ​​of each sample in each dimension. S202. Calculate the fusion feature vector of each sample based on the normalized data vector of each sample.

[0007] Preferably, the formula for calculating the fusion feature vector of each sample in S202 is as follows: , In the formula, For the first The fusion feature vector of each sample; Index the data source; The total number of data sources; For the first Adaptive weighting coefficients for each data source; For the first A feature transformation function specific to a data source; For the first Among the samples belonging to the th , Feature vectors of each data source

[0008] As a further improvement of the present invention: step S3 includes: S301. Calculate the local sparsity estimate of each sample based on the fused feature vector of each sample; S302. Calculate the gradient magnitude of the feature gradient vector based on the fused feature vector of each sample; S303. Calculate the adaptive weight vector of each sample based on the gradient magnitude of the feature gradient vector of each sample; S304. Calculate the gradient enhancement vector of each sample based on the feature gradient vector of each sample and the fused feature vector of each sample; S305. Calculate the feature enhancement vector of each sample based on the fused feature vector of each sample, the adaptive weight vector of each sample, and the gradient enhancement vector of each sample.

[0009] Preferably, the formula for calculating the feature enhancement vector of each sample in S305 is as follows: , In the formula, For the first Feature enhancement vectors for each sample; For the first The fusion feature vector of each sample; This is element-wise multiplication; For the first An adaptive weight vector for each sample; For the first Gradient enhancement vector for each sample.

[0010] As a further improvement of the present invention: step S4 includes: S401. Calculate the temporal context vector of each sample based on the collection timestamp of each sample; S402. Calculate the local statistical feature vector of each sample based on the mean of each dimension within the window, the variance of each dimension within the window, and the skewness of each dimension within the window. S403. Calculate the multi-branch fusion feature vector of each sample based on the feature enhancement vector of each sample, the temporal context vector of each sample, and the local statistical feature vector of each sample; S404. Calculate the predicted state probability vector of each sample based on the temporal context-aware feature vector of each sample.

[0011] Preferably, the formula for calculating the multi-branch fusion feature vector of each sample in S403 is as follows: , In the formula, For the first Multi-branch fusion feature vector of each sample; For corresponding The adaptive gating vector; This is element-wise multiplication; The feature mapping function for the feature enhancement vector; For corresponding The adaptive gating vector; The feature mapping function for the temporal context vector; For the first The temporal context vector of each sample; For corresponding The adaptive gating vector; It is the feature mapping function of the local statistical feature vector; For the first Local statistical feature vectors of a sample.

[0012] As a further improvement of the present invention: step S5 includes: S501. Calculate the state recognition loss used to optimize the network's accuracy in classifying operating states; S502. Calculate the timing consistency loss; S503. Calculate the feature reconstruction loss; S504. Calculate the composite loss function based on state recognition loss, temporal consistency loss, and feature reconstruction loss.

[0013] Preferably, the formula for calculating the composite loss function is: , In the formula, It is a composite loss function; State recognition loss; This is the weighting coefficient for the time-series consistency loss; This results in a loss of time-series consistency. The weighting coefficients for the feature reconstruction loss; The loss is for feature reconstruction.

[0014] The beneficial effects of the present invention are as follows: The multi-dimensional monitoring method for real-time operation status of an intelligent fusion terminal provided by the present invention has high accuracy in status identification.

[0015] First, S2 proposes an adaptive weighted multi-source data fusion method. By dynamically calculating the weights of each data source and combining source-specific feature transformations, it overcomes the limitations of traditional methods that simply splice data of different scales and types. S202 performs interval normalization on multi-dimensional data and constructs an adaptive weighted fusion function to eliminate dimensional differences and fuse multi-source heterogeneous data.

[0016] Secondly, based on the fused feature data, S3 uses local density estimation and spatiotemporal gradient information to construct an adaptive weight feature enhancement mechanism. This allows it to dynamically adjust the saliency of features according to the sparsity and drastic changes in feature distribution, thereby smoothing normal steady-state features while enhancing the expression of abnormal and abrupt features, thus improving the state recognition model's ability to perceive subtle state changes.

