A cloud computing-based electric energy monitoring system and method

Abstract

This invention belongs to the field of power monitoring technology, specifically relating to a cloud-based power monitoring system and method. It constructs a hybrid channel feature vector through wavelet packet decomposition and link index fusion; predicts the probability distribution of future channel states using neural network processing; performs online dictionary learning and sparse coding on power waveforms to generate data value assessment vectors; dynamically decides on coding schemes, compression levels, and network interfaces based on a reinforcement learning model, combining channel prediction and data value; and performs state monitoring and health assessment in the cloud based on received data-driven digital twin models. Through edge-cloud intelligent collaboration, it achieves reliable, real-time, and efficient transmission and intelligent analysis of power monitoring data in complex industrial environments with strong interference, effectively supporting predictive maintenance and improving the safety and economy of the power system.

Description

Technical Field

[0001] This invention belongs to the field of power monitoring technology, specifically relating to a cloud computing-based power monitoring system and method. Background Technology

[0002] Power monitoring systems are a key technological means to ensure the safe and economical operation of power systems. Their core function lies in the real-time acquisition, analysis, and processing of critical electrical parameters such as voltage, current, power, and power quality. With the development of the Industrial Internet and smart grids, traditional localized, isolated monitoring systems can no longer meet the needs of large-scale, cross-regional, and refined management. The introduction of cloud computing technology has brought a new paradigm to power monitoring.

[0003] However, when facing advanced application scenarios such as predictive maintenance with extremely high requirements for real-time performance, reliability, and in-depth analysis, especially in complex environments such as rail transit and data centers, the transmission of massive amounts of high-frequency sampling data (such as MHz-level waveform data used to capture voltage sags and harmonic events) to the cloud faces severe challenges due to the presence of strong electromagnetic interference and dynamic changes in network topology in scenarios like rail transit. The transmission mechanisms of existing systems relying on public wireless networks or general industrial networks are difficult to guarantee deterministic low latency and high reliability delivery of critical data in scenarios with strong electromagnetic interference and dynamic changes in network topology. Data packets are prone to loss, out-of-order delivery, or severe delays in complex environments, resulting in time gaps or distortions in the data stream received by the cloud, which cannot respond in a timely manner to the technical requirements of timely response to transient time in the application scenarios. In addition, the existing architecture tends to upload unprocessed raw streams or simple aggregated data directly to the cloud for computation. This firstly puts enormous pressure on network bandwidth. Moreover, for predictive maintenance that requires rapid response (such as early fault identification based on instantaneous characteristics), the latency of data transmission, coupled with the latency of cloud task scheduling and processing, makes the overall system response time too long, failing to meet the real-time requirements of equipment protection or emergency control, and losing the lead time for prediction. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention provides a cloud-based power monitoring system and method. This system solves the problem in existing power monitoring systems for rail transit and data centers that combine cloud processing. Due to strong electromagnetic interference, dynamic changes in network topology, and the large amount of unprocessed data uploaded from the edge, it is impossible to simultaneously guarantee the deterministic transmission of key monitoring data and the real-time nature of analysis and decision-making, leading to the failure of subsequent predictive maintenance.

[0005] The objective of this invention can be achieved through the following technical solution: a cloud computing-based power monitoring system, the system comprising: The data acquisition and processing module is used to acquire raw waveform data of power monitoring in real time, decompose the raw waveform data using wavelet packets, and construct a hybrid channel feature vector by combining communication link statistical indicators. The channel prediction module is used to input the hybrid channel feature vector into a channel prediction model constructed based on a neural process network to obtain a predicted distribution of future channel states. The coding evaluation module is used to perform sparse coding based on online dictionary learning on the raw waveform data and generate a data value evaluation vector based on the difference between the calculated sparse coding coefficients and the historical benchmark. The transmission control module is used to select the optimal transmission decision for the current data batch through a reinforcement learning policy network based on the predicted channel state, data value evaluation vector, and current communication link state. The optimal transmission decision includes the coding scheme, compression level, and network interface. The data transmission module encodes the data using a rate-distortion-optimized layered encoder based on the optimal transmission decision, and uploads the encoded data packets to the cloud processing module. The cloud processing module is used to monitor the electrical behavior of the monitoring nodes based on the encoded data and output status prediction and health assessment results.

[0006] Preferably, in the data acquisition and processing module, calculating the hybrid channel feature vector includes: The original waveform data is decomposed into N-level wavelet packets, and the energy proportion, information entropy, and distortion degree of the sub-band coefficients corresponding to the preset interference frequency band of each sub-band in at least the last level are calculated to obtain the time series characteristics. The time-series features and communication link statistics are used to generate a hybrid channel feature vector using a feature fusion algorithm.

