Real-time energy consumption analysis method based on artificial intelligence

Abstract

This invention belongs to the field of energy consumption analysis technology and provides a real-time energy consumption analysis method based on artificial intelligence. The method includes: spatiotemporal alignment and fusion of multi-source heterogeneous data to construct a comprehensive dataset; extracting data for the current time period from the comprehensive dataset; mapping the original data within the sliding window to a high-dimensional feature space using a KPCA model and performing dimensionality reduction to extract several nonlinear principal components; when an update is needed, using an incremental learning algorithm to update the kernel matrix and feature subspace of the KPCA model online, calculating the variance contribution rate of each nonlinear principal component; determining the key energy consumption influencing factors for the current time period based on the variance contribution rate of each nonlinear principal component, and identifying the main energy consumption patterns based at least on the numerical distribution, correlation, and changing trends of the key energy consumption influencing factors. This invention utilizes an improved PCA to more accurately analyze the main energy consumption patterns in ring spinning scenarios.

Description

Technical Field

[0001] This invention relates to the field of energy consumption analysis technology, and more specifically, to a real-time energy consumption analysis method based on artificial intelligence. Background Technology

[0002] The ring spinning process is lengthy, encompassing numerous steps such as opening and cleaning, carding, drawing, roving, spinning, and winding, involving a large number of energy-intensive equipment. Not only are main equipment like spinning frames and winding machines major energy consumers, but auxiliary equipment such as air conditioning systems and air compressors also account for a significant portion of energy consumption. Currently, while energy consumption data collection is possible, it is largely limited to simple measurement and statistics, resulting in severe data silos. It fails to deeply integrate and correlate data with equipment operating parameters, production process data (such as speed, draft ratio, temperature, and humidity), environmental data, and logistics status, making it impossible to accurately analyze the main energy consumption patterns.

[0003] Principal Component Analysis (PCA), a classic method for data dimensionality reduction and feature extraction, transforms multiple correlated variables into a few uncorrelated composite variables (principal components) through linear transformation, simplifying dataset structure. Theoretically, PCA can be used to analyze energy consumption data and identify major energy consumption patterns. However, traditional PCA is a linear, static, global modeling method, while spinning production is a dynamic process with frequent fluctuations in operating conditions (such as product variety changes, environmental temperature and humidity variations, and equipment performance degradation), causing changes in data feature distribution. Static PCA models quickly become ineffective, leading to decreased accuracy. Furthermore, the energy consumption relationships between equipment in a spinning system are complex, with numerous nonlinear couplings (such as the nonlinear relationship between air conditioning energy consumption and outdoor weather, and the heat dissipation of workshop equipment). Traditional linear PCA struggles to effectively extract these features, resulting in the loss of key information and analysis results that deviate from reality.

[0004] In summary, how to improve the traditional PCA to accurately analyze the main energy consumption patterns in ring spinning scenarios is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an artificial intelligence-based real-time energy consumption analysis method and a non-transient computer-readable storage medium.

[0006] This invention provides an artificial intelligence-based real-time energy consumption analysis method, comprising the following steps: real-time acquisition of multi-source heterogeneous data from the entire ring spinning process, including energy consumption data, equipment operating parameters, production process data, and environmental data; spatiotemporal alignment and fusion of the multi-source heterogeneous data to construct a comprehensive dataset; extraction of data for the current time period from the comprehensive dataset based on a sliding window of a set time length; mapping the original data within the sliding window to a high-dimensional feature space using a KPCA model and performing dimensionality reduction to extract several nonlinear principal components that can characterize the current main energy consumption pattern; when an update opportunity is reached, online updating of the kernel matrix and feature subspace of the KPCA model using an incremental learning algorithm, and calculation of the variance contribution rate of each nonlinear principal component; wherein, the update opportunity is determined based on the energy consumption pattern failure semantics analyzed from the new data within the sliding window; determination of key energy consumption influencing factors for the current time period based on the variance contribution rate of each nonlinear principal component, and identification of the main energy consumption pattern based at least on the numerical distribution, correlation, and changing trend of the key energy consumption influencing factors.

[0007] The present invention also provides a non-transient computer-readable storage medium having instructions embodied thereon that can be executed by one or more processors to perform a method as described in any of the preceding claims.

[0008] This invention accurately extracts the complex coupling characteristics between equipment and the environment in the ring spinning process through multi-source heterogeneous data fusion and KPCA nonlinear dimensionality reduction, overcoming the shortcomings of traditional linear methods in modeling nonlinear relationships. Furthermore, by employing a sliding window and incremental learning mechanism, the KPCA model can dynamically adapt to fluctuations in operating conditions, solving the problem of rapid failure of static models due to product changes and environmental variations. Finally, the data features are transformed into interpretable key influencing factors and energy consumption patterns, providing real-time and reliable decision-making basis for refined energy-saving control, significantly improving the accuracy, adaptability, and engineering practical value of energy consumption analysis. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a real-time energy consumption analysis method based on artificial intelligence disclosed in an embodiment of the present invention. Detailed Implementation

[0010] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0011] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0012] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0013] like Figure 1 As shown, this embodiment of the invention discloses an artificial intelligence-based real-time energy consumption analysis method 100, which includes the following steps: S10, real-time collection of multi-source heterogeneous data of the entire ring spinning process, including energy consumption data, equipment operating parameters, production process data and environmental data, and spatiotemporal alignment and fusion of multi-source heterogeneous data to construct a comprehensive dataset.

[0014] In this step, the solution of the present invention is applied to an energy consumption analysis server or edge computing terminal deployed in the ring spinning workshop. It communicates with various sensors, equipment control systems and production management systems (MES) through the Internet of Things interface, and receives and integrates massive amounts of data from all processes such as opening and cleaning, carding, drawing, roving, spinning and winding in real time.

[0015] Specifically, energy consumption data includes the instantaneous and cumulative values ​​of electrical energy consumption of the main equipment (such as spinning machines and winding machines) and auxiliary equipment (such as air conditioning systems, air compressors, and water pumps) in each process; equipment operating parameters include the real-time speed, load rate, and start / stop status of the equipment; production process data includes the process settings such as the variety being produced, yarn count, draft ratio, and twist; and environmental data includes the temperature and humidity inside and outside the workshop, dew point temperature, and the pressure and flow rate of the air compressor system.