[0017] Furthermore, S4 constructs a multi-branch fusion graph attention recurrent network based on feature enhancement vectors, combined with temporal context information and local statistical features, to simultaneously capture long-term dependencies, local mutation features, and structural relationships between samples in the state sequence, thereby achieving high-precision operation state recognition. Finally, S5 integrates state recognition loss, temporal consistency loss and feature reconstruction loss to construct a composite loss function. Through joint optimization using a multi-task learning strategy, the state recognition model can accurately classify while its internal feature representation conforms to the physical laws of state changes in intelligent fusion terminals. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the principle of the present invention; Figure 2 This is an overall architecture diagram of a multi-dimensional monitoring system for the real-time operating status of intelligent fusion terminals. Figure 3 Flowchart for data collection and training dataset construction for smart terminals; Figure 4 Flowchart for the construction of the multidimensional data normalization and fusion module. Detailed Implementation

[0019] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0020] like Figure 1 , Figure 2 , Figure 3 , Figure 4 As shown, the multi-dimensional monitoring method for real-time operation status of an intelligent fusion terminal provided by the present invention includes: S1. In order to effectively monitor the real-time operating status of the intelligent fusion terminal, multiple real-time operating status data with timestamps are collected from the intelligent fusion terminal at fixed time intervals. Each real-time operating status data with timestamps is a sample. Each sample includes multiple dimensions of data from multiple data sources. The operating status of each sample is labeled based on the actual performance of the intelligent fusion terminal at the time corresponding to each sample, domain knowledge, and preset threshold rules. If necessary, it can be manually reviewed by those skilled in the art. All labeled samples are used as the training dataset for the status recognition model. The data collection process covers multiple levels of terminal operation. By integrating the terminal's built-in monitoring agent, sensor interface, and system performance counter, it continuously collects real-time indicators from different data sources. Data from each data source is synchronously collected and timestamped at fixed time intervals (e.g., once per second) to form the original multi-dimensional time series (i.e., real-time operating status data). Data sources include: computing resource data sources, network performance data sources, sensor data sources, system status data sources, and application performance data sources. The dimensions of computing resource data sources include: CPU utilization (unit: %, percentage), memory usage (unit: MB or GB), and disk read / write speed (unit: MB / s). The dimensions of network performance data sources include: network latency (unit: ms), packet loss rate (unit: %), and throughput (unit: Mbps). The dimensions of sensor data sources include: device temperature (unit: °C, degrees Celsius), battery power (unit: %), and screen brightness (unit: cd / m²). The dimensions of the system status data source include: the number of currently running processes (unit: number) and system load (unit: dimensionless, representing the average number of processes). Typical dimensional attributes of application performance data sources include: application response time (unit: ms) and frame rate (unit: FPS).

[0021] The original observations of each sample across multiple dimensions are directly collected by various sensors or monitoring systems of the intelligent fusion terminal. The physical meaning is the central processing unit utilization, network latency, or device temperature mentioned above. The operating status is predefined based on the combined characteristics of data from various dimensions and historical operating experience. The operating status completely corresponds to the target of the subsequent status identification module. The operating status includes: "normal", "low load", "high load", "overload", "abnormal", "fault", "startup", "shutdown" and "transitional state".

