An artificial intelligence-based label production energy consumption monitoring method, device and medium

Abstract

This invention discloses an artificial intelligence-based method, equipment, and medium for monitoring energy consumption in tag production, relating to the field of production monitoring technology. The method includes: collecting tag production data, calibrating it, and defining a set of energy consumption monitoring nodes; defining anomaly factors and screening a set of steady-state second points; selecting energy consumption monitoring variables based on the anomaly factors, constructing anomaly scores for each anomaly factor, and converting them into activation vectors to generate state indices; dividing the data into windows and obtaining baseline sampling sets for each node to establish a window residual matrix; calculating the anomaly source intensity to establish a hypothetical observation model; processing the node residuals within the window and outputting multi-dimensional matrix features; constructing a parallel three-branch network to analyze the multi-dimensional matrix features, obtaining fused features, and mapping them to state probabilities and power predictions; and triggering early warnings based on the state probabilities output by the parallel three-branch network. This invention effectively improves the accuracy and timeliness of anomaly analysis in the tag production process.

Description

Technical Field

[0001] This invention relates to the field of production monitoring technology, and in particular to a method, equipment and medium for monitoring energy consumption in production based on artificial intelligence. Background Technology

[0002] As a crucial component of the packaging and printing industry, label printing and die-cutting composite production lines have undergone continuous evolution in recent years, focusing on integrated flexographic / offset printing, high-speed multi-color printing, refined zoning of drying systems, and deep integration of Industrial Internet of Things (IIoT) and Manufacturing Execution System (MES). Driven by new energy and "dual-carbon" policies, enterprises have significantly increased their demand for quantitative management of unit product energy consumption, order energy consumption, and process energy consumption. Energy consumption monitoring is gradually shifting from traditional "general metering + post-event reporting" to "itemized metering + online diagnostics + predictive optimization." Simultaneously, the label production process exhibits typical multivariate coupling characteristics: linear speed, tension, printing pressure, drying temperature zone settings, and fan load all exhibit thermal inertia and hysteresis in the time domain; different order materials (such as PP, PET, and paper base) and plate complexity cause systematic changes in drying load, transmission load, and die-cutting resistance, resulting in nonlinear, non-stationary, and periodic fluctuations in energy consumption. To improve monitoring accuracy and interpretability, existing research and engineering practices have begun to introduce data-driven methods. For example, energy consumption feature vectors are constructed using high-frequency electrical parameters, equipment operating status sequences, and quality inspection features. Then, energy consumption prediction, anomaly detection, and energy consumption attribution are achieved through machine learning or deep networks. Among these, multi-source heterogeneous data alignment, operating condition identification, time-series feature extraction, and multi-dimensional feature fusion have become key technology chains for highly reliable energy consumption monitoring systems. However, most existing tag production solutions are based on total energy consumption statistics or simple sub-meter data, lacking factor-based modeling oriented towards anomaly mechanisms. This makes it difficult to distinguish the contributions of different types of anomalies such as drying overload, die-cutting stagnation, transmission coupling malfunction, and idling standby, thus failing to achieve accurate maintenance and management. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, this invention provides a method, equipment, and medium for monitoring energy consumption in tag production based on artificial intelligence, which solves the problem that existing technologies lack factor-based modeling of anomaly mechanisms and cannot achieve precise maintenance and management.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an artificial intelligence-based method for monitoring energy consumption in tag production, comprising, Collect label production data for calibration and define a set of energy consumption monitoring nodes; Define anomaly factors and filter the set of steady-state second points. Select energy consumption monitoring variables based on the anomaly factors, construct anomaly scores for each anomaly factor, and convert them into activation vectors to generate state indexes. Data is divided into windows and a window residual matrix is ​​established by obtaining a baseline sampling set for each node. The intensity of the anomaly source is calculated to establish a hypothetical observation model. The residuals of nodes within the window are processed and multi-dimensional matrix features are output. A parallel three-branch network is constructed to analyze the multi-dimensional matrix features, obtain fused features, and map them into state probabilities and power predictions. The warning is triggered based on the state probability output by the parallel three-branch network.