[0007] Preferably, the formula for calculating the distortion degree of the sub-band coefficient is: ; In the formula, D represents the subband coefficient distortion degree, and M represents the set of subband indices covering the interference frequency band. It is the variance of the coefficients of subband j. It is the average variance under normal historical conditions.

[0008] Preferably, the channel prediction module includes constructing a channel prediction model using a variational conditional neural network, the model structure of which includes: The temporal context observation coding layer is used to obtain the channel observation set at historical moments. For each observation pair in the channel observation set, feature encoding is performed, and then the encoded features are globally aggregated to obtain the global context representation. A variational latent channel mode distribution layer is used to generate a latent probability distribution based on the global context representation and to sample latent variables from the latent probability distribution; The future channel quality probability prediction decoding layer is used to obtain the channel input set at future time, decode the channel input at each future time in combination with latent variables, and output the probability distribution parameters of the channel quality at the corresponding time. The probability distribution parameters are the predicted distribution of the future channel state.

[0009] Preferably, the encoding evaluation module includes a data preprocessing unit, a dictionary encoding unit, and an evaluation vector generation unit that are sequentially connected in communication, wherein: The data preprocessing unit is used to segment the original waveform signal into analysis windows of fixed duration and perform denoising and baseline correction. The dictionary encoding unit is used to establish and update the dynamically adaptive feature dictionary, and to represent the preprocessed electrical waveform data as a sparse linear combination of atoms in the dictionary through a sparse encoding algorithm. The evaluation vector generation unit is used to analyze sparse coding results, quantify the degree of anomaly and importance of the current data window, and generate standardized data value evaluation vectors.

[0010] Preferably, the dictionary encoding unit is used to perform the following process: A feature dictionary is constructed, which includes predefined base atoms and adaptive atoms obtained through online learning. The base atoms are generated based on typical waveform features of the power system, and the adaptive atoms are dynamically generated based on actual data during operation. An online dictionary learning algorithm is used to periodically update the dictionary atoms using newly collected normal operating condition data; Sparse coding is performed using an orthogonal matching pursuit algorithm to obtain sparse coding coefficients. The sparse coding process aims to minimize a preset objective function to obtain sparse coefficients that can accurately represent the original waveform with the fewest dictionary atoms.

[0011] Preferably, the formula for calculating the objective function is: ; In the formula, Let be the objective function. For error reconstruction term, The sparsity regularization term constrains the sparsity of the sparse coding coefficients using the L1 norm. Z is the dictionary matrix, and T is the total number of time windows. Let t be the vector of the t-th data window. Let be the sparse coding coefficients corresponding to the t-th data window. This is the regularization parameter.

[0012] Preferably, the evaluation vector generation unit is used to perform the following process: Based on sparsity, the sparsity is compared with the historical sparsity benchmark, and the relative rate of change is calculated as the sparsity mutation degree. Identify the rare atoms in the currently activated dictionary atoms and calculate the weighted activation energy of these rare atoms; Calculate the coefficient energy distribution entropy using the Shannon entropy formula; The calculated sparsity mutation degree, weighted activation energy, and coefficient energy distribution entropy are combined in sequence and normalized to obtain the value assessment vector.

[0013] Preferably, the dictionary encoding unit includes iteratively solving the objective function using the block coordinate descent method.

[0014] A cloud computing-based power monitoring method, comprising the following steps: S1: Real-time acquisition of raw waveform data for power monitoring, decomposition of raw waveform data using wavelet packets, and construction of hybrid channel feature vectors by combining communication link statistical indicators; S2: Input the hybrid channel feature vector into the channel prediction model constructed based on neural process network to obtain the predicted distribution of future channel states; S3: Perform sparse coding based on online dictionary learning on the original waveform data, and generate a data value assessment vector based on the difference between the calculated sparse coding coefficients and the historical benchmark. S4: Based on the predicted channel state, data value assessment vector, and current communication link state, the optimal transmission decision for the current data batch is selected through a reinforcement learning policy network. The optimal transmission decision includes the coding scheme, compression level, and network interface. S5: Based on the optimal transmission decision, the data is encoded using a rate-distortion optimized hierarchical encoder, and the encoded data packet is uploaded to the cloud processing module. S6: Monitor the electrical behavior of the monitoring nodes based on the encoded data, and output status prediction and health assessment results.