[0022] S2. Based on the data of each sample in each dimension, obtain the normalized data vector of each sample, calculate the fusion feature vector of each sample according to the normalized data vector of each sample, and complete the construction of the multidimensional data normalization and fusion module. S201. Perform independent normalization on the original observations of each sample in each dimension to obtain the normalized values ​​of each sample in each dimension. The normalized values ​​of each sample in each dimension fall within the range of... Within the interval, the normalized data vector of each sample is calculated based on the normalized values ​​of each sample in each dimension; data from different dimensions are mapped to a unified numerical interval to eliminate scale differences caused by different dimensions and numerical ranges. Normalization is achieved by mini-maximum normalization; No. The sample at the th The original observations in each dimension are represented as follows: , No. The sample at the th The normalized value in each dimension is represented as follows: ,in, For sample index; For dimension indexing; No. Normalized data vectors of samples The calculation formula is: , In the formula, For the first The normalized value of each sample in the first dimension; For the first The normalized value of each sample in the second dimension; For the first The sample at the th The normalized values ​​in each dimension range from 1. This represents the relative position of the original value in this dimension and is used to eliminate dimensional differences. If all values ​​in a dimension are the same, resulting in a denominator of zero, then all normalized values ​​in that dimension are defined as 0. For the first The sample at the th Normalized values ​​in each dimension; This is a transpose operation; S202. Calculate the fusion feature vector of each sample based on the normalized data vector of each sample, and complete the adaptive weighted fusion processing; based on the multidimensional data obtained after minimum-maximum normalization processing, adopt a fusion function that can adaptively weigh the importance of different data sources, and combine the specificity of the data source to perform feature transformation, so as to enhance the expressive power of the fusion features and the sensitivity to local fluctuations. The fused feature vectors of each sample integrate feature information from different data sources. The dimension of the fused feature vector for each sample can be set. For 30; the The fusion feature vector of each sample The calculation formula is: , In the formula, Index the data source; This represents the total number of data sources; the default value is 5. For the first The adaptive weight coefficient of each data source is a scalar used to dynamically characterize the feature contribution of the data source in the current sample. Specifically, it can adaptively adjust the contribution of the data source in the fused features according to the degree to which the current sample deviates from the typical pattern of its data source. For the first A specific feature transformation function for each data source will transform the dimension of... The input subvector is mapped to a dimension of The feature vectors are used to extract more effective source-specific features; For the first Among the samples belonging to the th , Feature vectors from data sources, with dimension . And satisfy Where D is the total dimension of the normalized data vector, which is the sum of all original dimensions; The feature transformation function is implemented using a multilayer perceptron neural network, whose structure includes an input layer, a hidden layer, and an output layer. The dimension of the input layer is... , corresponding to the The feature vectors from each data source are used to learn the non-linear feature interactions within the data source through a hidden layer followed by a ReLU non-linear activation function. The number of neurons in the hidden layer is set to 128, and the output layer dimension is [missing value]. The learned source-specific features are mapped to a unified feature space; No. Adaptive weight coefficients for each data source The calculation formula is: , In the formula, It is a natural exponential function; For the first The weight decay coefficient for each data source is an adjustable hyperparameter. The larger the value, the more sensitive the weights are to changes in distance. It is set independently based on the empirical distribution of the characteristics of each data source or through optimization using a validation set. For example, a larger value can be set for data sources with small numerical ranges and concentrated distributions. (e.g., 2.0 or 5.0) This makes its weights more sensitive to small changes. For data sources with large numerical ranges or dispersed distributions, a smaller value can be set. (e.g., 0.5 or 1.0); It is an L2 norm; For the first The cluster center vectors of the feature vectors of the nth data source are used to perform clustering on the nth data source using the K-means clustering algorithm. The data is calculated from all samples of a data source, representing a typical pattern of data within that data source. S3. Calculate the feature enhancement vector of each sample based on the fused feature vector, the adaptive weight vector of each sample, and the gradient enhancement vector of each sample, and complete the construction of the feature enhancement module based on adaptive weights; S301. Calculate the local sparsity estimate of each sample based on the fusion feature vector of each sample to complete the local density estimation; The estimated local sparsity of each sample is a scalar; a larger value indicates that the sample is more sparse in the feature space, and is more likely to be an anomaly or in a state transition region. To quantify the sparsity of each sample's distribution in its feature space, the K-nearest neighbor algorithm is used to calculate the local sparsity. Local sparsity estimate of each sample The calculation formula is: , In the formula, The number of nearest neighbors is a preset hyperparameter used to define the range of "local"; an example value is 10. Nearest neighbor index; To be in the feature space with Euclidean distance Feature vectors of the nearest samples; For indexing functions, it means returning the index. The first sample The index of the nearest neighbor samples.