[0006] As a preferred embodiment of the AI-based label production energy consumption monitoring method of the present invention, the step of collecting label production data for calibration and defining the energy consumption monitoring node set refers to collecting production data on the PLC, MES, and sub-meters of the label production line. Images of finished label products are collected using visual sensors, and defects in the finished labels are analyzed using a pre-trained visual detection model to calculate the pass rate as a quality characteristic. By comparing with the master clock to obtain the correction timestamp, a 1-second master time grid is established, and production data and quality characteristics are combined into multi-source data and aligned to the grid. Select energy consumption monitoring nodes from the label production line and form a node set S.

[0007] As a preferred embodiment of the AI-based label production energy consumption monitoring method of the present invention, the following steps are described: defining anomaly factors and screening a set of steady-state second points; selecting energy consumption monitoring variables based on the anomaly factors; constructing anomaly scores for each anomaly factor and converting them into activation vectors to generate state indexes; and defining anomaly factors based on the energy consumption monitoring node set. ; Within each order, scan the second grid and filter the set of steady-state second points. ; Select energy consumption monitoring variables based on abnormal factors In the steady-state second point set Calculate the steady-state median With MAD; Robust Z-score of energy consumption monitoring variables is defined based on steady-state median and MAD. ; Anomaly scores are constructed for each anomaly factor based on the robust Z-score. ; For each anomalous factor, an activation threshold and a release threshold are set, and the corresponding anomalous score is binarized into an activation variable. ; Combining all activation variables into a vector And generate a state index through binary encoding. .

[0008] As a preferred embodiment of the AI-based tag production energy consumption monitoring method of the present invention, wherein: the step of dividing the data by window and obtaining a baseline sampling set for each node to establish a window residual matrix, and calculating the anomaly source intensity to establish a hypothetical observation model refers to the second-grid time point... Based on the time taken from the start of the order, the time is 1 second. Ending seconds Generate window start sequence ; For each window's starting point, divide the window set into a fixed set of N windows and define the baseline replay time. Obtain the baseline sample set for each node s ; Compute node s in time Constructing a window residual matrix from residual power ; Use MAD to estimate the noise scale in the baseline sample set; Based on anomaly scoring Calculate the strength of the anomaly source ; A hypothetical observation model is established by combining the intensity of the anomaly source.

[0009] As a preferred embodiment of the AI-based tag production energy consumption monitoring method of the present invention, the step of processing the node residuals within the window and outputting multi-dimensional matrix features includes: After noise normalization of the residual at each node within the window, symmetrical amplitude limiting is applied. For each window w and node s, the node energy characteristics are obtained by statistically averaging the node residuals after symmetric clipping. Logarithmic compression of node energy features yields node features. And form the output matrix ; The node similarity weights are calculated using the RBF kernel, and a continuous adjacency matrix is ​​constructed. ; Continuous adjacency matrix Binarize into a binary adjacency matrix ; After removing the mean from the node residuals based on the node energy characteristics, symmetrical amplitude limiting is applied. Calculate the sample covariance of the nodal residuals after symmetric clipping. And construct the sample covariance matrix The sample covariance matrix is ​​shrunk. The shrunken sample covariance matrix is ​​normalized to the correlation coefficient matrix. And constitute a tensor ; Read the PLC status in each window and convert it into a numerical value. ; Constructing the additional vector at each time step ; The original input vector is obtained by concatenating the node residual vector and the additional vector. After noise normalization and symmetric clipping of the original input vector, linear embedding is performed and a linear embedding matrix is ​​constructed. .

[0010] As a preferred embodiment of the AI-based tag production energy consumption monitoring method of the present invention, wherein: the construction of a parallel three-branch network to analyze multi-dimensional matrix features to obtain fused features and map them to state probabilities and power predictions refers to the construction of a parallel three-branch network, including a GAT network, a CNN network and a Transformer network; Output matrix and binary adjacency matrix Inputting a two-layer GAT network yields node embeddings, which are then processed using global max pooling to obtain graph-level vectors. ; tensor After inputting into the CM-CNN convolutional network, the vector is flattened and reduced to one dimension to obtain a vector. ; Linear embedding matrix Inputting the Transformer temporal branch yields a contextualized sequence representation, which is then flattened into a one-dimensional vector. ; The concatenation feature is obtained by concatenating the output vectors of the three-branch network. And integrate them; Based on the fusion results, a probability mapping is performed, and the state probability and power prediction are output.