[0015] The beneficial effects of this invention are as follows: This invention fundamentally solves the deterministic and real-time challenges of critical monitoring data transmission in complex industrial environments by achieving cross-layer deep perception and probabilistic prediction of channel interference characteristics at the edge, as well as unsupervised intelligent evaluation of the value of power data. Secondly, based on a dynamic transmission decision mechanism using multi-objective reinforcement learning, it achieves intelligent joint optimization of coding strategies, compression levels, and network interfaces. While ensuring the priority and reliable transmission of high-value data, it reduces overall network bandwidth pressure and shortens the latency between the edge and the cloud, thereby ensuring the deterministic transmission of critical monitoring data, the real-time nature of analysis and decision-making, and the timeliness required for subsequent predictive maintenance in the cloud. Attached Figure Description

[0016] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0017] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a schematic diagram illustrating the execution process of the dictionary encoding unit in the system of the present invention; Figure 3 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0019] Please see Figures 1-3 This embodiment provides a cloud-based power monitoring system, comprising a data acquisition and processing module, a channel prediction module, a coding evaluation module, a transmission control module, and a data transmission module deployed in a local processor at the edge, and a cloud processing module deployed in the cloud. The system includes: The data acquisition and processing module is used to acquire raw waveform data for power monitoring in real time, and to decompose the raw waveform data using wavelet packets, communication link statistical indicators, and construct a hybrid channel feature vector. Edge devices continuously monitor their wireless communication interfaces, capturing baseband or radio frequency signals from the physical layer. These signals typically contain the expected communication signals (such as preambles or pilot sequences for Wi-Fi, 4G / 5G data packets) as well as various electromagnetic noises and interference superimposed on them. In environments such as rail transit, these interferences may originate from frequency converters of traction motors, high-power switching power supplies, etc. The system digitally samples a segment of continuous time-domain signal to obtain a discrete time series. The discrete time series is input into a wavelet packet decomposition algorithm with a preset depth of N layers (e.g., 6 layers) to obtain a complete binary tree structure. A set of preset high-pass and low-pass filters are used to iteratively filter and downsample the signal, decomposing the original signal into multiple fine sub-bands in the time-frequency plane. Each sub-band corresponds to a specific frequency range and contains the coefficients of the signal changing with time within that frequency band. Compared with traditional Fourier transform or wavelet transform, wavelet packet decomposition can provide more uniform and flexible time-frequency resolution, and is particularly good at capturing non-stationary and transient interference features. Extract subband features. For all subbands of the last level (Nth level) obtained from the decomposition, calculate the following features for each subband (i.e., leaf node) of the last level, such as multiple subbands of the Nth level: Energy percentage: Calculate the ratio of the sum of squares of all coefficients in each sub-band (representing the energy of that frequency band) to the total energy of all sub-bands. This is used to clearly reveal the distribution of energy generated by interference in different frequency bands. For example, if the energy in a certain frequency band (such as the harmonic of the corresponding IGBT switching frequency) increases significantly, it indicates the existence of this type of periodic pulse interference. Information entropy: The entropy value of the probability distribution of the amplitude of each sub-band coefficient is calculated. A high entropy value indicates that the signal of that sub-band is complex and has strong randomness, which may correspond to broadband noise or chaotic interference; a low entropy value indicates that the signal has strong regularity, which may correspond to periodic interference or the signal itself. The system calculates the subband coefficient distortion. It predefines one or more hazardous frequency bands requiring special attention, such as interference bands prone to communication failures identified based on historical fault data. Specific subbands covering these hazardous frequency bands are located, and the statistical characteristics (such as variance) of their coefficients are calculated, along with the distortion degree compared to the baseline value recorded under historical clean conditions. This distortion degree directly quantifies the severity of interference in the current hazardous frequency band. The interference bands prone to communication failures identified based on historical fault data are then designated as... The formula for calculating the subband coefficient distortion degree D is: ; In the formula, D represents the distortion degree, and M represents the set of sub-band indices covering the interference frequency band. It is the variance of the coefficients of subband j. It is the average variance under normal historical conditions; Traditional communication link statistics, including Received Signal Strength (RSSI), Bit Error Rate (BER), and Retransmission Rate (RTR), are obtained from the data link layer (layer 2 of the OSI model). These metrics are then normalized with time-frequency features extracted from the original waveform data (energy proportion of each subband, entropy, and distortion degree of a specific frequency band) to eliminate the influence of dimensions. A feature fusion algorithm is used to fuse these features into a hybrid channel feature vector. By fusing the time-frequency features (energy distribution, entropy, distortion degree) that characterize the physical nature of interference with the link layer metrics (RSSI, BER, RTR) that characterize communication performance into a unified feature vector, this vector serves as the input to the subsequent intelligent decision-making model, describing the comprehensive channel state that affects the data transmission quality of power monitoring.

[0020] This paper deeply couples communication deep signal analysis (wavelet packet decomposition) with reliable transmission of power monitoring data. Traditional power monitoring systems only monitor data transmission quality at the link layer BER or simple RSSI. By extracting the spectrum composition of interference, the problem of strong interference is solved.