[0023] S302. Calculate the gradient magnitude of the feature gradient vector based on the fused feature vector of each sample to obtain feature gradient information; The dimension of the feature gradient vector of each sample is The feature vector is characterized by the specific changes in each dimension over time. In order to capture the drastic changes of features in the time series, the gradient vector and gradient magnitude of each sample feature vector relative to the previous time point are calculated. For samples indexed in time order, their feature gradient vectors are calculated by first-order discrete difference. No. The feature gradient vector of each sample The calculation formula is: , In the formula, For the first The fusion feature vector of each sample; Specifically, for the first sample, the definition is... It is a zero vector; The gradient magnitude of the feature gradient vector of each sample is a scalar, representing the overall drasticness of the feature vector's change over time; a larger value indicates a more rapid state change. The definition of the... The gradient magnitude of the feature gradient vector of the i-th sample is the L2 norm of that gradient vector. Gradient magnitude of the feature gradient vector of each sample The calculation formula is: ; S303. Calculate the adaptive weight vector of each sample based on the gradient magnitude of the feature gradient vector of each sample; The value of each element in the adaptive weight vector of each sample is in Within the interval, it is used to modulate the feature vector element by element; the smaller the value, the higher the degree of enhancement of the corresponding dimension. Based on the gradient magnitudes of the local sparsity estimate and the feature gradient vector, an adaptive weight vector is generated through linear combination and passed through the Sigmoid activation function to achieve element-wise modulation; The adaptive weight vector of each sample The calculation formula is: , In the formula, Use the Sigmoid activation function; This is the sparsity sensitivity coefficient, used to control the strength of the influence of local sparsity on the weights; an example value is 0.5. This is a sparse weight vector with dimension . , is used to map local sparsity estimates to each dimension to adjust the degree of sparsity influence on different dimensions, and is a trainable parameter; This is the gradient sensitivity coefficient, used to control the strength of the gradient vector's influence on the weights; an example value is 0.3. Gradient weight vector, dimension 1 , is used to map gradient magnitude to each dimension to adjust the degree of influence of gradient information on different dimensions, and is a trainable parameter; S304. Calculate the gradient enhancement vector of each sample based on the feature gradient vector of each sample and the fused feature vector of each sample; The gradient enhancement vector dimension for each sample is It is used to additively enhance the original features to highlight features that change dramatically; The feature gradient vector is smoothed and compressed element-wise using the hyperbolic tangent function to limit the range of gradient values ​​and preserve sign information. The smoothed gradient vector is then multiplied element-wise with the fused feature vector, and combined with global enhancement coefficients to obtain a gradient enhancement vector, which is used to additively enhance the original features, highlighting dimensions with drastic changes; Gradient enhancement vector for each sample The calculation formula is: , In the formula, This is the global enhancement coefficient, used to control the overall magnitude of gradient enhancement; an example value is 0.1. It is a hyperbolic tangent function, and its output range is within... This is used to smoothly compress gradient magnitudes; S305. Calculate the feature enhancement vector of each sample based on the fused feature vector of each sample, the adaptive weight vector of each sample, and the gradient enhancement vector of each sample, and complete the feature enhancement operation; The feature enhancement vectors of each sample are smoothed in regions with dense and stable features, and enhanced in regions with sparse or rapidly changing features, with a dimension of [missing information]. ; The adaptive weight vector is fused with the feature vector and multiplied element-wise to suppress or preserve features based on sparsity and gradient information. This is then combined with the gradient enhancement vector to obtain the feature enhancement vector. Feature enhancement vector of each sample The calculation formula is: , In the formula, This is element-wise multiplication.

[0024] S4. Calculate the multi-branch fusion feature vector of each sample based on the feature enhancement vector of each sample, the temporal context vector of each sample, and the local statistical feature vector of each sample; calculate the predicted state probability vector of each sample based on the temporal context-aware feature vector of each sample, and complete the construction of the running state recognition module. S401. Calculate the temporal context vector for each sample based on its acquisition timestamp; the dimension of the temporal context vector for each sample is... It is used to encode the periodicity of time and introduce prior knowledge of external load to help identify typical operating modes in different time periods; To incorporate external periodic factors at the sampling time and prior knowledge of typical load patterns into the recognition process, the timestamp information of the samples is encoded. Temporal context vector of each sample The calculation formula is: , In the formula, It is a sine function used to map periodic time information to a continuous numerical space; For timestamp The hour information extracted from it is an integer ranging from 0 to 23; For the first The collection timestamp of each sample is recorded by the data acquisition system; It is a cosine function used to map periodic time information to a continuous numerical space, and is used in conjunction with a sine function to avoid numerical discontinuity; For timestamp The extracted day of the week information, where 0 represents Monday and 6 represents Sunday, with values ​​ranging from 0 to 6. For timestamps The corresponding load pattern coefficient vector has a dimension of Its value is predefined based on historical experience or domain knowledge, and is used to adjust the weights of relevant features according to different time periods. For example, the vector of working hours on weekdays may take the value of , This is a fixed dimension for the time-series context vector, with a default value of 4.

