Multi-device energy consumption monitoring and management method based on Internet of Things

Abstract

The invention relates to the technical field of energy management of the Internet of Things, and discloses a multi-device energy consumption monitoring and management method based on the Internet of Things. The method comprises the following steps: collecting multi-device energy consumption data through an intelligent sensing node, processing the data, dynamically dividing the data into low, medium and high energy consumption state intervals, and generating a multi-dimensional feature matrix; monitoring a state interval switching frequency in real time, calculating a moving average value and comparing the moving average value with a self-adaptive threshold value to trigger an abnormal mark; matching the feature matrix with a historical mode, extracting an abnormal feature component, and performing electrical parameter calibration; and finally, constructing an equipment energy consumption association network, executing reverse path search based on the calibrated features to position abnormal source nodes, and generating a regulation and control instruction. According to the invention, the transient abnormity of the equipment state can be effectively captured to realize early warning, the root equipment of the systematic energy consumption abnormity can be accurately positioned, and the monitoring accuracy and the management efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) energy management technology, specifically to a method for monitoring and managing the energy consumption of multiple devices based on the Internet of Things. Background Technology

[0002] Current equipment energy consumption monitoring systems generally rely on installing smart meters or sensors at the device end to collect basic parameters such as power, current, and voltage. These systems typically use fixed thresholds for over-limit alarms or employ simple statistical analysis methods to identify energy consumption points that significantly deviate from normal levels. For group monitoring of multiple devices, the common practice is to analyze the data of each device independently or perform simple summation calculations, lacking in-depth analysis of the interrelationships and influences between devices.

[0003] Existing technical solutions have shortcomings. Fixed threshold alarm mechanisms are insensitive to slowly occurring energy consumption drift or instantaneous, intermittent anomalies, easily leading to missed alarms. Analysis models based on individual devices cannot effectively address systemic failures where an anomaly in one device triggers a chain reaction affecting related devices. When anomalies manifest as frequent switching between different energy consumption levels rather than consistently exceeding a certain threshold, traditional static analysis methods struggle to capture this dynamic, unstable behavior. Regarding locating the root cause of anomalies, existing technologies often focus on detecting the anomaly itself, failing to quickly and accurately pinpoint the initial source device causing the entire energy consumption network anomaly at the system level, resulting in low maintenance efficiency and a lack of targeted control measures.

[0004] In complex multi-device operation scenarios, early signs of energy consumption anomalies often manifest as rapid oscillations in operating status rather than absolute exceedances of limits. Existing methods lack the specialized monitoring and analysis capabilities for this dynamic behavior, failing to issue early warnings before the fault escalates. Due to electrical connections or process logic coupling between devices, an anomaly in one node can propagate along a specific path. Conventional technologies do not construct such a relational network model, meaning that after an anomaly occurs, maintenance personnel still need to rely on experience to troubleshoot step by step, failing to achieve automatic and rapid location of the anomaly source and making it difficult to meet the precision and efficiency requirements of modern intelligent energy management. Summary of the Invention

[0005] The purpose of this invention is to provide a method for monitoring and managing the energy consumption of multiple devices based on the Internet of Things, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for monitoring and managing the energy consumption of multiple devices based on the Internet of Things, the method comprising: By continuously collecting power, current and voltage data from multiple devices through intelligent sensor nodes deployed on the device side, the collected data is synchronized in time and cleaned of outliers, and a standardized energy consumption data stream is output. Based on standardized energy consumption data streams, energy consumption state intervals are dynamically divided according to preset energy consumption thresholds, including low energy consumption intervals, medium energy consumption intervals, and high energy consumption intervals. Time-domain energy consumption features, frequency-domain energy consumption features, and statistical energy consumption features are extracted for each energy consumption state interval and fused to generate a multi-dimensional energy consumption feature matrix. The switching frequency between energy consumption state intervals is monitored in real time. By comparing the moving average of the switching frequency with an adaptive threshold, an abnormal state marking process is triggered. When the switching frequency is abnormal, the current energy consumption state interval is marked. The multi-dimensional energy consumption feature matrix is ​​matched with the device's historical energy consumption pattern library. Based on the matching results, the multi-dimensional energy consumption feature matrix of the normal energy consumption state interval is decomposed to extract abnormal energy consumption feature components. Based on the impedance and energy efficiency characteristics in the equipment electrical parameter library, the amplitude of the abnormal energy consumption characteristic components is calibrated and the phase is compensated to obtain the calibrated abnormal energy consumption characteristic components. Construct an equipment energy consumption correlation network, perform reverse path search in the equipment energy consumption correlation network based on the calibrated energy consumption anomaly feature components, locate the energy consumption anomaly source equipment node, calculate the optimization priority according to the node influence degree, and generate equipment control instructions.