[0021] The channel prediction module is used to input the hybrid channel feature vector into a channel prediction model based on a neural process network to obtain a predicted distribution of future channel states. Here, the neural network (NP network) is a deep learning model that combines the flexibility of neural networks with the probabilistic advantages of Gaussian processes. First, a temporal context observation encoder learns to extract a latent probability distribution characterizing the dynamic properties of the channel from historical observation data. Then, based on this latent probability distribution, a decoder predicts the mean and variance of the output value for future inputs, outputting a predicted distribution of the channel score at future times. Preferably, a variational conditional neural network (VCNP) is used to construct the channel prediction model, and the model structure includes: The temporal context observation coding layer is used to obtain the channel observation set at historical moments. For each observation pair in the channel observation set, feature encoding is performed, and then the encoded features are globally aggregated to obtain the global context representation r. The feature encoding of the temporal context observation encoding layer is implemented through the encoder network h, specifically a multilayer perceptron (MLP). The formula for calculating the encoded features of a single observation pair at time i is: ; In the formula, For the tth Channel inputs at time i (such as channel state and transmission parameters). For the tth Overall channel quality score at time i (historical output). The mean value of all single-observation features is taken to obtain the context global representation r, which is used to integrate all historical channel information; A variational latent channel mode distribution layer is used to learn the parameters of the latent probability distribution from the global context representation r through two independent MLP parameterization networks. The parameters include a mean network and a variance network. Then, the latent probability distribution is sampled normally to obtain the latent variable z. In the formula, The context global representation r is obtained by processing it using the mean network MLP. To obtain the global context representation r using a variance network MLP, a variational latent channel mode distribution layer is used to capture implicit dynamic patterns (such as channel fading periods and interference correlations) not explicitly expressed in historical channel observations, thereby improving the model's generalization ability. The future channel quality probability prediction decoding layer is used to obtain the channel input set at the future time t+i. It combines the latent variable z to decode the channel input at each future time and outputs the probability distribution parameters of the channel quality at the corresponding time. The probability distribution parameters include the prediction mean and the prediction variance. The normal distribution of the prediction mean and the prediction variance is the channel quality prediction distribution at the future time. The channel quality prediction distribution is used to characterize the predicted channel quality value and uncertainty at the future time.

[0022] The training and inference of a neural network model includes the following processes: Training phase: The NP network is trained offline using the acquired historical channel data. The training data consists of a series of time segments, each containing a mixed feature vector (as context) and the corresponding channel quality label (as observation) for a past period. The network learns to infer the pattern of channel changes from the context. In the online inference phase: At the current moment, the system inputs the observed hybrid feature vector and its corresponding channel quality score from the previous moment as context points into the NP network. Simultaneously, it inputs the hybrid feature vectors from the next few moments (which can be extrapolated through time series or set as variables to be predicted) as target points. The NP network outputs the predicted distribution of channel quality scores for these future moments, typically represented as a mean (the most likely channel quality) and a variance (the uncertainty of the prediction). The larger the variance, the more uncertain the prediction. The channel prediction model obtains the distribution of future channel quality by using historical channel observations, generating a global representation in the coding layer, sampling implicit variables in the latent distribution layer, and combining the probability prediction of future inputs and outputs in the decoding layer. By predicting the distribution of channel quality scores for future moments, the system quantitatively assesses the risk of future channels, thereby obtaining the communication quality of the channel in the future and determining the data transmission method.

[0023] The encoding evaluation module performs sparse coding based on online dictionary learning on the raw waveform data. Based on the difference between the pattern of the obtained sparse coding coefficients and the historical benchmark, it generates a data value evaluation vector. The encoding evaluation module includes a data processing unit, a dictionary coding unit, and an evaluation vector generation unit. Each unit performs the following functions: The data preprocessing unit is used to collect raw waveform data from power monitoring equipment (such as merging units and smart terminals), including three-phase voltage and current signals. These raw waveform data are continuously input at a high sampling rate (such as 1MHz). The data is first divided into analysis windows of fixed duration (such as each power frequency cycle or a 10ms window) and then denoised and baseline corrected. The dictionary encoding unit is used to build and update a dynamically adaptive feature dictionary, and to represent the preprocessed electrical waveform data as a sparse linear combination of atoms in the dictionary through a sparse encoding algorithm. The feature dictionary initially includes basic atoms and adaptive atoms. The basic atoms are pre-trained based on typical waveform features of the power system, such as standard sinusoidal fundamental wave, common harmonic waveforms, and typical voltage sag / surge patterns. The adaptive atoms are dynamically learned and generated based on actual data during operation. The dictionary is stored in matrix form, with each atom being a standardized typical waveform segment. The dictionary size is configurable (e.g., 256 atoms), and the length of each atom matches the analysis window. The online dictionary update process includes: First, construct a first-in-first-out (FIFO) buffer to store the data from the most recent M analysis windows. and the corresponding sparse coding coefficients ; A dictionary update is triggered every T analysis windows or when the cache is full; Let an objective function be defined as the update objective, the objective being to update the sparse coding coefficients given the current set of sparse coding coefficients. In this case, we need to find a better dictionary Z that minimizes the reconstruction error of all samples while maintaining the normalization constraint of atoms. For each real-time acquired raw waveform data, the sparse encoding process of the dictionary encoding unit includes: first, preprocessing the raw waveform signal, and then solving for the sparse representation. The objective function of dictionary learning optimization in online scenarios is:

[0024] In the formula, Let be the objective function. This is the error reconstruction term, used to measure the difference between the linear combination of the dictionary matrix and sparse coding coefficients and the original data window. The smaller the value, the more accurate the reconstruction. As a sparse regularization term, the sparsity of the sparse coding coefficients is constrained by the L1 norm, ensuring that reconstruction can be achieved with only a small number of dictionary atoms, thus fulfilling the core requirement of sparse representation. Z is a dictionary matrix, which consists of K elements, each containing an atom. It is a normalized waveform prototype, where T is the total number of time windows. That is, in an online scenario, the time-series data to be processed is divided into T consecutive windows. Let be the vector of the t-th data window, and be the input data to be reconstructed from the dictionary. Let be the sparse coding coefficients corresponding to the t-th data window, representing the reconstruction using linear combinations of dictionary atoms. The weights must satisfy the sparsity property. This is a regularization parameter used to control the sparsity of the sparse coding coefficients. The larger the value, the fewer non-zero values ​​there are in the sparse coding coefficients. The objective function is solved iteratively using the block coordinate descent method: Based on a fixed dictionary matrix, the corresponding sparse coding coefficients are updated. For the fixed sparse coding coefficients, the dictionary atoms are updated column by column. For each atom, the set of all sample indices that use the atom is found. The residual matrix of the dictionary atom is calculated using the sample index set and the atom is updated. The above update steps are iterated until the objective function converges or the preset maximum number of iterations is reached. Elimination and updating of dictionary atoms: The system records the historical activation frequency of each atom. The historical activation frequency is the frequency at which the coefficient of the atom is non-zero in the sparse coding. If the historical activation frequency of an atom is lower than the set threshold for a long time, it is considered "outdated" or "invalid" and removed from the dictionary. When the system detects a new waveform pattern that is persistent and cannot be well sparsely represented by the existing dictionary, it will extract new candidate atoms from the samples of these new patterns through methods such as clustering, replace the eliminated atoms, and keep the dictionary size K unchanged. For each newly arrived analysis window of data By solving the following constraints, the sparsest representation coefficients are found under the current dictionary Z. :

[0025] In the formula, This represents the L0 norm, which is the number of non-zero elements. It is a preset reconstruction error tolerance; The Orthogonal Matching Pursuit (OPM) algorithm is used for an approximately efficient iterative solution, ultimately outputting sparse coding coefficients. ; The evaluation vector generation unit is used to analyze sparse coding results, quantify the anomaly degree and importance of the current data window, generate a standardized data value evaluation vector, and extract features from the sparse coefficient vector based on the sparse coding pattern that can quantify the potential importance or anomaly degree of the current data window, and encapsulate them into a value evaluation vector. This includes the following processes: Calculate the total number of non-zero coefficients after this encoding, which is the sparsity of the signal. Low sparsity means that the signal structure is simple and may be composed of only a few main waveform patterns superimposed; high sparsity means that the signal components are complex and require more basic atoms to describe. The sparsity abruptness is calculated to describe the abrupt changes in signal complexity. The current sparsity is compared with the historical sparsity baseline (absolute difference), and the relative rate of change is calculated to obtain the sparsity abruptness. This determines whether there is an anomaly in the current sparsity. The historical baseline is obtained by continuously tracking the sparsity of all signals over a period of time using the exponential moving average algorithm. It is used to quantify the typical level of signal complexity under normal operating conditions. During long-term operation, atoms in the dictionary are assigned different rarity weights based on their frequency of use. Atoms that frequently appear in normal waveforms (such as fundamental atoms and common harmonic atoms) have lower weights, while atoms that are rarely activated and usually only appear in specific faults or special events are assigned a rarity weight for each dictionary atom. This weight is inversely proportional to the activation frequency of the atom in historical normal data. When the rarity weight is greater than a set threshold, it is identified as a rare atom. The part of the currently activated dictionary atoms that belongs to rare atoms is identified, and the weighted activation energy of these rare atoms is calculated. The weighted activation energy is used to characterize the frequency of occurrence of rare waveforms. The activation intensity of these rare atoms (the square of their corresponding coefficients) is calculated and weighted summation is performed to quantify the activation energy of rare atoms using weighted summation. Calculate the coefficient energy distribution entropy, analyze the energy distribution among the currently activated non-zero coefficients, normalize the square values ​​(representing energy) of all non-zero coefficients and treat them as a probability distribution. Based on this probability distribution, calculate the coefficient energy distribution entropy using the Shannon entropy formula to quantify the concentration of energy distribution. Finally, the sparsity mutation degree, rare atom activation energy, and coefficient energy distribution entropy calculated above are combined in sequence to generate an original value vector, which is then normalized so that the values ​​of each dimension fall within a similar range, making it easier for subsequent algorithms to process.