[0025] S402. Calculate the local statistical feature vector of each sample based on the mean, variance, and skewness of each dimension within the window for each sample; the dimension of the local statistical feature vector of each sample is... It is used to characterize the central tendency, dispersion, and distribution asymmetry of each dimension within a local time window. The dimension of the local statistical feature vector, i.e. ; To utilize the statistical properties of the original normalized data within a local time window to aid state identification, higher-order statistical moments are calculated for the data within the adjacent windows before and after each sample: For the Take a sample as the center and take the samples before and after it. For each sample of normalized data, the mean, variance, and skewness within a window are calculated independently for each dimension, and the statistics of all dimensions are concatenated to form a local statistical feature vector; Local statistical feature vectors of a sample The calculation formula is: , In the formula, This is a vector concatenation operation with a lower bound of j=1 and an upper bound of D; For the first The first sample The mean of each dimension within the window is used to measure the average level of the data within a local window. For the first The first sample The variance of each dimension within the window is used to measure the degree of data fluctuation within a local window. For the first The first sample The skewness of each dimension within the window is used to measure the asymmetry of data distribution within a local window; The calculation formula is: , In the formula, The time window radius is a preset hyperparameter used to define the range of local statistical calculations; an example value is 5. For samples within the window relative to the center sample The offset index, with a value range of 1. to ; For the first The sample at the th Normalized values ​​in each dimension; The calculation formula is: ; The calculation formula is: ; S403. In order to make comprehensive use of feature enhancement information, temporal context information and local statistical information, a gated fusion module is adopted to dynamically weigh the contribution of the three information sources to state recognition. Based on the feature enhancement vector of each sample, the temporal context vector of each sample and the local statistical feature vector of each sample, the multi-branch fusion feature vector of each sample is calculated to complete the multi-branch feature fusion. The dimension of the multi-branch fusion feature vector of each sample is It is a high-dimensional abstract feature vector that integrates feature enhancement vector, temporal context vector, and local statistical feature vector information; the first Multi-branch fusion feature vector of each sample The calculation formula is: , In the formula, For corresponding The adaptive gated vector, with dimension Each element has a value between 0 and 1, and is used to dynamically adjust the contribution of the feature enhancement branch to each dimension in the final fused feature; The feature mapping function for the feature enhancement vector is implemented through a separate linear layer, which transforms the input from dimension [1]. Mapping to Dimensions ; For corresponding The adaptive gated vector, with dimension An adaptive gating vector used to dynamically adjust the contribution weights of temporal context branches in each dimension of the final fused feature; The feature mapping function for temporal context vectors is implemented through a separate linear layer, transforming the input from a dimension... Mapping to Dimensions ; For corresponding The adaptive gated vector, with dimension An adaptive gating vector used to dynamically adjust the contribution weights of local statistical feature branches in each dimension of the final fused feature; The feature mapping function for local statistical feature vectors is implemented through an independent linear layer, transforming the input from dimension... Mapping to Dimensions ,in, To unify the dimension of the feature vectors, it is an adjustable hyperparameter, with an example value of 64; the linear layer, also known as the fully connected layer, is used to project input vectors of different dimensions onto a unified feature space dimension; , and The attention scalar is determined by three information sources. The specific calculation method is as follows: First, an attention scalar is calculated for each branch. This is achieved by performing a dot product between a learnable query and the transformed keys of each branch, followed by normalization. Then, it is normalized using the Softmax function. Finally, the scalar is expanded into a dimension. The vectors are used to ensure that the sum of the corresponding elements of the three gated vectors is 1 in each dimension, thus achieving dynamic weighted fusion.