[0007] Preferably, the time synchronization and outlier cleaning include: The power data, current data, and voltage data are timestamped separately, and the data with different sampling rates are aligned to a unified time axis using a linear interpolation method; Calculate the sliding window variance of the aligned data and remove data points whose variance exceeds three times the standard deviation. The minimum mean square error filter in the adaptive filtering algorithm is used to smooth the discarded data and generate a standardized energy consumption data stream.

[0008] Preferably, the extraction of time-domain energy consumption features, frequency-domain energy consumption features, and statistical energy consumption features includes: For the power data in each energy consumption state interval, the mean, variance and peak factor in the time domain are calculated as time-domain energy consumption characteristics. Perform a fast Fourier transform on the power data and extract the amplitude of the main frequency components as the frequency domain energy consumption characteristics. Skewness and kurtosis of statistical current data are used as statistical energy consumption characteristics; The time-domain energy consumption characteristics, frequency-domain energy consumption characteristics, and statistical energy consumption characteristics are combined row by row to form a multi-dimensional energy consumption characteristic matrix.

[0009] Preferably, the time-domain energy consumption characteristics also include the zero-crossing rate of the power data, the frequency-domain energy consumption characteristics also include the spectral centroid of the current data, and the statistical energy consumption characteristics also include the coefficient of variation of the voltage data.

[0010] Preferably, the process for triggering the abnormal state marking includes: Set a dynamic time window, count the number of times the energy consumption state interval switches within the window, and calculate the switching frequency; The adaptive threshold is updated using an exponentially weighted moving average method. When the switching frequency exceeds the adaptive threshold, an abnormal state flag is activated. Compare the cosine similarity of the multi-dimensional energy consumption feature matrices of adjacent energy consumption state intervals. If the similarity is lower than a preset threshold, the current energy consumption state interval is marked as abnormal.

[0011] Preferably, the cosine similarity of the multi-dimensional energy consumption feature matrices comparing adjacent energy consumption state intervals includes: Extract power feature vectors and current feature vectors from a multi-dimensional energy consumption feature matrix; Calculate the cosine of the angle between the power eigenvector and the current eigenvector in adjacent intervals, respectively. A weighted average of the cosine values ​​of the angle between power and current is used to obtain a comprehensive similarity index.

[0012] Preferably, the feature decomposition includes: Calculate the Euclidean distance between the multi-dimensional energy consumption feature matrix and the templates in the device's historical energy consumption pattern library, and use this distance as the similarity matching result; If the Euclidean distance is less than the matching threshold, principal component analysis is used to reduce the dimensionality of the multi-dimensional energy consumption feature matrix and extract the main components as energy consumption anomaly feature components. If the Euclidean distance is greater than or equal to the matching threshold, independent component analysis is used to separate out the abnormal features and generate energy consumption abnormal feature components.

[0013] Preferably, the amplitude calibration and phase compensation include: Based on the impedance frequency response in the equipment electrical parameter library, adjust the amplitude attenuation of the power characteristic in the abnormal energy consumption characteristic component; Phase alignment of current characteristics is performed based on the power factor in energy efficiency characteristics; The calibrated power and current characteristics are reintegrated with the voltage characteristics to output the calibrated energy consumption anomaly characteristic components.

[0014] Preferably, the reverse path search includes: Analyze the edge weights and node connectivity in the device energy consumption correlation network; Starting from the device node corresponding to the calibrated energy consumption anomaly feature component, traverse the upstream nodes in reverse order along the edge weights; Calculate the cumulative anomaly score for each upstream node, and locate the energy consumption anomaly source device node based on the score. The impact is calculated by combining node depth and anomaly score to determine optimization priorities.

[0015] Preferably, the size of the dynamic time window is adaptively adjusted according to the device type and running time, wherein for high-power devices, the time window is shortened to improve sensitivity, and for low-power devices, the time window is extended to smooth fluctuations.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This technology monitors the switching frequency between energy consumption state intervals in real time and triggers anomaly flags by comparing the calculated moving average with an adaptive threshold. It shifts the monitoring focus from static energy consumption values ​​to dynamic state transition behaviors. The adaptive threshold dynamically adjusts the judgment benchmark based on historical system operating data, avoiding misjudgments or missed judgments caused by fixed thresholds that are not adapted to changing operating conditions. The moving average processing of state switching frequencies smooths short-term fluctuations and highlights true trend anomalies. This technology effectively captures early fault symptoms such as transient instability and intermittent oscillations in equipment energy consumption, enabling early warning of potential risks and overcoming the limitation of conventional methods that are only effective for stable out-of-limit anomalies.