[0026] By processing a high-dimensional waveform vector of length L as input through online dictionary learning, sparse coding, and pattern analysis, the output is a low-dimensional (three-dimensional) value assessment vector containing multiple informational elements. This value assessment vector compactly encodes the anomalies and potential importance of the current data window, providing the most crucial judgment basis for subsequent dynamic transmission decisions. The entire process enables data understanding and evaluation at the edge, reduces computational complexity at the edge through sparse coding algorithms, improves processing speed, and ensures that only high-value data needs to be transmitted in its entirety, while low-value data can be significantly compressed or only a statistical summary is transmitted. In actual rail transit scenarios, this reduces the amount of data uploaded and improves the capture rate of important events during data acquisition and uploading.

[0027] The transmission control module is used to select the optimal transmission decision for the current data batch through a reinforcement learning policy network based on the predicted channel state, data value assessment vector, and current communication link state. The optimal transmission decision includes the encoding scheme, compression level, and network interface. The channel state includes the channel quality prediction mean vector and variance vector. The communication link state includes the data interface type, current signal strength, historical packet loss rate, bandwidth and current queue length. Combined with the data value assessment vector, after unifying the dimensions of multiple data input to the transmission control module, the feature vectors are concatenated to obtain a fixed-dimensional vector state S. A policy network is constructed using the Deep Reinforcement Learning (DRL) framework to output the optimal transmission decision based on the state S. The DRL employs a fully connected, dual-branch network with an attention mechanism. The backbone branch extracts global features of the state, while the attention branch assigns a higher vector to the data value and channel mean, reinforcing the influence of core factors on the decision. The network input is the state vector S, and the output is the probability distribution of the action space. Corresponding to the three dimensions of optimal transmission decision, a discrete action space is defined and constructed. The discrete action space includes coding scheme, compression level and network interface. The coding scheme includes wavelet packet coding, sparse dictionary coding and hybrid coding. A multi-objective reward function is constructed to guide the network to learn the optimal decision. The multi-objective reward function integrates data value matching reward, channel adaptation reward and link utilization reward and is generated by weighted fusion using a weighted fusion formula. The policy network outputs the probability of each action through the Softmax function, selects the action combination with the highest probability as the optimal transmission decision for the current data batch, and collects post-transmission feedback (such as actual distortion and transmission success rate) to update network parameters, minimize the loss function of long-term cumulative reward, and achieve iterative optimization of the policy. The data transmission module, based on the optimal transmission decision, uses a rate-distortion-optimized layered encoder to encode the data, uploads the encoded data packets to the cloud, and realizes intelligent compression encoding of the data, optimally balancing information fidelity and compression rate within a limited bitrate budget.

[0028] The data transmission module includes a data layering unit and a layered encoding unit. The data layering unit is used to divide the data to be transmitted into two layers: Feature layer: Composed of core feature data such as sparse coefficient vectors generated by calculation. This is the minimum necessary information for subsequent cloud processing modules to perform status analysis and fault diagnosis, and it must be transmitted with the highest possible fidelity. Residual layer: The difference between the original waveform data and the waveform reconstructed using the feature layer data. It contains details, noise, and subtle early fault features (such as the extremely high frequency components of the arc) that the feature layer failed to capture. Since the total transmission bandwidth (bit rate) is limited, the bit rate allocated to each layer is optimized using the particle swarm optimization algorithm to determine the bit rate budget for each layer. After determining the bitrate budget for each layer, the feature layer data and residual layer data are processed using state-of-the-art compression coding techniques (such as transform coding, quantization, and entropy coding). Coding parameters (such as quantization step size) are adjusted in real-time based on the allowable distortion limit for that layer (determined by the rate-distortion function of the allocated bitrate) to ensure optimal coding quality at a given bitrate. The encoded feature layer data, residual layer data (if a bitrate has been assigned), and necessary metadata (such as logical sequence number, data value score, dictionary version identifier used for sparse coding, feature digest hash, etc.) are encapsulated into data packets that conform to the network protocol. The logical sequence number and feature digest hash carried in the data packets enable the cloud processing module to accurately reassemble the temporal order of the data stream when a small amount of packet loss or out-of-order delivery occurs. Based on the context information of the preceding and following data packets, the module performs a certain degree of intelligent interpolation reconstruction of the lost data packet content based on electrophysical laws, thereby minimizing the impact of time gaps on continuous analysis.