[0016] Since the aforementioned data originates from different heterogeneous systems, their sampling frequencies, timestamp precision, and data formats vary. Therefore, spatiotemporal alignment and fusion processing are necessary. For example, using a unified millisecond-level timestamp as a benchmark, various time series data are resampled and interpolated to ensure that all data points are strictly synchronized in the time dimension. Simultaneously, based on the physical location and process logic relationships of equipment and sensors, data associations are established in the spatial dimension (e.g., associating the energy consumption of a spinning machine with the temperature and humidity data of its corresponding air conditioning supply area), ultimately forming a multi-dimensional, highly consistent comprehensive dataset.

[0017] S20, based on a sliding window of a set time length, extract data for the current time period from the comprehensive dataset; the KPCA model maps the original data within the sliding window to a high-dimensional feature space and performs dimensionality reduction to extract several nonlinear principal components that can characterize the current main energy consumption pattern.

[0018] In this step, considering that the spinning production process is continuous and the operating conditions fluctuate, the present invention employs a sliding window mechanism for data analysis. Specifically, based on a preset window length (e.g., 30 minutes or the time of one production batch), the latest data for the current period is dynamically extracted from the comprehensive dataset, ensuring that the analysis always focuses on the most recent production status and avoids interference from outdated data.

[0019] Subsequently, the KPCA (Kernel Principal Component Analysis) model processes the data within the window. Unlike traditional linear PCA, KPCA uses nonlinear kernel functions such as the Gaussian kernel (RBF) to map the original linearly inseparable multivariate data into a higher-dimensional feature space. In this higher-dimensional space, complex nonlinear coupling relationships between variables (such as the nonlinear relationship between air conditioning energy consumption and outdoor temperature, and equipment heat dissipation) become linearly separable.

[0020] Performing linear PCA in this high-dimensional space reduces its dimensionality and extracts several nonlinear principal components. These nonlinear principal components are a combination of the original variables, representing the feature combination that contributes the most variance to the data within the current sliding window, and can characterize the most important energy consumption patterns. For example, the first principal component represents the "production load and output" pattern, and the second principal component represents the "environmental temperature and humidity control" pattern.

[0021] S30, when the update timing is reached, the kernel matrix and feature subspace of the KPCA model are updated online using an incremental learning algorithm, and the variance contribution rate of each nonlinear principal component is calculated; wherein, the update timing is determined based on the energy consumption mode failure semantics derived from the analysis of new data within the sliding window.

[0022] In this step, the present invention does not use a fixed KPCA model, but introduces an online model update mechanism, specifically as follows: By continuously monitoring the data flowing into the sliding window, the differences between the new data and the main energy consumption patterns described by the current KPCA model are analyzed. When situations occur, such as product changes, drastic changes in ambient temperature and humidity, or a significant decrease in equipment efficiency, the feature distribution of the new data will drift, and the original model will be unable to accurately fit it, i.e., energy consumption pattern failure semantics will be generated. Once such semantics is identified (e.g., model reconstruction error continuously exceeds the limit), the model update procedure is automatically triggered.

[0023] During updates, an incremental learning algorithm is employed. This algorithm does not require saving all historical data and recalculating the entire model; instead, it efficiently updates the kernel matrix and eigenspace of the KPCA model online based on new window data. This allows the KPCA model to quickly adapt to new operating conditions while maintaining its analytical accuracy. After the model update, the variance contribution rate of each nonlinear principal component is recalculated to reflect the importance ranking of various energy consumption patterns under the new operating conditions.

[0024] S40. Determine the key energy consumption influencing factors for the current period based on the variance contribution rate of each nonlinear principal component, and identify the main energy consumption patterns based at least on the numerical distribution, correlation and changing trend of the key energy consumption influencing factors.

[0025] In this step, the principal components are first sorted according to their variance contribution rates calculated in step S30, and the top few nonlinear principal components with the highest contribution rates are selected as the key nonlinear principal components for the current period. For each key nonlinear principal component, by analyzing its load vector, it is possible to identify which original variables (such as "spinning machine speed", "return air humidity", and "air compressor pressure") have the greatest influence weight on the nonlinear principal component. These variables are the key energy consumption influencing factors.

[0026] Ultimately, by comprehensively examining the numerical distribution of these key factors (whether the current value is within the efficient range), their correlations (how the factors influence each other), and their trends (whether they are rising, falling, or fluctuating), the main energy consumption patterns that can be described in natural language or business terminology can be identified. For example, the output might state: "The current main energy consumption pattern is 'high-load production mode under high humidity conditions,' with the spinning machine's energy consumption accounting for 45%, and air conditioning dehumidification energy consumption increasing by 15% year-on-year due to the rise in outdoor dew point temperature." This output of the main energy consumption pattern serves as a basis for production management personnel to make energy-saving control decisions.

[0027] This invention accurately extracts the complex coupling characteristics between equipment and the environment in the ring spinning process through multi-source heterogeneous data fusion and KPCA nonlinear dimensionality reduction, overcoming the shortcomings of traditional linear methods in modeling nonlinear relationships. Furthermore, by employing a sliding window and incremental learning mechanism, the KPCA model can dynamically adapt to fluctuations in operating conditions, solving the problem of rapid failure of static models due to product changes and environmental variations. Finally, the data features are transformed into interpretable key influencing factors and energy consumption patterns, providing real-time and reliable decision-making basis for refined energy-saving control, significantly improving the accuracy, adaptability, and engineering practical value of energy consumption analysis.

[0028] As an example, based on a sliding window of a set time length, data for the current period is extracted from the comprehensive dataset, including: S201, real-time monitoring of production change signals in the comprehensive dataset, including at least one of product variety change signals, process parameter adjustment signals, and equipment start / stop status change signals.

[0029] In this step, the comprehensive dataset constructed in step S10 is continuously scanned to extract feature signals that can characterize significant changes in production status. Specifically, the production change signals mainly include: product variety change signals: by monitoring work order information from the production management system (MES) or equipment process parameter setting unit, changes in key information such as the yarn variety and count being produced are identified. For example, switching from producing 40-count pure cotton yarn to 80-count combed cotton yarn will directly lead to changes in parameters such as equipment speed and draft ratio, thereby affecting energy consumption patterns.

[0030] Process parameter adjustment signals: These are signals that monitor changes in the setpoints of key process parameters in the equipment control system in real time, such as the spindle speed and twist of a ring spinning frame, or the draft ratio of a roving frame, when these settings exceed preset thresholds. Even without changing the product type, such adjustments can alter the equipment's operating load and energy consumption characteristics.