[0017] A device energy consumption correlation network is constructed, and a reverse path search is performed within this network based on abnormal feature components calibrated with electrical parameters. The correlation network can formally express the mutual influence relationships between devices. The reverse path search algorithm starts from the node where an anomaly is detected and works backward to deduce the most probable path of anomaly propagation, thereby locating the root cause device. Combined with node impact degree calculation, the critical nodes with the greatest impact on system stability can be identified. This technology achieves a leap from discovering anomalies to locating their root causes, accurately identifying the initial fault point that triggers systemic problems, providing a direct basis for targeted control and improving operational efficiency. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the IoT-based multi-device energy consumption monitoring and management method described in this invention. Figure 2 Flowchart for time synchronization and outlier cleaning; Figure 3 This is a flowchart for cosine similarity comparison. Figure 4 A graph showing the energy consumption state switching frequency and adaptive threshold monitoring. Figure 5 A comparison chart of the contribution rates of variance in the multi-dimensional energy consumption feature decomposition. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides a method for monitoring and managing the energy consumption of multiple devices based on the Internet of Things (IoT). The method includes: continuously collecting power, current, and voltage data from multiple devices using intelligent sensor nodes deployed on the devices; performing time synchronization and outlier cleaning on the collected data; and outputting a standardized energy consumption data stream. Based on the standardized energy consumption data stream, dynamically dividing energy consumption state intervals according to a preset energy consumption threshold, including low, medium, and high energy consumption intervals; extracting time-domain energy consumption features, frequency-domain energy consumption features, and statistical energy consumption features for each energy consumption state interval; and fusing them to generate a multi-dimensional energy consumption feature matrix. The method also includes real-time monitoring of the switching frequency between energy consumption state intervals and comparing the moving average of the switching frequency with an adaptive threshold to trigger... An abnormal state marking process is implemented, which identifies the current energy consumption state range when the switching frequency is abnormal. A multi-dimensional energy consumption feature matrix is ​​matched with the equipment's historical energy consumption pattern library for similarity analysis. Based on the matching results, the multi-dimensional energy consumption feature matrix of the normal energy consumption state range is decomposed to extract abnormal energy consumption feature components. Based on the impedance and energy efficiency characteristics in the equipment's electrical parameter library, the amplitude and phase of the abnormal energy consumption feature components are calibrated to obtain calibrated abnormal energy consumption feature components. An equipment energy consumption association network is constructed, and based on the calibrated abnormal energy consumption feature components, a reverse path search is performed in the network to locate the energy consumption anomaly source equipment node. The optimization priority is calculated based on the node's influence, and equipment control commands are generated.

[0021] Example 1: See Figure 2In practical implementation, the power, current, and voltage data collected by the intelligent sensing nodes are each appended with a high-precision timestamp, and the timestamp information is synchronized with Coordinated Universal Time (UTC). For data sequences with inconsistent sampling rates, such as power data at 1kHz and current data at 500Hz, a linear interpolation method is used to map all data points onto a unified millisecond-level time axis. This unified millisecond-level time axis is a virtual time reference line, with milliseconds as the smallest unit of time, starting at the beginning of the data acquisition cycle and ending at the end of the data acquisition cycle. The interval between each point on the time axis is strictly 1 millisecond. The raw data collected by the intelligent sensing nodes is timestamped, but the sampling time and sampling interval may differ for different physical quantities (power, current, voltage). The mapping process first defines the raw data sequence for the physical quantity to be interpolated, which consists of a set of discrete data pairs. For each target time point on the unified millisecond-level time axis, the algorithm searches for two actual sampling points immediately before and after the target time point in the raw data sequence of that physical quantity. Specifically, the timestamp of the preceding actual sampling time point is less than or equal to the target time point, while the timestamp of the following actual sampling time point is greater than the target time point. Based on the values ​​of these two actual sampling points, an estimated value for the target time point is calculated using linear interpolation. Linear interpolation utilizes the proportion of the time difference between the target time point and the preceding actual sampling time point to the total time interval between the two actual sampling time points, proportionally allocating the change in value within that time interval to estimate the value at the target time point. This process is performed independently for power data, current data, and voltage data, ultimately ensuring that these three physical quantities have a corresponding calculated value at every millisecond on a unified millisecond-level time axis, achieving strict alignment of multi-source data in the time dimension.

[0022] The linear interpolation method uses the values ​​of the two preceding and following actual sampling points to perform linear calculations for each missing time point to insert an estimated value, thereby achieving time synchronization alignment of multi-source data. In some embodiments, the width of the sliding window can be configured to 100 data points, and the variance of the data within each window is calculated. The variance calculation adopts a standard statistical method, namely the average of the sum of squared deviations of each data point from the mean of the data within the window. The overall standard deviation of the variance of the entire data sequence is calculated, and the window variance value is compared with three times the overall standard deviation. If the variance value of a certain window exceeds this threshold, all data points within that window are determined to be outliers and are removed. It can be understood that the minimum mean square error filter minimizes the mean square error between the output signal and the desired signal by iteratively adjusting the filter coefficients, thereby achieving smooth processing of the data after outlier removal and generating a smooth, standardized energy consumption data stream.