[0026] S404. Construct a hybrid network containing graph attention layer and recurrent layer to model the fused feature sequence in order to capture the spatial correlation and temporal dependence between samples, and calculate the predicted state probability vector of each sample based on the temporal context-aware feature vector of each sample. The dimension of the predicted state probability vector for each sample is Each element represents the probability that a sample belongs to the corresponding state category; the... The predicted state probability vector of each sample The calculation formula is: , In the formula, The Softmax activation function normalizes the network output into a probability distribution. Here is the weight matrix of the output layer, with dimension 1. , are trainable parameters, where, The total number of predefined operating status categories, specifically including "Normal", "Low Load", "High Load", "Overload", "Abnormal", "Fault", "Startup", "Stop" and "Transitional State"; For the first The temporal context-aware feature vector of the sample is the _th _th _th_ ... The final temporal context-aware feature vector after each sample passes through a bidirectional gated recurrent unit network has a dimension of . ; Let be the bias vector of the output layer, with dimension . , are trainable parameters; Based on the multi-branch fusion feature vector of each sample, the first Temporal context-aware feature vectors of each sample The generation process includes the following steps: (1) Continuous Fusion features of individual samples Viewed as a fully connected graph, a multi-head graph attention network is used to calculate the node feature representation of each sample after being updated by the graph attention network layer. The node feature representation of each sample after being updated by the graph attention network layer is aggregated with the information of the neighboring nodes in the current sequence graph structure. Representing the The node feature representation of each sample after being updated by the graph attention network layer; in, The length of the input sequence is the number of consecutive samples used to construct the graph and perform time series modeling. For the first The multi-branch fusion feature vector of each sample is the fusion feature vector of the starting sample in the sequence; (2) The first Each sample, with its updated node representation sequence based on its neighbor information, is input into a bidirectional gated recurrent unit network. Temporal dependencies are captured from both the forward and backward directions, and the final hidden states from both directions are concatenated to obtain... This enables the network to simultaneously model the structural relationships and short- and long-term temporal dynamics between samples; S5. Calculate the composite loss function based on state recognition loss, temporal consistency loss, and feature reconstruction loss; S501. Calculate the state recognition loss used to optimize the network's accuracy in classifying operating states; A focus loss with class balancing weights is used as the basic recognition loss to address the potential class imbalance problem in the runtime state data, while also making the state recognition model training more focused on difficult-to-classify samples; state recognition loss The calculation formula is: , In the formula, The total number of samples contained in a training batch; Index for status categories; For the first The class state balance weight is set as the reciprocal of the frequency of samples of that class in the training set, which is used to alleviate the class imbalance problem. The value is obtained by statistical analysis of the training dataset. For the first The sample was predicted as the first... The probability of a class state, i.e., the probability vector of the predicted state. The first vector One element; The focus parameter is an adjustable hyperparameter used to reduce the contribution of easily classified samples to the total loss, making the training more focused on difficult-to-classify samples. At this point, the loss degenerates into a weighted cross-entropy loss, with an example value of 2.0; For the first The sample belongs to the first The actual labels for class states are encoded using one-hot encoding. It is a logarithmic function, with the default base being the natural constant; S502. Calculate the temporal consistency loss to encourage the state recognition model to maintain a smooth state distribution when the original features change gradually, thus avoiding unreasonable and drastic jumps. Temporal consistency loss The calculation formula is: , In the formula, This is a very small positive integer used to prevent the denominator from being zero and to ensure numerical stability. An example value is 0.0001. It is an L1 norm; For the first The predicted state probability vector of each sample; For the first The predicted state probability vector of each sample; The absolute value of the change in the key physical quantity to be reconstructed, specifically the gradient magnitude of the feature gradient vector, is denoted as for the i-th sample. ; This is a temperature coefficient used to adjust the scale of the influence of original feature differences on the strength of smoothness constraints. S503. Calculate the feature reconstruction loss. The feature reconstruction loss is used to constrain the intermediate fused features of the network to contain key physical representation information to guide the direction of feature learning. Feature Reconstruction Loss The calculation formula is: , In the formula, d represents the key physical quantity to be reconstructed, specifically referring to the local sparsity estimate, denoted as for the i-th sample. ; This is the first independent multilayer perceptron, whose input is the fused feature vector. The output is a key physical quantity The predicted value; For the target physical quantity The variance over the entire training set is used to standardize the reconstruction error; It is a second independent multilayer perceptron, whose input is the fused feature vector. The output is the absolute value of the change in the physical quantity. The predicted value; The absolute value of the change in the target physical quantity The variance over the entire training set is used to standardize the reconstruction error; and Before training the state recognition model begins, all samples in the entire training set are pre-statistically analyzed. and The calculated constants are used to standardize the reconstruction error, making the reconstruction errors of physical quantities with different dimensions comparable and stabilizing the training process.