[0031] Equipment start / stop status change signals: Monitor the operating status signals of major energy-consuming equipment (such as spinning machines, winding machines, and air compressors). The batch start / stop of one or more pieces of equipment (such as collective shutdown during shift handover or restart after equipment failure) will instantly change the total load of the workshop and the trend of changes in ambient temperature and humidity, thereby causing sudden changes in energy consumption patterns.

[0032] By monitoring one or more of the above signals, it is possible to determine in a timely and accurate manner whether production has entered a transitional operating condition.

[0033] S202, when a production change signal is detected, the set time length of the sliding window is automatically adjusted to the first preset time length; when no production change signal is detected and production is in a stable state, the set time length of the sliding window is automatically adjusted to the second preset time length, which is greater than the first preset time length.

[0034] In this step, the memory length of the KPCA model is adaptively optimized. The core principle is to analyze data of different time lengths for different production states in order to balance the sensitivity and stability of the model in the best way.

[0035] Specifically, when the aforementioned production change signal is detected, it indicates that the current energy consumption pattern is changing rapidly or is about to change. At this time, the set time length of the sliding window should be immediately adjusted to a shorter first preset duration (e.g., 5-15 minutes). A shorter time window means that the KPCA model focuses more on recent data, can quickly forget the old pattern before the change, and capture and learn the new energy consumption characteristics generated after the change, thereby improving the model's tracking sensitivity and response speed during the transition phase of operating conditions.

[0036] Conversely, when no production change signals are detected and the system determines that production is in a stable state, it indicates that the current energy consumption pattern has a certain degree of persistence and regularity. In this case, the sliding window's set time length is adjusted to a longer second preset duration (e.g., 30-60 minutes or longer). A longer time window contains richer historical data, allowing the KPCA model to learn based on more samples. The extracted principal components and energy consumption patterns are more statistically significant and stable, effectively smoothing out interference from short-term minor fluctuations and improving the reliability and signal-to-noise ratio of the analysis results.

[0037] Through this dynamic adjustment mechanism, this embodiment can intelligently adapt to changes in production rhythm, ensuring the optimality of analysis results whether in a rapidly changing transition period or a stable period of smooth operation.

[0038] As an example, the KPCA model maps the original data within the sliding window to a high-dimensional feature space and performs dimensionality reduction to extract several nonlinear principal components that can characterize the current main energy consumption patterns, including: S203, which calculates the kernel matrix of the original data within the sliding window based on the Gaussian radial basis kernel function, and nonlinearly maps the original data to a high-dimensional feature space.

[0039] This step uses kernel tricks to indirectly transform the original linearly inseparable spinning data into linearly separable features in a high-dimensional space without directly constructing a high-dimensional feature space (avoiding the problem of dimensional explosion). This allows for a precise characterization of the complex nonlinear relationship between energy consumption and process and environmental parameters (such as the stepwise increase in air conditioning energy consumption with rising outdoor temperature and humidity, and the nonlinear increase in spinning machine energy consumption with increasing spindle speed).

[0040] Ring spinning data is characterized by strong local correlations and complex global distributions. For example, the energy consumption is relatively small when the workshop temperature is between 25-28℃, but the energy consumption of air conditioning and dehumidification increases sharply above 28℃. This locally sensitive nonlinear relationship needs to be captured by a kernel function with local response characteristics. The expression for the Gaussian radial basis function (RBF kernel) is: In the formula, For any two original data samples within the sliding window, For Euclidean distance, This is the kernel width parameter.

[0041] The core advantages of the Gaussian radial basis kernel function are: (1) Local sensitivity: When the difference between two samples (such as the current energy consumption and process parameters and the historical energy consumption and process parameters) is small (the Euclidean distance is close), the kernel function value is close to 1, indicating that the two are closely related in the high-dimensional space; when the difference is large, the kernel function value is close to 0, the correlation is low, and it can accurately capture the characteristics of energy consumption mutation caused by small changes in parameters in spinning data.

[0042] (2) No dimensional limitations: It can adapt to high-dimensional data that integrates energy consumption, equipment operation, process and environment within the sliding window. For example, the sample data includes parameters of 10+ dimensions such as spinning machine energy consumption, spindle speed, draft ratio, workshop temperature, and air compressor pressure, without the need to simplify the original data dimensions.

[0043] The process of constructing the kernel matrix is ​​as follows: Assuming the sliding window contains n original data samples (each sample is a set of synchronously collected multi-source data), the kernel matrix K is a... A symmetric matrix, where the element in the i-th row and j-th column is... That is, the result calculated using the Gaussian radial basis kernel function mentioned above. and The inner product in a high-dimensional feature space. For example, when analyzing the relationship between the energy consumption of a spinning machine and the spindle speed and workshop humidity, the original data may show a nonlinear characteristic of linear correlation at low spindle speeds and accelerated energy consumption at high spindle speeds. Through kernel matrix calculation, this nonlinear relationship can be transformed into a linear correlation in a high-dimensional space.

[0044] S204, the kernel matrix is ​​centered and its eigenvalues ​​and eigenvectors are solved. The eigenvalues ​​are sorted from largest to smallest, and the eigenvectors corresponding to the first k largest eigenvalues ​​are selected to form the projection matrix.

[0045] This step begins with kernel matrix centering. Since KPCA indirectly maps to a high-dimensional space through kernel functions, the original kernel matrix cannot guarantee that the high-dimensional data will have a zero mean (i.e., the data distribution center deviates from the origin). For example, a spinning workshop has fixed baseline values ​​such as equipment standby energy consumption and basic air conditioning supply energy consumption. Without correction, subsequent analysis might misinterpret the shift in these baseline values ​​as changes in energy consumption patterns. Therefore, it is necessary to use a formula... , ( for The kernel matrix is ​​then centered (using a matrix of all ones). It can effectively eliminate the influence of fixed benchmarks such as equipment standby power consumption and default temperature and humidity in the workshop, allowing subsequent analysis to focus on the dynamic energy consumption differences that change with production conditions (such as product changes and process adjustments).

[0046] Next, the eigenvalues ​​and eigenvectors are solved. The centered kernel matrix... Essentially, this reflects the covariance relationship of spinning data in high-dimensional space, and the eigenvalues ​​are solved. With corresponding feature vectors This is the key to extracting the core features of energy consumption.