[0023] In practice, for power data segments that fall within a specific energy consumption range after preprocessing, the calculation of time-domain energy consumption characteristics is immediately initiated. The mean represents the average power level of the data segment, the variance reflects the degree of power fluctuation, and the peak factor characterizes the signal spike characteristics by the ratio of the peak value to the root mean square value. The zero-crossing rate is calculated as the number of times the statistical power data segment crosses the zero level. A Fast Fourier Transform is performed on the same power data segment to convert the time-domain signal into a frequency-domain representation, and the amplitudes corresponding to the three frequency components with the largest amplitudes are extracted as frequency-domain energy consumption characteristics. The calculation of the spectral centroid is based on the spectrum of the current data, representing the energy center of the frequency distribution. Statistical energy consumption characteristics calculate the skewness of the current data to measure the asymmetry of the data distribution and the kurtosis to measure the steepness of the distribution shape. The coefficient of variation of the voltage data is calculated by the ratio of the standard deviation to the mean, and is used to assess the relative volatility of the voltage. After all features are calculated, they are combined according to a predetermined row order, with the time-domain energy consumption characteristics occupying the first few rows of the matrix, the frequency-domain energy consumption characteristics in the middle, and the statistical energy consumption characteristics at the end, thus forming a multi-dimensional energy consumption characteristic matrix. The mathematical representation of the multidimensional energy consumption feature matrix is ​​as follows:

[0024] in: Represents a multi-dimensional energy consumption feature matrix. Represents the time-domain energy consumption feature vector. Represents the frequency domain energy consumption feature vector. This represents the statistical energy consumption feature vector, with the semicolon indicating row concatenation. It can be understood that the time-domain energy consumption feature vector contains values ​​such as mean, variance, peak factor, and zero-crossing rate; the frequency-domain energy consumption feature vector contains values ​​such as the amplitude of the main frequency components and the spectral centroid; and the statistical energy consumption feature vector contains values ​​such as skewness, kurtosis, and coefficient of variation. The structure of the multi-dimensional energy consumption feature matrix ensures that features from different sources have fixed positions in the matrix, facilitating subsequent similarity matching and feature decomposition operations. In some embodiments, the order of feature calculation and the matrix concatenation method can be adjusted according to the actual application scenario, but the core is to retain information from the time domain, frequency domain, and statistical dimensions.

[0025] Example 2: See Figure 3In practical implementation, setting a dynamic time window is the initial step in triggering the abnormal state marking process. The size of the dynamic time window is adaptively adjusted according to the device type and runtime. For high-power devices, the time window is shortened to improve sensitivity, while for low-power devices, the time window is extended to smooth fluctuations. Within the dynamic time window, the number of switching between energy consumption state intervals is counted, and the switching frequency is calculated as the number of switching times per unit time. The adaptive threshold is updated using an exponentially weighted moving average method. This method uses a smoothing factor to combine the current switching frequency with the adaptive threshold from the previous moment, thereby generating a new adaptive threshold value. When the switching frequency exceeds the current adaptive threshold, the abnormal state marking process is initiated. The cosine similarity of the multi-dimensional energy consumption feature matrices of adjacent energy consumption state intervals is compared. Power feature vectors and current feature vectors are extracted from the multi-dimensional energy consumption feature matrices, and the cosine value of the angle between the power feature vector and the current feature vector within adjacent intervals is calculated. The formula for calculating the cosine value of the angle is:

[0026] in: Represents the cosine value of the included angle. and The feature vectors representing adjacent intervals (such as power feature vectors or current feature vectors) are represented by a dot operation, where the dot operation represents the dot product of the vectors, and the double vertical lines represent the Euclidean norm of the vectors. A weighted average is calculated on the cosine values ​​of the angles between the power feature vectors and the current feature vectors, with the weights based on a preset importance of the feature vectors, to obtain a comprehensive similarity index. If the comprehensive similarity is lower than a preset threshold, the current energy consumption state interval is marked as abnormal.