[0029] The cloud processing module is used to monitor the electrical behavior of the corresponding monitoring nodes based on the encoded data, and output status prediction and health assessment results. When the cloud receives the encoded data packets uploaded by the edge nodes, it performs data decoding and integrity verification, time domain alignment and multi-source fusion, and feature extraction and enhancement. Based on the monitoring indicators, it uses a pre-trained monitoring model to monitor the acquired and uploaded data.

[0030] Among them, data decoding and integrity verification include identifying the encoding scheme and compression level according to the packet header information, executing the corresponding decoding algorithm, using error correction codes to detect and correct errors that may occur during transmission, and for partially lost data packets, using conditional generative adversarial networks to intelligently interpolate and reconstruct them based on the context information and time series patterns of the preceding and following data, in order to restore data integrity to the greatest extent. Temporal alignment and multi-source fusion include: precise timestamp alignment of data uploaded from different edge nodes to eliminate timing misalignment caused by network latency, and spatial alignment and correlation analysis of data from multiple monitoring points on the same electrical circuit (such as incoming, outgoing, and load ends). Feature extraction and enhancement include: extracting electrical features from the decoded data, including but not limited to: effective voltage / current values, harmonic content, power factor, three-phase imbalance, voltage fluctuation and flicker; performing inverse transformation and enhancement on the sparse features extracted from the edge side; and combining the more powerful computing capabilities of the cloud to perform deeper feature mining.

[0031] A cloud computing-based power monitoring method, comprising the following steps: S1: Real-time acquisition of raw waveform data for power monitoring, decomposition of raw waveform data using wavelet packets, and construction of hybrid channel feature vectors by combining communication link statistical indicators; S2: Input the hybrid channel feature vector into the channel prediction model constructed based on neural process network to obtain the predicted distribution of future channel states; S3: Perform sparse coding based on online dictionary learning on the original waveform data, and generate a data value assessment vector based on the difference between the calculated sparse coding coefficients and the historical benchmark. S4: Based on the predicted channel state, data value assessment vector, and current communication link state, the optimal transmission decision for the current data batch is selected through a reinforcement learning policy network. The optimal transmission decision includes the coding scheme, compression level, and network interface. S5: Based on the optimal transmission decision, the data is encoded using a rate-distortion optimized hierarchical encoder, and the encoded data packet is uploaded to the cloud processing module. S6: Monitor the electrical behavior of the monitoring nodes based on the encoded data, and output status prediction and health assessment results.

[0032] This invention fundamentally solves the deterministic and real-time challenges of critical monitoring data transmission in complex industrial environments by achieving cross-layer deep perception and probabilistic prediction of channel interference characteristics at the edge, as well as unsupervised intelligent evaluation of the value of power data. Secondly, based on a dynamic transmission decision mechanism using multi-objective reinforcement learning, it achieves intelligent joint optimization of coding strategies, compression levels, and network interfaces. While ensuring the priority and reliable transmission of high-value data, it reduces overall network bandwidth pressure and shortens the latency between the edge and the cloud, thereby ensuring the deterministic transmission of critical monitoring data, the real-time nature of analysis and decision-making, and the timeliness required for subsequent predictive maintenance in the cloud.

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A cloud computing-based power monitoring system, characterized in that: The system includes: The data acquisition and processing module is used to acquire raw waveform data of power monitoring in real time, decompose the raw waveform data using wavelet packets, and construct a hybrid channel feature vector by combining communication link statistical indicators. The channel prediction module is used to input the hybrid channel feature vector into a channel prediction model constructed based on a neural process network to obtain a predicted distribution of future channel states. The coding evaluation module is used to perform sparse coding based on online dictionary learning on the raw waveform data and generate a data value evaluation vector based on the difference between the calculated sparse coding coefficients and the historical benchmark. The transmission control module is used to select the optimal transmission decision for the current data batch through a reinforcement learning policy network based on the predicted channel state, data value evaluation vector, and current communication link state. The optimal transmission decision includes the coding scheme, compression level, and network interface. The data transmission module encodes the data using a rate-distortion-optimized layered encoder based on the optimal transmission decision, and uploads the encoded data packets to the cloud processing module. The cloud processing module is used to monitor the electrical behavior of the monitoring nodes based on the encoded data and output status prediction and health assessment results.

2. The cloud computing-based power monitoring system according to claim 1, characterized in that: In the data acquisition and processing module, calculating the hybrid channel feature vector includes: The original waveform data is decomposed into N-level wavelet packets, and the energy proportion, information entropy, and distortion degree of the sub-band coefficients corresponding to the preset interference frequency band of each sub-band in at least the last level are calculated to obtain the time series characteristics. The time-series features and communication link statistics are used to generate a hybrid channel feature vector using a feature fusion algorithm.