[0011] As a preferred embodiment of the AI-based tag production energy consumption monitoring method of the present invention, wherein: the step of determining the triggering of the early warning based on the state probability output by the parallel three-branch network refers to calculating the probability ratio statistic based on the window state probability output by the parallel three-branch network. The probability ratio statistic is compared with the threshold. The comparison is performed, and if the value exceeds the threshold, an alarm is triggered and the abnormal state with the highest probability is output based on the state probability.

[0012] As a preferred embodiment of the AI-based tag production energy consumption monitoring method of the present invention, wherein: the energy consumption of a single tag is calculated in seconds based on power prediction, and the total energy consumption of the order is aggregated and output within the order to form an energy consumption monitoring list.

[0013] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the artificial intelligence-based tag production energy consumption monitoring method described in the first aspect of the present invention.

[0014] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the artificial intelligence-based tag production energy consumption monitoring method described in the first aspect of the present invention.

[0015] The beneficial effects of this invention are as follows: by defining the set of data collection calibration and energy consumption monitoring nodes, screening abnormal factors and steady-state second points, constructing window residuals and hypothetical observation models, performing fusion analysis of multidimensional matrix features through parallel three-branch networks and outputting state probabilities and power predictions, and triggering early warnings based on state probabilities, the problem of controllable early warning and reliable monitoring of online energy consumption anomalies is solved, effectively improving the accuracy and timeliness of anomaly analysis in the tag production process. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the AI-based tag production energy consumption monitoring method in Example 1.

[0018] Figure 2 This is a flowchart of the input and processing of the parallel three-branch network in Example 1. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0022] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides an artificial intelligence-based method for monitoring energy consumption in tag production, comprising the following steps: S1. Collect tag production data for calibration and define a set of energy consumption monitoring nodes; S1.1 Production data is collected from the PLC, MES, and sub-meters on the label production line. The PLC collects data by configuring OPC UA subscription on the gateway: pulling key points every 1 second: equipment status (running / standby / stopping), line speed setting / actual, temperature zone setting / actual, printing pressure setting / actual, etc. The MES collects data by checking the MES task status (new work order / change order / parameter change) on the gateway every 1 second and pulling: order number, material, width, number of colors, page ID, and planned output. The sub-meters collect power data at 10Hz and cache the raw 10Hz data in a 1-second bucket. When the second boundary is reached, the data is aggregated and output. Images of finished label products are collected using visual sensors, and defects in the finished labels are analyzed using pre-trained visual detection models (such as YOLOv5). The pass rate is then calculated as a quality characteristic. S1.2. Obtain the correction timestamp by comparing it with the master clock, establish a 1-second master time grid, and align the production data and quality characteristics into the grid as multi-source data:

[0023] in To correct the timestamp, As a reference starting point, the preferred time is the start time of the work order. The grid period is per second, preferably 1 second. For raster index, This represents the k-th second of the raster timeframe; It should be noted that each sub-meter collects M power data points in the k-th second bucket. ,in When sampling at 10Hz, M is 10. Before aggregation and output when the second boundary is reached, the power data needs to be de-extreme Windsorized to suppress transient communication glitches.

[0024] in and Power data within the second bucket and Quantiles The power data is after Windsorization; S1.3 Select energy consumption monitoring nodes from the label production line and form a node set S, where there are a total of 12 energy consumption monitoring nodes. The label production line is fixedly divided, including: s=1, the total power of the entire production line; s=2, unroll drive; s=3…8, color group motor; s=9,10, Drying fan / heating; s=11, die-cutting motor; s=12, winding drive.

[0025] S2. Define abnormal factors and filter the set of steady-state second points. Select energy consumption monitoring variables according to the abnormal factors, construct an abnormal score for each abnormal factor and convert it into an activation vector to generate a state index. S2.1 Define anomaly factors based on the energy consumption monitoring node set. P=4, where the outlier factor This indicates a drying anomaly, corresponding to nodes 9 and 10, representing deviations in temperature zone and drying power. The abnormality in the die-cutting load corresponds to node 11, indicating abnormal power spikes or fluctuations. This indicates a printing transmission malfunction, corresponding to nodes 3-8, representing a power-speed coupling error. This indicates an idling / standby abnormality, corresponding to node 1, representing a mismatch between the status and energy consumption; Within each order, scan the second grid and filter the set of steady-state second points. ; It should be noted that the steady-state second point set needs to meet the following conditions: the PLC is in running status, the pass rate is greater than 95%, and the linear speed is greater than the minimum speed threshold. , The preferred speed is 30 m / min. Speeds below this are mostly used for machine setup / belt threading / transition, and using this as a baseline will contaminate the threshold. S2.2 Select energy consumption monitoring variables based on abnormal factors This refers to power data used to identify anomalies, such as drying power, die-cutting power, and color group power, which correspond to anomaly factors and are located in the steady-state second point set. Calculate the steady-state median With MAD:

[0026]

[0027] Robust Z-score of energy consumption monitoring variables is defined based on steady-state median and MAD. :

[0028] in The scaling factor. To prevent the protection constant from being zero in the denominator, it is preferably 10. -3 ; Specifically, the scaling factor b is preferably 1.4826, so that MAD is comparable to the standard deviation under normal conditions; taking a larger value will reduce sensitivity, while taking a smaller value will increase sensitivity but make it prone to false alarms. S2.3 Construct anomaly scores for each anomaly factor based on the robust Z-score. ; Abnormal factors in drying : First, based on the power of node 9 and node 10 power Calculate the total drying power by summing the results. The total drying power is converted into a robust Z-score as the drying anomaly score. ; Die-cut load anomaly factor : Obtain the power of node 11 As die-cutting power And recently The standard deviation is calculated per second as the short-window fluctuation amplitude. Die-cutting power and short window fluctuation amplitude Convert to robust Z-score and Robust Z-score and Perform weighted sampling to obtain anomaly scores for die-cutting load. ; Abnormal factors of printing transmission : Summing the power of nodes 3-8 yields the printing power. In the steady-state second point set Fit a linear coupled baseline using the least squares method and calculate the coupling residuals. Printing power and coupling residual Convert to robust Z-scores and perform weighted summation to form a printing transmission anomaly score. ; Abnormal factors for idling / standby : The total power of node 1 is converted into a robust Z-score and used as an idling / standby anomaly score. ; S2.4. Set activation and release thresholds for each anomalous factor and binarize the corresponding anomalous score into an activation variable. :

[0029] in The activation variable represents the time when the p-th type of anomaly occurs. Whether it is activated, As the activation threshold, To release the threshold, To ensure continuous confirmation for the specified number of seconds, 3 seconds is preferred. For indicator functions; It should be noted that the startup threshold The preferred value is 3, corresponding to a significant deviation of approximately 3 robust standard deviations, which can suppress false alarms due to normal fluctuations and release the threshold. The preferred value is 2, which creates hysteresis and prevents the score from fluctuating around the threshold, causing frequent switching. The number of consecutive confirmation seconds is also considered. The preferred value is 3, which filters out single-second burrs; Combining all activation variables into a vector And generate a state index through binary encoding. :

[0030] State index This represents the energy consumption network status number.

[0031] S3. Divide the data into windows and obtain the baseline sampling set for each node to establish a window residual matrix. Calculate the anomaly source intensity and establish a hypothetical observation model. S3.1, Time points in seconds grid Based on the time taken from the start of the order, the time is 1 second. Ending seconds Generate window start sequence :

[0032] Where H is the step size, which is fixed at 50. The starting number of the window; For each window's starting point, divide the window set into a fixed set of N=200 windows, and define the baseline replay time. Obtain the baseline sample set for each node s :

[0033] in For the i-th baseline time in the baseline sampling set, For time linear velocity, For time The pass rate The minimum pass rate is 95%, preferably 95%. It should be noted that the baseline review time Setting it to 600s is sufficient to cover thermal inertia and short-term fluctuations, stabilizing the baseline. The baseline sampling set needs to have at least 300 time samples. If there are not enough, the baseline is frozen to the previous time baseline to avoid baseline drift caused by sparse samples. Compute node s in time Constructing a window residual matrix from residual power :

[0034]

[0035] in For residual power, The median of the baseline sample set; S3.2. Using MAD to estimate the noise scale in the baseline sample set:

[0036]

[0037] in For node s in time The noise scale; Based on anomaly scoring Calculate the strength of the anomaly source :

[0038] S3.3. Establish a hypothetical observation model based on the intensity of the anomaly source:

[0039] in Let be the residual of node s at the nth sampling point within the window. For noise terms, , This indicates no anomalies, meaning all activation variables are 0; in this case, only noise exists. This indicates that at least one anomalous factor has an activation variable of 1. is the coupling coefficient from the anomaly factor p to node s; It should be noted that for each node s, the coupling coefficient needs to be adjusted using training data. Perform calibration:

[0040] in The regularization coefficient is preferably 0.1, which suppresses overfitting caused by excessively large coupling coefficients.