[0047] Here, the magnitude of the eigenvalue represents the magnitude of the data variance along the corresponding eigenvector direction: in a spinning scenario, the larger the variance, the better the combination of energy consumption features along that direction can distinguish different energy consumption modes. For example, the largest eigenvalue... The corresponding direction may represent energy consumption fluctuations caused by changes in production load (spindle speed, output) (large variance, large amount of information); minimum eigenvalue The corresponding direction may represent energy consumption noise caused by minor airflow disturbances in the workshop (small variance, small amount of information). The eigenvector is the direction with the largest data variance in high-dimensional space. Each eigenvector corresponds to a weighted combination of a set of original data samples. The core parameters associated with this direction (such as energy consumption, process, and environmental parameters) can be interpreted through the weights.

[0048] Next, the projection matrix is ​​constructed. Following the principle of preserving core information and eliminating noise, the eigenvectors corresponding to the top k largest eigenvalues ​​(typically requiring the cumulative variance contribution rate of the top k eigenvalues ​​to be ≥85% to ensure the retention of the vast majority of effective energy consumption information) are selected. ,constitute projection matrix The k value is set to 3-5 (adjusted based on actual data) to cover core energy consumption patterns such as production load-driven, environmental control-driven, and equipment efficiency-driven energy consumption, while avoiding excessive dimensions that could complicate the analysis.

[0049] S205, the centered kernel matrix is ​​projected onto the projection matrix to obtain several nonlinear principal components after dimensionality reduction.

[0050] In this step, the centered kernel matrix With projection matrix Perform matrix multiplication, that is ,get matrix Each column of the matrix This refers to the nonlinear principal components after dimensionality reduction. Among them, It is the first nonlinear principal component, corresponding to the largest eigenvalue, which contains the most energy consumption information; It is the k-th nonlinear principal component, containing less energy consumption information.

[0051] Unlike the linear combination of traditional linear PCA, the nonlinear principal components of KPCA are nonlinear combinations of the original multi-source data, which can accurately reflect the complex relationships in the spinning system. For example: the first nonlinear principal component. The formula "spinning machine energy consumption × 0.6 + spindle speed × 0.3 + workshop temperature × 0.1" (weights calculated using eigenvectors) represents the energy consumption pattern dominated by production load. When... An increase in the value indicates an increase in the current production load and an increase in the energy consumption of the spinning machine.

[0052] Second nonlinear principal component The formula "Air conditioning energy consumption × 0.5 + Outdoor humidity × 0.3 + Dew point temperature × 0.2" represents an energy consumption mode dominated by environmental control. When... An increase in the value indicates that the current air conditioning system is consuming more energy to maintain the temperature and humidity in the workshop.

[0053] As an example, an incremental learning algorithm is used to update the kernel matrix and feature subspace of the KPCA model online, including: S401, when a new batch of data flows into the sliding window, the rank-one update algorithm is used to quickly calculate the expanded new kernel matrix based on the original kernel matrix and the new batch of data.

[0054] In this step, ring spinning employs a sliding window mechanism (e.g., a 30-minute window), with each new data inflow being a small batch of samples (e.g., data collected over 1-5 minutes, denoted as m new samples). , (This represents the original number of samples), which falls under the low-rank incremental scenario. The core advantage of the rank-one update algorithm is that it reduces the complexity of matrix expansion from... Down to Only the kernel function values ​​of the original sample and the new sample need to be calculated, without having to repeatedly calculate the kernel function values ​​between the original samples (the original kernel matrix has already been stored), which can greatly reduce the amount of computation.

[0055] The construction process of the new kernel matrix is ​​as follows: (1) Define the data and the original kernel matrix: Let the original kernel matrix composed of n samples in the original sliding window be . The m new samples flowing into the sliding window are .

[0056] (2) Calculate the cross-kernel matrix: Calculate the cross-kernel matrix between the original sample and the new sample using the Gaussian radial basis kernel function. ,element ; and the kernel matrix between new samples ,element .

[0057] (3) Concatenation and expansion into a new kernel matrix: Following the logic of original sample-original sample, original sample-new sample, new sample-original sample, and new sample-new sample, the matrix is ​​concatenated into the expanded new kernel matrix. , dimension .

[0058] For example, when a workshop switches from producing 40-count pure cotton yarn to producing 80-count combed cotton yarn, triggering a new data inflow, the traditional method would require recalculating the kernel matrix of all 30-minute window data (approximately 1800 millisecond-level samples, a huge computational burden). However, the rank-one update algorithm only needs to calculate the cross kernel function value between the original samples in the first 25 minutes and the new samples in the last 5 minutes, reducing the computation time by more than 80%. This ensures that the model can complete the kernel matrix expansion within 10 seconds after the change in operating conditions, meeting the needs of real-time analysis.

[0059] S402, the new kernel matrix is ​​incrementally centered, and an incremental eigenspace update algorithm is used to approximately solve the eigenvalues ​​and eigenvectors of the new kernel matrix.

[0060] This step avoids the need for full centering and full eigenvalue solving of the expanded new kernel matrix (traditional methods have high computational complexity). Incremental operations are used to correct the kernel matrix and update the feature subspace, ensuring that the model can quickly adapt to the feature distribution of new data (such as energy consumption feature drift after variety change).

[0061] Specifically, this involves utilizing the centered intermediate results of the original kernel matrix (such as the mean vector of the original samples). Only the mean of the new sample is calculated. The correlation term with the original sample mean is determined by the formula. Incremental correction is achieved. Among them, This is the incremental correction term calculated based on the mean of the old and new samples.

[0062] Through the above incremental centralization, the benchmark offset of basic energy consumption of new equipment and basic energy consumption of air conditioning under new operating conditions in new samples can be quickly eliminated, ensuring that the kernel matrix after centralization still focuses on dynamic energy consumption differences. For example, the standby energy consumption of 80-count cotton yarn spinning machine in the new sample is higher than that of 40-count machine, and incremental centralization can correct this benchmark difference in real time.

[0063] Next, the incremental eigenspace update algorithm is as follows: Based on the original eigenspace (the first k eigenvectors of the original kernel matrix) and the feature perturbation brought by the new data, an incremental version of the power iteration method or the Lanczos algorithm is used. This only fine-tunes the original eigenspace, rather than solving for all eigenvalues ​​and eigenvectors of the new kernel matrix. The specific logic is: calculate the new kernel matrix... For the original eigenvector The disturbance effect is used to obtain a temporary vector. ;right Normalization is performed to obtain the updated feature vector. ;based on Calculate the corresponding new eigenvalues .

[0064] For example, when the temperature and humidity in the workshop environment suddenly rise (new data feature drift), the original feature vector cannot accurately represent the pattern of a surge in air conditioning energy consumption. The incremental algorithm only needs to update the feature vector and feature value through 3-5 iterations, reducing the calculation time from minutes to seconds for full solution, ensuring that the changes in energy consumption patterns dominated by environmental regulation can be captured in real time.