[0027] In some embodiments, the adaptive adjustment of the dynamic time window can be fine-tuned based on the real-time load of the device, for example, further shortening the time window during peak operation periods of high-power devices. Optionally, the value of the smoothing factor can be dynamically adjusted based on the variance of historical switching frequencies to improve the responsiveness of the adaptive threshold. It is understood that the weight settings in the weighted averaging process can be optimized based on device type to ensure the accuracy of the comprehensive similarity index. In some embodiments, the calculation of the cosine of the included angle can be extended to other feature vectors such as voltage feature vectors, but the core remains power and current. Optionally, the value of the preset threshold can be dynamically updated based on historical operating data of the device using machine learning methods. It is understood that the triggering conditions for abnormal state marking can also be combined with other indicators such as abrupt changes in energy consumption data for comprehensive judgment. After the abnormal state marking is completed, the system adds a status identifier label to each energy consumption state interval, with the label value being normal or abnormal, and binds and stores the label with the corresponding multi-dimensional energy consumption feature matrix. Subsequent steps only perform similarity matching, feature decomposition, amplitude calibration, and phase compensation on the multi-dimensional energy consumption feature matrix of energy consumption state intervals labeled as abnormal. For energy consumption state intervals labeled as normal, the subsequent processing is skipped directly, and only the standardized energy consumption data stream is retained for updating the historical pattern library. This achieves accurate correlation between the abnormal labeling results and subsequent steps, avoiding invalid calculations.

[0028] See Figure 4 This diagram presents a core visualization achievement of multi-device energy consumption monitoring and management methods based on the Internet of Things (IoT), focusing on the dynamic monitoring of energy consumption state switching frequency and adaptive thresholds. The horizontal axis represents time, and the vertical axis represents switching frequency. A combination of solid black lines, dashed gray lines, and gray filled areas clearly illustrates the dynamic changes in energy consumption states. The technical value of this diagram lies in overcoming the insensitivity of traditional fixed thresholds to slow drift and intermittent anomalies. Through a combination of dynamic thresholds and frequency monitoring, it accurately captures the transient instability of device energy consumption, providing intuitive evidence for early fault warnings. This represents a key upgrade from static numerical monitoring to dynamic behavioral analysis, effectively supporting the timely identification of abnormal states and subsequent root cause localization in multi-device energy consumption networks.

[0029] Example 3: Before entering the feature decomposition step, the system first reads the status identifier tags bound to each energy consumption state interval. Only multi-dimensional energy consumption feature matrices with abnormal tags are selected for the similarity matching step; if a certain energy consumption state interval is labeled as normal, Euclidean distance calculation and subsequent feature decomposition operations are not performed, and it is directly determined that there are no abnormal energy consumption feature components in that interval, without the need for subsequent calibration and anomaly source localization. Through this filtering logic, it is ensured that all subsequent processing steps revolve around the anomaly labeling results, forming a complete logical chain of anomaly labeling and targeted processing. In specific implementation, the similarity matching process is achieved by calculating the Euclidean distance between the multi-dimensional energy consumption feature matrix and the pre-stored template matrix in the device's historical energy consumption pattern library. The formula for calculating the Euclidean distance is:

[0030] in: Represents Euclidean distance. The total number of dimensions of the feature vector. Represents the first element in the current multi-dimensional energy consumption feature matrix. 1 eigenvalue, The template matrix representing the historical energy consumption pattern of the device is the first one. Each eigenvalue. The calculated Euclidean distance. The similarity matching result is compared with a preset matching threshold. If the Euclidean distance is less than the matching threshold, the current multi-dimensional energy consumption feature matrix is ​​determined to be similar to the historical normal pattern. Principal component analysis (PCA) is then used to reduce the dimensionality of the multi-dimensional energy consumption feature matrix. PCA projects the original features onto a new orthogonal basis through linear transformation and selects the top features according to their variance contribution rate. Each principal component is used as an energy consumption anomaly feature component. If the Euclidean distance is greater than or equal to the matching threshold, it indicates that there is a significant difference between the current mode and the historical normal mode. In this case, independent component analysis is used to separate the anomaly features. Independent component analysis finds a linear transformation that makes the output components statistically independent, thereby separating the independent source signals from the observed signals and identifying the components related to abnormal operating conditions as energy consumption anomaly feature components.

[0031] In some embodiments, the matching threshold can be determined based on the statistical quantile of the Euclidean distance distribution of the normal pattern in the device's historical operating data. Optionally, the number of principal components retained in the principal component analysis. This can be automatically determined by the cumulative variance contribution rate exceeding a specific percentage (e.g., 85%). It is understood that the number of components separated in independent component analysis can be consistent with the original dimensions of the multi-dimensional energy consumption feature matrix to ensure information integrity. In some embodiments, the priority of using principal component analysis or independent component analysis can be preset for specific types of devices. Optionally, the calculation of Euclidean distance can consider weighting different feature dimensions to reflect their differences in importance. It is understood that the template matrix in the device historical energy consumption pattern library can store multiple versions according to different operating conditions of the device (e.g., light load, full load) to improve matching accuracy.