[0027] S504. Calculate the composite loss function based on state recognition loss, temporal consistency loss, and feature reconstruction loss; The composite loss function serves as the overall optimization objective for training the state recognition model. It comprehensively improves the classification accuracy, prediction smoothness, and feature interpretability of the state recognition model by jointly minimizing the state recognition error, the temporal jitter of state prediction, and the reconstruction error of key physical information from fused features. The calculation formula is: , In the formula, This is the weighting coefficient for the time consistency loss, an adjustable hyperparameter used to control the strength of this loss's impact on the total loss; an example value is 0.1. is the weighting coefficient for the feature reconstruction loss, which is an adjustable hyperparameter, with an example value of 0.05; S6. Randomly sample a sample from the training dataset, and repeat steps S2, S3, S4, and S5 in sequence until the final goal is achieved, thus obtaining the trained state recognition model; The ultimate goal is to minimize the composite loss function; Forward propagation is performed according to the complete process defined above: S2, S3, S4, and S5 are executed sequentially to obtain the predicted state probability distribution of all samples in the current batch and the composite loss value of the current batch; training adopts a variant optimization algorithm of stochastic gradient descent, calculates the gradient based on the composite loss value and performs backpropagation to update all trainable parameters in the state recognition model.

[0028] During training, the performance of the state recognition model is periodically evaluated using an independent validation set, and metrics such as loss and state recognition accuracy on the validation set are monitored. The stopping criterion for training the state recognition model is primarily based on the validation set performance, employing an early stopping strategy. Specifically, training is terminated when the composite loss on the validation set no longer decreases over several consecutive training epochs, and the model is rolled back to the snapshot of state recognition model parameters with the lowest validation loss. This avoids overfitting and obtains the state recognition model with the best generalization ability. Additionally, a maximum number of training epochs can be set as an auxiliary stopping criterion.

[0029] S7. Collect new real-time operating status data, input the new real-time operating status data into the trained status recognition model, output the predicted status probability vector of the new real-time operating status data, take the operating status with the highest probability value as the real-time operating status of the intelligent fusion terminal, and display and record it visually. Early warning rules can be set to trigger alarm notifications when the real-time operating status is "overload," "abnormal," or "fault," facilitating timely inspection and maintenance by staff.

Claims

1. A method for multi-dimensional monitoring of the real-time operating status of an intelligent fusion terminal, characterized in that, include: S1. Collect multiple real-time operating status data with timestamps from the intelligent fusion terminal at fixed time intervals. Each real-time operating status data with timestamps is a sample. Each sample includes multiple dimensions of data from multiple data sources. Label the operating status of each sample and use all labeled samples as training dataset. S2. Based on the data of each sample in each dimension, obtain the normalized data vector of each sample, and calculate the fusion feature vector of each sample based on the normalized data vector of each sample. S3. Calculate the feature enhancement vector of each sample based on the fused feature vector of each sample, the adaptive weight vector of each sample, and the gradient enhancement vector of each sample; S4. Calculate the multi-branch fusion feature vector of each sample based on the feature enhancement vector of each sample, the temporal context vector of each sample, and the local statistical feature vector of each sample; calculate the predicted state probability vector of each sample based on the temporal context-aware feature vector of each sample. S5. Calculate the composite loss function based on state recognition loss, temporal consistency loss, and feature reconstruction loss; S6. Repeatedly sample a sample randomly from the training dataset and execute S2, S3, S4, and S5 in sequence until the final goal is achieved, and the trained state recognition model is obtained. S7. Collect new real-time operating status data, input the new real-time operating status data into the trained state recognition model, output the predicted state probability vector of the new real-time operating status data, and take the operating status with the highest probability value as the real-time operating status of the intelligent fusion terminal.