[0065] S403, based on the updated eigenvalues, redetermine the number k of principal components to be retained, and update the projection matrix.

[0066] This step dynamically adjusts the number of principal components k to be retained based on the updated eigenvalue distribution, ensuring that the projection matrix always focuses on the energy consumption mode with the most information under the current operating conditions, thus avoiding redundancy of principal components or loss of key information due to changes in operating conditions.

[0067] Specifically, calculate all updated eigenvalues The cumulative variance contribution rate, i.e. Select the one that satisfies The minimum value of k is used as the new number of principal components. Understandably, the threshold can be adjusted according to the precision requirements of the spinning scenario; for example, in high-precision spinning, the threshold can be adjusted from 85% to 90%.

[0068] For example, when production switches to complex varieties (such as 100-count high-count yarn), process parameters (draft ratio, twist) change more frequently, increasing the dimensions of energy consumption characteristics. The updated characteristic value may show a situation where the cumulative contribution rate of the first four characteristic values ​​reaches 85%. In this case, k is adjusted from 3 to 4 to ensure coverage of additional energy consumption patterns caused by equipment precision loss. When production enters a stable period (such as continuous production of the same variety for 8 hours), energy consumption characteristics tend to be simpler, and the updated characteristic value may show a cumulative contribution rate of the first two characteristic values ​​reaching 90%. In this case, k is adjusted from 3 to 2 to eliminate redundant minor perturbation principal components and improve analysis efficiency.

[0069] Next, update the projection matrix: select the eigenvectors corresponding to the k largest eigenvalues ​​after the update. Replace the original projection matrix Construct a new projection matrix (dimension is) ).

[0070] Updated projection matrix It can be directly used in the subsequent step S40 to identify key energy consumption influencing factors. For example, when k is adjusted to 4 and the energy consumption increases due to wear of the spinning machine rollers corresponding to the 4th principal component, the projection matrix can include this feature in the analysis to accurately locate the key influencing factor of roller speed deviation, providing a basis for equipment maintenance decisions and avoiding the omission of key energy consumption problems caused by the traditional fixed k value.

[0071] As an example, the timing of the update is determined based on the energy consumption pattern failure semantics derived from the analysis of new data within the sliding window, including: real-time calculation of the squared prediction error (SPE) statistic of the newly flowing sample data into the sliding window based on the current KPCA model; and calculation of dynamic control limits for the SPE statistic based on historical window data.

[0072] In this step, the KPCA model maps the high-dimensional original data to low-dimensional principal components using a projection matrix (preserving the main energy consumption information). Simultaneously, it generates residuals, which are the information in the original data not covered by the principal components (small noise under normal operating conditions, significant bias when the model fails). The SPE statistic is essentially a quantification of these residuals, calculated using the following formula: ;in, For a single sample of data newly flowing into the sliding window (a high-dimensional vector that integrates energy consumption, process, and environmental parameters). To be The reconstructed samples are obtained by backprojecting the principal components of the current KPCA model back into the original data space. The SPE value is the square of the Euclidean distance. The smaller the SPE value, the closer the new sample is to the normal pattern of the current model (the reconstructed sample is less different from the original sample); the larger the SPE value, the further the new sample deviates from the normal pattern (the reconstructed sample is more different from the original sample), which may indicate that the energy consumption pattern has failed.

[0073] In ring spinning, new sample data consists of multi-source heterogeneous data acquired at the millisecond level (e.g., a vector composed of "spinning machine energy consumption 2.5kW + spindle speed 18000r / min + workshop temperature 27℃ + air compressor pressure 0.6MPa"). During calculation, the new sample must first be projected onto the current KPCA model to obtain the corresponding nonlinear principal components (e.g., the first three principal components). Then, the sample is reconstructed by inversely calculating the transpose of the principal components and the projection matrix. Finally, the squared error between the two is calculated as the SPE value. For example, when a sudden air conditioning failure in the workshop causes a sharp increase in humidity, the deviation between "air conditioning energy consumption of 3.2kW + workshop humidity of 75%" in the new sample and "air conditioning energy consumption of 2.0kW under normal humidity" in the current model is significant. At this time, the SPE value will rise sharply from the historical stable value (such as 50) to more than 150, which directly reflects the deviation of the model.

[0074] Understandably, since spinning production is a continuous and dynamic process, new sample data continuously flows into the sliding window (e.g., one sample per second). The SPE statistic needs to be calculated in real-time upon sample inflow (with calculation time controlled within 100ms) to avoid delays in semantic determination of failure due to computational latency. Deploying lightweight computing modules on edge computing terminals allows direct connection to sensor data streams, ensuring that the SPE statistic is generated synchronously with sample acquisition.

[0075] Meanwhile, based on the idea of ​​statistical process control (SPC), the dynamic control limits are calculated in real time using the SPE statistics data of the historical sliding window that are judged to be normal working conditions. The specific steps are as follows: (1) From the most recent M historical sliding windows (M ranges from 10 to 20, such as 10 30-minute windows), the window data that have not triggered failure semantics and contain stable production signals (such as no product change or no process adjustment) are selected, and the SPE values ​​of all samples are extracted to form a normal SPE dataset.

[0076] (2) Fit the probability distribution of the normal SPE dataset by kernel density estimation (SPE data in spinning scenarios usually follow a non-normal distribution, and kernel density estimation is more accurate than the traditional normal distribution assumption). Take the 95th or 99th quantile of the distribution as the current dynamic control limit (e.g., the 95th quantile represents that the probability of the SPE value exceeding the limit under normal working conditions is only 5%, and exceeding it is highly likely to be abnormal).

[0077] (3) For each new normal operating condition window, its SPE data is included in the historical dataset, and the dynamic control limit is recalculated to realize the dynamic adjustment of the control limit as the normal energy consumption mode slowly drifts (such as the energy consumption benchmark slowly rising due to equipment aging).

[0078] For example, when production switches from "40-count pure cotton yarn" (the 95th percentile of the normal SPE dataset is 60) to "80-count combed cotton yarn" (the equipment speed is higher, energy consumption fluctuates more, and the 95th percentile of the normal SPE dataset rises to 85), the dynamic control limit will automatically adjust from 60 to 85. If a fixed threshold of 70 is used, SPE=75 will be mistakenly judged as a failure when producing 40-count yarn, and SPE=80 will be missed as normal when producing 80-count yarn. The dynamic control limit can completely avoid such problems and ensure the consistency of the judgment standard under different operating conditions.