[0032] See Figure 5 This figure is the core visualization result of the eigenvalue decomposition process. The horizontal axis represents the feature components, and the vertical axis represents the variance contribution rate. Different gray-scale bars are used to present the differences in variance contribution between principal component analysis (PCA) and independent component analysis (ICA). The figure visually compares the characteristics of the two eigenvalue decomposition methods: PCA focuses more on variance interpretation, with the preceding components concentrating most of the variance; ICA, on the other hand, distributes the variance more evenly across components and has a greater advantage in separating non-Gaussian and independent features. This provides a basis for method selection when extracting energy consumption anomaly feature components. When the Euclidean distance between the multi-dimensional energy consumption feature matrix and the historical pattern library is less than the matching threshold, PCA is used for dimensionality reduction to extract the principal components; when the Euclidean distance is greater than or equal to the matching threshold, ICA is used to separate the anomaly features, thereby accurately extracting energy consumption anomaly feature components and providing crucial data support for subsequent anomaly source localization and equipment control.

[0033] Example 4: In specific implementation, the amplitude calibration process begins by querying the equipment electrical parameter database. This database stores the impedance frequency response data of various devices, which describes the impedance characteristics of the device at different frequencies. Based on the impedance frequency response, the power characteristics in the abnormal energy consumption component are adjusted by amplitude attenuation. The adjustment formula is as follows:

[0034] in: This represents the calibrated power characteristic value. Represents the original power characteristic value. This represents the standard impedance value at the reference frequency in the equipment's electrical parameter library. This represents the impedance value obtained from the impedance frequency response data at the actual operating frequency.

[0035] This adjustment compensates for power measurement deviations caused by frequency variations. Phase alignment is performed on the current characteristics, based on energy efficiency features in the device's electrical parameter library, which include power factor information. The power factor is used to calculate the phase difference compensation angle between the current and voltage. This compensation angle is obtained by applying an inverse cosine function to the power factor value. Subsequently, the current characteristic data is phase-rotated to align with the reference voltage waveform. The calibrated power characteristics and phase-aligned current characteristics are then reintegrated with the original voltage characteristics. Since the power, current, and voltage characteristics have all been mapped to a unified millisecond-level time axis using linear interpolation in the initial processing, each characteristic has a corresponding value at every millisecond on the time axis. Therefore, the first step in reintegration is to verify the strict alignment of these characteristic values ​​in the time dimension, ensuring that for any given millisecond, the power, current, and voltage characteristic values ​​originate from the same sampling time. The re-integration operation combines the calibrated power characteristic value, the phase-aligned current characteristic value, and the original voltage characteristic value at each time point into an ordered triplet. The order of the triplet is usually fixed as power, current, and voltage. These triplets are arranged in chronological order according to the time series to form a new multidimensional data sequence, namely the calibrated energy consumption anomaly characteristic component.

[0036] In some embodiments, the reintegration process may introduce a data scaling step to unify the dimensions of different features. For example, power, current, and voltage feature values ​​may be divided by their respective reference values ​​(e.g., rated power, rated current, and rated voltage), converting all feature values ​​into dimensionless relative values ​​to avoid biases in subsequent analysis due to differences in magnitude. Optionally, the scaled feature values ​​may be further converted to a logarithmic scale to compress the dynamic range of the data and improve numerical stability. This dimensionality processing helps to fairly compare the contributions of different features in the reverse path search. In some embodiments, the reintegrated, calibrated energy consumption anomaly feature components may be encapsulated in a data structure, such as a two-dimensional array, where each row corresponds to a millisecond time point and each column corresponds to a feature type, with the array's row indices aligned with the time point order of the unified millisecond-level time axis. This data structure may include an array of timestamps and device identifiers to facilitate data source tracking. This structured format facilitates direct reading and processing by subsequent algorithms. Optionally, the reintegration process may include consistency checks, such as calculating the theoretical relationship between power, current, and voltage characteristics at each time point (e.g., power should equal the cosine of current multiplied by voltage, but considering that calibration and phase alignment errors have been corrected, only a rough verification is performed here) to detect any anomalous inconsistencies that may occur during the integration process. Consistency checks help ensure the reliability of the reintegration and prevent data mistransmission. It can be understood that the final output of the reintegration is a calibrated energy consumption anomaly characteristic component, which, as a whole, retains the temporal relationship between power, current, and voltage characteristics, providing accurate input for subsequent reverse path searching in the device energy consumption correlation network. The voltage characteristic itself is not corrected but participates in the integration as a benchmark. The integration process ensures that the characteristics remain synchronized in the time series, ultimately outputting the calibrated energy consumption anomaly characteristic component.

[0037] In some embodiments, the equipment electrical parameter library can be organized in tabular form for easy and quick lookup. Optionally, the power factor value can be dynamically updated based on the real-time operating status of the equipment to improve phase alignment accuracy. It is understood that impedance value lookup can be performed by interpolating environmental parameters such as operating temperature to improve accuracy. Refer to Table 1, which shows a simplified equipment electrical parameter table: Table 1: Equipment Electrical Parameters Table Equipment Model Reference frequency (Hz) Standard impedance (Ω) Power factor M001 50 10.2 0.92 M002 60 8.5 0.88 M003 50 15.0 0.95 In some embodiments, the phase rotation operation can be implemented using complex multiplication in digital signal processing. Optionally, the re-integration process may include a normalization step for the eigenvalues ​​to ensure dimensional consistency. It is understood that the calibrated energy consumption anomaly eigencomponents will be used for subsequent reverse path searching, and their accuracy directly affects the reliability of anomaly source localization.