3. The power monitoring system based on cloud computing according to claim 2, characterized in that: The formula for calculating the distortion degree of the sub-band coefficient is: ； In the formula, D represents the subband coefficient distortion degree, and M represents the set of subband indices covering the interference frequency band. It is the variance of the coefficients of subband j. It is the average variance under normal historical conditions.

4. The cloud computing-based power monitoring system according to claim 1, characterized in that: The channel prediction module includes a channel prediction model constructed using a variational conditional neural network. The model structure of the channel prediction model includes: The temporal context observation coding layer is used to obtain the channel observation set at historical moments. For each observation pair in the channel observation set, feature encoding is performed, and then the encoded features are globally aggregated to obtain the global context representation. A variational latent channel mode distribution layer is used to generate a latent probability distribution based on the global context representation and to sample latent variables from the latent probability distribution; The future channel quality probability prediction decoding layer is used to obtain the channel input set at future time, decode the channel input at each future time in combination with latent variables, and output the probability distribution parameters of the channel quality at the corresponding time. The probability distribution parameters are the predicted distribution of the future channel state.

5. The cloud computing-based power monitoring system according to claim 1, characterized in that: The encoding evaluation module includes a data preprocessing unit, a dictionary encoding unit, and an evaluation vector generation unit, which are connected in sequence via communication. The data preprocessing unit is used to segment the original waveform signal into analysis windows of fixed duration and perform denoising and baseline correction. The dictionary encoding unit is used to establish and update the dynamically adaptive feature dictionary, and to represent the preprocessed electrical waveform data as a sparse linear combination of atoms in the dictionary through a sparse encoding algorithm. The evaluation vector generation unit is used to analyze sparse coding results, quantify the degree of anomaly and importance of the current data window, and generate standardized data value evaluation vectors.

6. A cloud computing-based power monitoring system according to claim 5, characterized in that: The dictionary encoding unit is used to perform the following process: A feature dictionary is constructed, which includes predefined base atoms and adaptive atoms obtained through online learning. The base atoms are generated based on typical waveform features of the power system, and the adaptive atoms are dynamically generated based on actual data during operation. An online dictionary learning algorithm is used to periodically update the dictionary atoms using newly collected normal operating condition data; Sparse coding is performed using an orthogonal matching pursuit algorithm to obtain sparse coding coefficients. The sparse coding process aims to minimize a preset objective function to obtain sparse coefficients that can accurately represent the original waveform with the fewest dictionary atoms.

7. A cloud computing-based power monitoring system according to claim 6, characterized in that: The formula for calculating the objective function is as follows: ； In the formula, Let be the objective function. For error reconstruction term, The sparsity regularization term constrains the sparsity of the sparse coding coefficients using the L1 norm. Z is the dictionary matrix, and T is the total number of time windows. Let t be the vector of the t-th data window. Let be the sparse coding coefficients corresponding to the t-th data window. This is the regularization parameter.

8. A cloud computing-based power monitoring system according to claim 5, characterized in that: The evaluation vector generation unit is used to perform the following process: Based on sparsity, the sparsity is compared with the historical sparsity benchmark, and the relative rate of change is calculated as the sparsity mutation degree. Identify the rare atoms in the currently activated dictionary atoms and calculate the weighted activation energy of these rare atoms; Calculate the coefficient energy distribution entropy using the Shannon entropy formula; The calculated sparsity mutation degree, weighted activation energy, and coefficient energy distribution entropy are combined in sequence and normalized to obtain the value assessment vector.

9. A cloud computing-based power monitoring system according to claim 6, characterized in that: The dictionary encoding unit includes iteratively solving the objective function using the block coordinate descent method.

10. A cloud computing-based power monitoring method, applied to a cloud computing-based power monitoring system as described in claims 1-9, characterized in that: The method includes the following steps: S1: Real-time acquisition of raw waveform data for power monitoring, decomposition of raw waveform data using wavelet packets, and construction of hybrid channel feature vectors by combining communication link statistical indicators; S2: Input the hybrid channel feature vector into the channel prediction model constructed based on neural process network to obtain the predicted distribution of future channel states; S3: Perform sparse coding based on online dictionary learning on the original waveform data, and generate a data value assessment vector based on the difference between the calculated sparse coding coefficients and the historical benchmark. S4: Based on the predicted channel state, data value assessment vector, and current communication link state, the optimal transmission decision for the current data batch is selected through a reinforcement learning policy network. The optimal transmission decision includes the coding scheme, compression level, and network interface. S5: Based on the optimal transmission decision, the data is encoded using a rate-distortion optimized hierarchical encoder, and the encoded data packet is uploaded to the cloud processing module. S6: Monitor the electrical behavior of the monitoring nodes based on the encoded data, and output status prediction and health assessment results.

Images