[0079] Understandably, to avoid deviations in dynamic control limit calculations due to insufficient historical data (such as during the first production run of a new workshop or new product), industry benchmark data can be introduced as an initial supplement. For example, normal SPE distribution data from the same industry producing the same product can be imported as initial dynamic control limits. After accumulating 3-5 local normal window data, the calculation can be switched to pure local data to ensure accuracy in the initial production phase.

[0080] When the SPE statistic is detected to exceed the dynamic control limit for a preset number of consecutive times, an energy consumption mode failure semantic is determined to be generated, and the KPCA model update program is triggered according to the energy consumption mode failure semantic.

[0081] In this step, the ring spinning workshop experiences numerous short-term random disturbances. For example, a momentary drop in grid voltage can cause the spinning machine's energy consumption to briefly rise to 3.0kW (normally 2.5kW). At this time, the SPE value of a single sample may exceed the dynamic control limit, but this disturbance only lasts for 1-2 samples and does not indicate a true failure of the energy consumption pattern. If an update is triggered for only a single instance of exceeding the limit, it will lead to frequent model reconstruction (e.g., 5-6 updates per hour), which not only wastes computational resources but may also reduce the model's generalization ability due to over-adaptation to short-term disturbances.

[0082] By using the above-mentioned judgment logic that continuously exceeds the preset number of times, such interference can be effectively filtered out, ensuring that an update is triggered only when the deviation persists (i.e., the mode truly fails).

[0083] Specifically, the preset number of times needs to be determined based on the sampling frequency of the spinning data and the rate of change of operating conditions, and is usually set to 3-5 times (e.g., when the sampling frequency is 1 second / sample, 3 consecutive exceedances are considered a failure, and the corresponding deviation lasts for 3 seconds, which is sufficient to eliminate instantaneous disturbances). For example, when the energy consumption pattern changes due to product variety changes in the workshop, the SPE value will continuously exceed the control limit (e.g., 10 consecutive samples exceed the limit). In this case, reaching the preset number of 3 times is sufficient to determine a failure and trigger an update in time. It can be understood that when the condition of "N consecutive times (N=3-5) SPE statistic > dynamic control limit" is met, an energy consumption pattern failure semantic is formally determined. The physical meaning of this semantic is: the energy consumption pattern represented by the current KPCA model can no longer adapt to the new data, and the model needs to be updated through incremental learning to match the new energy consumption pattern.

[0084] Once an energy consumption mode failure semantic is determined, a trigger signal is automatically sent to the KPCA model incremental update module to initiate the kernel matrix expansion-feature subspace update-projection matrix reconstruction process, so as to ensure that the model can quickly adapt to the new energy consumption mode.

[0085] As an example, the dynamic control limit for calculating SPE statistics based on historical window data includes: obtaining the SPE statistics dataset of the most recent M historical sliding windows, filtering out the SPE data under stable production conditions to form a local normal SPE dataset; when the number of samples in the local normal SPE dataset is lower than a preset threshold, introducing industry benchmark SPE data as a supplement to form a hybrid SPE dataset.

[0086] In this step, a pool of SPE statistics data containing the most recent M historical sliding windows (M ranging from 10 to 20 windows) is continuously maintained. From this data pool, by querying work order status and equipment process parameter change records in the Production Management System (MES), time periods marked as stable production conditions (such as no product changes, no process parameter adjustments, and continuous equipment operation) are selected, and all SPE statistics corresponding to these time periods are extracted to form a clean local normal SPE dataset.

[0087] The sample size N_local of the local normal SPE dataset is monitored in real time. When N_local is lower than the preset threshold N_min (e.g., N_min = 150 samples, approximately equivalent to 1-2 complete stable production windows), the local normal SPE dataset is deemed insufficient for reliable statistical estimation. In this case, industry benchmark SPE data (including statistical characteristics such as mean, variance, and distribution quantiles) matching the current production product and equipment model is retrieved from the cloud-based industry database and mixed with the local normal SPE dataset to form a hybrid SPE dataset. This compensates for the insufficient local data volume and ensures the stability of kernel density estimation.

[0088] Based on the hybrid SPE dataset, the kernel density estimation method is used to calculate the dynamic control limits; and as local normal SPE data continues to accumulate, the weight of the industry benchmark data is dynamically adjusted. When the amount of local normal SPE data reaches a sufficient threshold, the dynamic control limits are calculated entirely using local data.

[0089] In this step, a Gaussian kernel function is used to estimate the kernel density of the mixed SPE dataset. The bandwidth parameter h can be automatically calculated using Silverman's rule. Finally, the 99th percentile of the estimated distribution is taken as the dynamic control limit Q_α currently in use.

[0090] Simultaneously, weighting factors W_local (local data weight, initial value is N_local / (N_local+N_industry)) and W_industry (industry data weight, initial value is 1-W_local) are introduced. Here, N_local is the number of samples in the local normal SPE dataset, and N_industry is the number of samples in normal SPE data obtained from external industry databases that match the current production situation (e.g., the same yarn type, similar equipment model).

[0091] As production continues, local normal SPE data accumulates, and N_local gradually increases. The weight of local data is dynamically increased according to W_local=min(1,N_local / N_sufficient), while the weight of industry data is correspondingly decreased (W_industry=1-W_local), and the weighted kernel density estimate is recalculated. When N_local≥N_sufficient, W_local=1, W_industry=0, completely discarding industry benchmark data and calculating dynamic control limits solely based on pure local data.

[0092] As an example, the sufficient threshold is determined at least based on the complexity of the product mix and the frequency of changes in operating conditions.

[0093] In this embodiment, the sufficiency threshold N_sufficient is not a fixed value, but a variable intelligently set based on the dynamic characteristics of the current production environment. Specifically, the sufficiency threshold is determined based on the following factors: First, the complexity of the product varieties is considered. A variety complexity classification database is maintained, and different yarn varieties are divided into multiple complexity levels (such as ordinary, medium, and advanced) according to their production process difficulty and energy consumption fluctuation range. For example, producing 40-count pure cotton yarn belongs to the ordinary level (complexity coefficient C=1.0), and the threshold is set to N_base=500; producing 80-count combed cotton yarn belongs to the medium level (complexity coefficient C=1.5), and the threshold is adjusted to N_base×C=750; producing blended yarn of 100 count or higher belongs to the advanced level (complexity coefficient C=2.0), and the threshold is set to N_base×C=1000. The higher the complexity, the more complex the energy consumption pattern, requiring more local data to accurately capture its statistical characteristics.