[0038] Example 5: The calibrated energy consumption anomaly feature components used in this step all originate from energy consumption state intervals marked as anomalous, i.e., target data obtained after screening, feature decomposition extraction, amplitude calibration, and phase compensation through anomaly marking processes. If the energy consumption state interval corresponding to a certain device node is not marked as anomalous, there is no corresponding calibrated energy consumption anomaly feature component, and this node does not participate in the reverse path search, thus clarifying the direct correlation between anomaly source location and the previous anomaly marking steps. In specific implementation, the parsing of the device energy consumption association network is a prerequisite for performing the reverse path search. The device energy consumption association network is represented by a graph structure, where nodes represent devices, and edges represent the energy consumption influence relationships between devices. The parsing process includes reading the weight value of each edge and the connection relationship of each node. The edge weights are usually set based on the power flow or control dependency relationships between devices. Starting from the device node corresponding to the calibrated energy consumption anomaly feature component, this node is marked as the starting node of the search, and the upstream nodes are traversed in reverse along the edge weight direction, i.e., the search is performed in the opposite direction to energy consumption propagation. During the traversal, each directly or indirectly connected upstream node is visited, and the cumulative anomaly score for each upstream node is calculated. The formula for calculating the cumulative anomaly score is as follows:

[0039] in: Representative node The cumulative abnormal score, Representative node Its own initial anomaly score (derived from the calibrated energy consumption anomaly feature component). Representative node The set of all downstream neighbor nodes, Represents the node Pointing to node edge weights, Representing downstream nodes The cumulative anomaly score is calculated. Based on the calculated cumulative anomaly score, all visited upstream nodes are sorted, and the node with the highest score is identified as the energy consumption anomaly source device node. The impact is calculated by combining node depth and anomaly score. Node depth refers to the minimum number of reverse edges traversed from the starting node to the current upstream node. Impact is a function of node depth and cumulative anomaly score; generally, shallower nodes with higher scores have greater impact. All relevant nodes are sorted according to their impact to determine optimization priorities.

[0040] In some embodiments, edge weights can be assigned based on the actual physical connection distance between devices or the logical control strength. Optionally, an initial anomaly score... Normalization can be applied to eliminate the influence of different device dimensions. It is understood that node depth calculation can employ a breadth-first search algorithm from graph theory. In some embodiments, the influence calculation function can be designed as a weighted form, for example, influence = cumulative anomaly score / (node ​​depth + 1). Optionally, the optimization priority can be determined by setting a minimum influence threshold, generating control instructions only for nodes exceeding the threshold. It is understood that the termination condition for reverse path search can be set when the increment of the cumulative anomaly score falls below a certain critical value or reaches a preset maximum search depth.

[0041] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring and managing the energy consumption of multiple devices based on the Internet of Things, characterized in that, The method includes: By continuously collecting power, current and voltage data from multiple devices through intelligent sensor nodes deployed on the device side, the collected data is synchronized in time and cleaned of outliers, and a standardized energy consumption data stream is output. Based on standardized energy consumption data streams, energy consumption state intervals are dynamically divided according to preset energy consumption thresholds, including low energy consumption intervals, medium energy consumption intervals, and high energy consumption intervals. Time-domain energy consumption features, frequency-domain energy consumption features, and statistical energy consumption features are extracted for each energy consumption state interval and fused to generate a multi-dimensional energy consumption feature matrix. The switching frequency between energy consumption state intervals is monitored in real time. By comparing the moving average of the switching frequency with an adaptive threshold, an abnormal state marking process is triggered. When the switching frequency is abnormal, the current energy consumption state interval is marked. The multi-dimensional energy consumption feature matrix is ​​matched with the device's historical energy consumption pattern library. Based on the matching results, the multi-dimensional energy consumption feature matrix of the normal energy consumption state interval is decomposed to extract abnormal energy consumption feature components. Based on the impedance and energy efficiency characteristics in the equipment electrical parameter library, the amplitude of the abnormal energy consumption characteristic components is calibrated and the phase is compensated to obtain the calibrated abnormal energy consumption characteristic components. Construct an equipment energy consumption correlation network, perform reverse path search in the equipment energy consumption correlation network based on the calibrated energy consumption anomaly feature components, locate the energy consumption anomaly source equipment node, calculate the optimization priority according to the node influence degree, and generate equipment control instructions.