2. The method for multi-dimensional monitoring of the real-time operating status of an intelligent fusion terminal according to claim 1, characterized in that, Step S2 includes: S201. Perform independent normalization on the original observation values ​​of each sample in each dimension to obtain the normalized values ​​of each sample in each dimension, and calculate the normalized data vector of each sample based on the normalized values ​​of each sample in each dimension. S202. Calculate the fusion feature vector of each sample based on the normalized data vector of each sample.

3. The method for multi-dimensional monitoring of the real-time operating status of an intelligent fusion terminal according to claim 2, characterized in that, The formula for calculating the fusion feature vector of each sample in S202 is as follows: ， In the formula, For the first The fusion feature vector of each sample; Index the data source; The total number of data sources; For the first Adaptive weighting coefficients for each data source; For the first A feature transformation function specific to a data source; For the first Among the samples belonging to the th , Feature vectors of each data source.

4. The method for multi-dimensional monitoring of the real-time operating status of an intelligent fusion terminal according to claim 1, characterized in that, The steps in S3 include: S301. Calculate the local sparsity estimate of each sample based on the fused feature vector of each sample; S302. Calculate the gradient magnitude of the feature gradient vector based on the fused feature vector of each sample; S303. Calculate the adaptive weight vector of each sample based on the gradient magnitude of the feature gradient vector of each sample; S304. Calculate the gradient enhancement vector of each sample based on the feature gradient vector of each sample and the fused feature vector of each sample; S305. Calculate the feature enhancement vector of each sample based on the fused feature vector of each sample, the adaptive weight vector of each sample, and the gradient enhancement vector of each sample.

5. The method for multi-dimensional monitoring of the real-time operating status of an intelligent fusion terminal according to claim 4, characterized in that, The formula for calculating the feature enhancement vector of each sample in S305 is as follows: ， In the formula, For the first Feature enhancement vectors for each sample; For the first The fusion feature vector of each sample; This is element-wise multiplication; For the first An adaptive weight vector for each sample; For the first Gradient enhancement vector for each sample.

6. The method for multi-dimensional monitoring of the real-time operating status of an intelligent fusion terminal according to claim 1, characterized in that, The steps in S4 include: S401. Calculate the temporal context vector of each sample based on the collection timestamp of each sample; S402. Calculate the local statistical feature vector of each sample based on the mean of each dimension within the window, the variance of each dimension within the window, and the skewness of each dimension within the window. S403. Calculate the multi-branch fusion feature vector of each sample based on the feature enhancement vector of each sample, the temporal context vector of each sample, and the local statistical feature vector of each sample; S404. Calculate the predicted state probability vector of each sample based on the temporal context-aware feature vector of each sample.

7. The method for multi-dimensional monitoring of the real-time operating status of an intelligent fusion terminal according to claim 6, characterized in that, The formula for calculating the multi-branch fusion feature vector of each sample in S403 is as follows: ， In the formula, For the first Multi-branch fusion feature vector of each sample; For corresponding The adaptive gating vector; This is element-wise multiplication; The feature mapping function for the feature enhancement vector; For corresponding The adaptive gating vector; The feature mapping function for the temporal context vector; For the first The temporal context vector of each sample; For corresponding The adaptive gating vector; It is the feature mapping function of the local statistical feature vector; For the first Local statistical feature vectors of a sample.

8. The method for multi-dimensional monitoring of the real-time operating status of an intelligent fusion terminal according to claim 1, characterized in that, The steps in S5 include: S501. Calculate the state recognition loss used to optimize the network's accuracy in classifying operating states; S502. Calculate the timing consistency loss; S503. Calculate the feature reconstruction loss; S504. Calculate the composite loss function based on state recognition loss, temporal consistency loss, and feature reconstruction loss.

9. A method for multi-dimensional monitoring of the real-time operating status of an intelligent fusion terminal according to claim 8, characterized in that, The formula for calculating the composite loss function is as follows: ， In the formula, It is a composite loss function; State recognition loss; This is the weighting coefficient for the time-series consistency loss; This results in a loss of time-series consistency. The weighting coefficients for the feature reconstruction loss; The loss is for feature reconstruction.

Images