[0094] Secondly, consider the frequency of operating condition changes. By monitoring indicators such as the frequency of product changes and the number of process adjustments in historical production data, assess the stability of the current production environment. When the frequency of operating condition changes is high (e.g., more than 3 product changes per day), set the frequency coefficient F=0.8 and lower the sufficient threshold requirement (e.g., final threshold = N_base×C×F) to ensure that model localization can be completed within a limited stable production time. When the frequency of operating condition changes is low (e.g., 1 product change per week), set the frequency coefficient F=1.2 and increase the threshold requirement to obtain a more accurate localized model.

[0095] Ultimately, the sufficiency threshold N_sufficient is determined by adjusting the base value N_base using the complexity coefficient C and the frequency coefficient F, i.e., N_sufficient = N_base × C × F.

[0096] Through this dynamic adjustment mechanism, the optimal balance between model localization speed and accuracy can be achieved while ensuring statistical reliability.

[0097] As an example, the step of determining the key energy consumption influencing factors for the current period based on the variance contribution rate of each nonlinear principal component includes: S401, sorting the nonlinear principal components from largest to smallest according to their variance contribution rates, and selecting the top k nonlinear principal components as key principal components; for each key principal component, analyzing the weight coefficients of each original variable in its load vector, and selecting the original variables whose absolute values ​​of the weight coefficients are greater than a set threshold as key energy consumption influencing factors.

[0098] In this step, the variance contribution rate of each nonlinear principal component is calculated and sorted in descending order. The variance contribution rate characterizes the degree of variation in the original data captured by each nonlinear principal component; a larger contribution rate indicates that the energy consumption pattern represented by that nonlinear principal component is more important. K nonlinear principal components with a cumulative variance contribution rate reaching a preset requirement (e.g., ≥85%) are selected as key principal components. These key principal components collectively characterize the most dominant energy consumption pattern within the current sliding window period. For example, the first principal component represents the production load-dominated pattern, and the second principal component represents the environmental regulation-dominated pattern.

[0099] For each key principal component, its corresponding load vector is analyzed in depth. This load vector is defined in a high-dimensional feature space, where each element corresponds to a weighting coefficient of the contribution of an original monitoring variable (such as spinning machine speed, draft ratio, workshop temperature and humidity, air compressor pressure, etc.) to the principal component. These weighting coefficients are obtained by solving the kernel eigenvalue problem, and their absolute values ​​reflect the importance of the original variables in forming the principal component (i.e., the energy consumption pattern).

[0100] The raw variables in the load vector whose absolute values ​​of weight coefficients are greater than a set threshold (e.g., 0.3) are selected. These selected raw variables are the key energy consumption influencing factors that have the most significant impact on the current energy consumption pattern. For example, for the principal component of the "production load-dominated pattern", spinning machine speed (weight = 0.82) and twist (weight = 0.75) can be identified as key factors; for the principal component of the "environmental control-dominated pattern", return air humidity (weight = 0.68) and dew point temperature (weight = 0.61) may be identified as key factors.

[0101] S402, based on the sign of the weight coefficients of key energy consumption influencing factors, determines the direction of their correlation with the principal components.

[0102] In this step, the sign of the weighting coefficients is used to interpret the correlation direction between key influencing factors and principal components. A positive sign indicates that the variable is positively correlated with the principal component, meaning that increasing the value of the variable will lead to an increase in the principal component score (strengthening this energy consumption pattern); a negative sign indicates a negative correlation, meaning that increasing the value of the variable will lead to a decrease in the principal component score (weakening this energy consumption pattern).

[0103] For example, the workshop temperature is positively correlated with the principal component of the "air conditioning energy consumption mode" (weight = +0.71), while the outdoor temperature is negatively correlated with the principal component of the "heating energy consumption mode" (weight = -0.65).

[0104] This embodiment enables the precise identification of key factors that play a decisive role in the current energy consumption pattern and their direction of influence from dozens of monitored variables, providing a clear and actionable basis for subsequent energy consumption diagnosis and energy-saving optimization.

[0105] As an example, at least based on the numerical distribution, correlation, and changing trends of key energy consumption influencing factors, the main energy consumption patterns can be identified, including: S403, analyzing the relative position of the real-time values ​​of each key energy consumption influencing factor in its historical optimal operating range to obtain numerical distribution characteristics; establishing a dynamic correlation network among key energy consumption influencing factors, quantifying their coupling strength and direction of action to obtain correlation network characteristics; and extracting multi-scale changing trend characteristics of key energy consumption influencing factors, including short-term fluctuations, medium-term trends, and periodic patterns.

[0106] In this step, for each key energy consumption influencing factor, its historical optimal operating range under different production varieties and seasonal operating conditions is dynamically constructed (e.g., the efficient operating range of the spinning machine speed is 16000-18500 r / min). By calculating the deviation of the real-time value from the optimal range (e.g., the current value of 18200 r / min is at the upper edge of the optimal range), the operating efficiency status of each factor is evaluated.

[0107] Using time-varying correlation coefficients and Granger causality tests, a dynamic correlation network among key energy consumption influencing factors was constructed to quantify the real-time coupling relationship between factors. For example, the instantaneous correlation coefficient between air conditioning energy consumption and outdoor temperature reached 0.85, and a significant causal relationship existed between outdoor temperature and air conditioning energy consumption (p-value < 0.01).

[0108] By applying multi-scale time series analysis, we can simultaneously extract the short-term characteristics (such as the variance of fluctuations on a 5-minute scale), medium-term trends (such as the slope of linear trends on a 2-hour scale), and periodic patterns (such as the mean changes at the same time of day) of key energy consumption influencing factors, thus comprehensively characterizing the dynamic behavior of the factors.

[0109] S404, based on a pattern matching algorithm, integrates numerical distribution features, correlation network features, and multi-scale change trend features to generate interpretable descriptions of major energy consumption patterns.

[0110] Based on a pattern matching algorithm that combines rule-based reasoning and machine learning, the numerical distribution features, correlation network features, and multi-scale change trend features obtained in the first three steps are matched with the energy consumption pattern knowledge base. The final output is a pattern description that is both accurate and interpretable, such as "The current main energy consumption pattern is 'high-load production mode under high humidity environment', the energy consumption of spinning machines has increased to 45%, and the energy consumption of air conditioning dehumidification has increased by 15% year-on-year due to the increase in outdoor dew point temperature".

[0111] This invention also discloses a non-transient computer-readable storage medium having instructions embodied thereon that can be executed by one or more processors to perform the method described in any of the preceding embodiments.