2. The method for intelligent diagnosis and control of multi-device energy consumption based on the Internet of Things as described in claim 1, characterized in that, The time synchronization and outlier cleanup include: The power data, current data, and voltage data are timestamped separately, and the data with different sampling rates are aligned to a unified time axis using a linear interpolation method; Calculate the sliding window variance of the aligned data and remove data points whose variance exceeds three times the standard deviation. The minimum mean square error filter in the adaptive filtering algorithm is used to smooth the discarded data and generate a standardized energy consumption data stream.

3. The method for intelligent diagnosis and control of multi-device energy consumption based on the Internet of Things as described in claim 1, characterized in that, The extraction of time-domain energy consumption features, frequency-domain energy consumption features, and statistical energy consumption features includes: For the power data in each energy consumption state interval, the mean, variance and peak factor in the time domain are calculated as time-domain energy consumption characteristics. Perform a fast Fourier transform on the power data and extract the amplitude of the main frequency components as the frequency domain energy consumption characteristics. Skewness and kurtosis of statistical current data are used as statistical energy consumption characteristics; The time-domain energy consumption characteristics, frequency-domain energy consumption characteristics, and statistical energy consumption characteristics are combined row by row to form a multi-dimensional energy consumption characteristic matrix.

4. The method for intelligent diagnosis and control of multi-device energy consumption based on the Internet of Things as described in claim 3, characterized in that, The time-domain energy consumption characteristics also include the zero-crossing rate of power data, the frequency-domain energy consumption characteristics also include the spectral centroid of current data, and the statistical energy consumption characteristics also include the coefficient of variation of voltage data.

5. The method for intelligent diagnosis and control of multi-device energy consumption based on the Internet of Things as described in claim 1, characterized in that, The process for triggering the abnormal state marking includes: Set a dynamic time window, count the number of times the energy consumption state interval switches within the window, and calculate the switching frequency; The adaptive threshold is updated using an exponentially weighted moving average method. When the switching frequency exceeds the adaptive threshold, an abnormal state flag is activated. Compare the cosine similarity of the multi-dimensional energy consumption feature matrices of adjacent energy consumption state intervals. If the similarity is lower than a preset threshold, the current energy consumption state interval is marked as abnormal.

6. The method for intelligent diagnosis and control of multi-device energy consumption based on the Internet of Things as described in claim 5, characterized in that, The cosine similarity of the multi-dimensional energy consumption feature matrices comparing adjacent energy consumption state intervals includes: Extract power feature vectors and current feature vectors from a multi-dimensional energy consumption feature matrix; Calculate the cosine of the angle between the power eigenvector and the current eigenvector in adjacent intervals, respectively. A weighted average of the cosine values ​​of the angle between power and current is used to obtain a comprehensive similarity index.

7. The method for intelligent diagnosis and control of multi-device energy consumption based on the Internet of Things as described in claim 1, characterized in that, The feature decomposition includes: Calculate the Euclidean distance between the multi-dimensional energy consumption feature matrix and the templates in the device's historical energy consumption pattern library, and use this distance as the similarity matching result; If the Euclidean distance is less than the matching threshold, principal component analysis is used to reduce the dimensionality of the multi-dimensional energy consumption feature matrix and extract the main components as energy consumption anomaly feature components. If the Euclidean distance is greater than or equal to the matching threshold, independent component analysis is used to separate out the abnormal features and generate energy consumption abnormal feature components.

8. The method for intelligent diagnosis and control of multi-device energy consumption based on the Internet of Things as described in claim 1, characterized in that, The amplitude calibration and phase compensation include: Based on the impedance frequency response in the equipment electrical parameter library, adjust the amplitude attenuation of the power characteristic in the abnormal energy consumption characteristic component; Phase alignment of current characteristics is performed based on the power factor in energy efficiency characteristics; The calibrated power and current characteristics are reintegrated with the voltage characteristics to output the calibrated energy consumption anomaly characteristic components.

9. The method for intelligent diagnosis and control of multi-device energy consumption based on the Internet of Things as described in claim 1, characterized in that, The reverse path search includes: Analyze the edge weights and node connectivity in the device energy consumption correlation network; Starting from the device node corresponding to the calibrated energy consumption anomaly feature component, traverse the upstream nodes in reverse order along the edge weights; Calculate the cumulative anomaly score for each upstream node, and locate the energy consumption anomaly source device node based on the score. The impact is calculated by combining node depth and anomaly score to determine optimization priorities.

10. The method for intelligent diagnosis and control of multi-device energy consumption based on the Internet of Things as described in claim 5, characterized in that, The size of the dynamic time window is adaptively adjusted according to the device type and running time. For high-power devices, the time window is shortened to improve sensitivity, while for low-power devices, the time window is extended to smooth fluctuations.

Images