[0112] Although the invention has been specifically shown and described with reference to preferred embodiments, those skilled in the art will understand that various modifications in form and detail may be made without departing from the spirit and scope of the invention. Accordingly, the disclosed invention should be considered merely illustrative and limited only by the scope specified in the appended claims.

Claims

1. A real-time energy consumption analysis method based on artificial intelligence, characterized in that, The process includes the following steps: real-time acquisition of multi-source heterogeneous data from the entire ring spinning process, including energy consumption data, equipment operating parameters, production process data, and environmental data; spatiotemporal alignment and fusion of the multi-source heterogeneous data to construct a comprehensive dataset; extraction of data for the current time period from the comprehensive dataset based on a sliding window of a set time length; and mapping of the original data within the sliding window to a high-dimensional feature space using a KPCA model for dimensionality reduction to extract several nonlinear principal components that can characterize the current main energy consumption patterns. When the update timing is reached, the kernel matrix and feature subspace of the KPCA model are updated online using an incremental learning algorithm, and the variance contribution rate of each nonlinear principal component is calculated. The update timing is determined based on the energy consumption mode failure semantics derived from the analysis of new data within the sliding window. The key energy consumption influencing factors for the current period are determined based on the variance contribution rate of each nonlinear principal component, and the main energy consumption modes are identified based at least on the numerical distribution, correlation, and changing trends of the key energy consumption influencing factors.

2. The energy consumption real-time analysis method based on artificial intelligence according to claim 1, characterized in that: Based on a sliding window with a set time length, data for the current period is extracted from the comprehensive dataset, including: real-time monitoring of production change signals in the comprehensive dataset, including at least one of product variety change signals, process parameter adjustment signals, and equipment start / stop status change signals; when a production change signal is detected, the set time length of the sliding window is automatically adjusted to a first preset time length; when no production change signal is detected and production is in a stable state, the set time length of the sliding window is automatically adjusted to a second preset time length that is greater than the first preset time length.

3. The energy consumption real-time analysis method based on artificial intelligence according to claim 1, characterized in that: The KPCA model maps the original data within a sliding window to a high-dimensional feature space and performs dimensionality reduction to extract several nonlinear principal components that can characterize the current main energy consumption patterns. This includes: calculating the kernel matrix of the original data within the sliding window based on the Gaussian radial basis function, nonlinearly mapping the original data to a high-dimensional feature space; centering the kernel matrix and solving for its eigenvalues ​​and eigenvectors; sorting the eigenvalues ​​from largest to smallest and selecting the eigenvectors corresponding to the k largest eigenvalues ​​to form a projection matrix; and projecting the centered kernel matrix onto the projection matrix to obtain the dimensionality-reduced nonlinear principal components.

4. The energy consumption real-time analysis method based on artificial intelligence according to claim 3, characterized in that: An incremental learning algorithm is used to update the kernel matrix and feature subspace of the KPCA model online. This includes: when a new batch of data flows into the sliding window, the expanded new kernel matrix is ​​quickly calculated using the rank-one update algorithm based on the original kernel matrix and the new batch of data; the new kernel matrix is ​​incrementally centered, and the eigenvalues ​​and eigenvectors of the new kernel matrix are approximately solved using an incremental feature subspace update algorithm; the number of principal components k to be retained is redetermined based on the updated eigenvalues, and the projection matrix is ​​updated.

5. The energy consumption real-time analysis method based on artificial intelligence according to claim 1, characterized in that: The update timing is determined based on the energy consumption pattern failure semantics derived from the analysis of new data within the sliding window. This includes: real-time calculation of the squared prediction error (SPE) statistic of newly flowing sample data into the sliding window based on the current KPCA model; and calculation of the dynamic control limit of the SPE statistic based on historical window data. When the SPE statistic is detected to have exceeded the dynamic control limit for a preset number of consecutive times, an energy consumption pattern failure semantic is determined to have occurred, and the KPCA model update procedure is triggered based on the energy consumption pattern failure semantic.

6. The energy consumption real-time analysis method based on artificial intelligence according to claim 5, characterized in that: The dynamic control limits for SPE statistics are calculated based on historical window data, including: acquiring the SPE statistics datasets from the most recent M historical sliding windows, selecting SPE data under stable production conditions to form a local normal SPE dataset; when the number of samples in the local normal SPE dataset is lower than a preset threshold, introducing industry benchmark SPE data as a supplement to form a hybrid SPE dataset; calculating the dynamic control limits based on the hybrid SPE dataset using kernel density estimation; and dynamically adjusting the weight of the industry benchmark data as local normal SPE data accumulates, until the amount of local normal SPE data reaches a sufficient threshold, at which point the dynamic control limits are calculated entirely using local data.

7. The energy consumption real-time analysis method based on artificial intelligence according to claim 6, characterized in that: The sufficient threshold is determined based at least on the complexity of the product variety and the frequency of changes in operating conditions.

8. The energy consumption real-time analysis method based on artificial intelligence according to claim 1, characterized in that: The key energy consumption influencing factors for the current period are determined based on the variance contribution rate of each nonlinear principal component. This includes: sorting the nonlinear principal components from largest to smallest according to their variance contribution rates, and selecting the top k nonlinear principal components as key principal components; for each key principal component, analyzing the weight coefficients of each original variable in its loading vector, and selecting the original variables whose absolute values ​​of the weight coefficients are greater than a set threshold as key energy consumption influencing factors; and determining the correlation direction between the key energy consumption influencing factors and the principal components based on the sign of their weight coefficients.

9. The energy consumption real-time analysis method based on artificial intelligence according to claim 8, characterized in that: The main energy consumption patterns are identified based on at least the numerical distribution, correlation, and changing trends of key energy consumption influencing factors. This includes: analyzing the relative position of the real-time values ​​of each key energy consumption influencing factor within its historical optimal operating range to obtain numerical distribution characteristics; establishing a dynamic correlation network among key energy consumption influencing factors, quantifying their coupling strength and direction of action to obtain correlation network characteristics; extracting multi-scale changing trend characteristics of key energy consumption influencing factors, including short-term fluctuations, medium-term trends, and periodic patterns; and generating an interpretable description of the main energy consumption patterns based on a pattern matching algorithm, integrating numerical distribution characteristics, correlation network characteristics, and multi-scale changing trend characteristics.

10. A non-transient computer-readable storage medium having instructions embodied thereon, characterized in that, The instructions can be executed by one or more processors to perform the method for implementing any one of claims 1-9.

Images