[0041] S4. Process the node residuals within the window and output multi-dimensional matrix features. Construct a parallel three-branch network to analyze the multi-dimensional matrix features, obtain fused features, and map them into state probabilities and power predictions. S4.1. After noise normalization of the residuals at each node within the window, apply symmetrical amplitude limiting:

[0042]

[0043] in As a protection constant, The limiting factor is preferably 6, which limits the normalized residual to between -6 and 6. For each window w and node s, the node energy characteristics are obtained by statistically averaging the node residuals after symmetric clipping. Logarithmic compression of node energy features yields node features. And form the output matrix :

[0044] in The scaling factor is preferably 1, because of the node energy characteristics. Noise normalization and limiting have been performed, and the range is stable after logarithmic compression. S4.2 Calculate node similarity weights using the RBF kernel and construct a continuous adjacency matrix. :

[0045] in Let be the similarity weights between nodes i and j. and Let i and j be the node energy characteristics. Radius of influence of RBF; Continuous adjacency matrix Binarize into a binary adjacency matrix :

[0046] in This is the binarization threshold; S4.3. After removing the mean from the node residuals based on the node energy characteristics, apply symmetrical limiting:

[0047]

[0048] in The residuals after removing the mean. The mean-free residual after amplitude limiting The limiting threshold is preferably 3 times. It retains most of the true fluctuations while cutting off spikes; Calculate the sample covariance of the nodal residuals after symmetric clipping. And construct the sample covariance matrix Shrink the sample covariance matrix:

[0049] Where I is the identity matrix. The sample covariance matrix after shrinkage. The mean diagonal value of the covariance. The shrinkage coefficient is preferably 0.1, which improves numerical stability without significantly altering the structure. The shrunken sample covariance matrix is ​​normalized to the correlation coefficient matrix. And constitute a tensor ; S4.4 Read the PLC status in each window and convert it into numerical values. The PLC status includes two states: running, standby, and stationary. The running value is set to 1, the standby value to 0.5, and the stationary value to 0. Constructing the additional vector at each time step :

[0050] in Linear velocity, The average temperature range is obtained through temperature range setting or actual measurement. The original input vector is obtained by concatenating the node residual vector and the additional vector. , Using dimension , the original input vector is noise normalized and symmetrically clipped before linear embedding is performed, and a linear embedding matrix is ​​constructed. :

[0051] in This is the original input vector after amplitude limiting. For the embedding layer parameter matrix, For bias, d is the embedding dimension, which is fixed at 16. This is the embedding vector.

[0052] S4.5 Construct a parallel three-branch network, including a GAT network, a CNN network, and a Transformer network; Output matrix and binary adjacency matrix Inputting a two-layer GAT network (each layer has multi-head attention, and the outputs of each attention head are concatenated) yields node embeddings, which are then processed by global max pooling to obtain graph-level vectors. ; tensor After inputting into the CM-CNN convolutional network, the vector is flattened and reduced to one dimension to obtain a vector. ; It should be noted that the CM-CNN convolutional network contains two convolutional layers. Convolutional layer 1 uses 20 5*5 convolutional kernels and 2*2 max pooling, while convolutional layer 2 uses 20 3*3 convolutional kernels and 2*2 max pooling. Both convolutional layers 1 and 2 use LeakyReLU activation. Linear embedding matrix Inputting the Transformer temporal branch yields a contextualized sequence representation, which is then flattened into a one-dimensional vector. ; S4.6. Concatenate the output vectors of the three-branch network to obtain the concatenation features. And then merge them:

[0053] in and For fusion layer parameters, For integration and sharing representation; Based on the fusion results, a probability mapping is performed, and the state probability and power prediction are output:

[0054]

[0055]

[0056] in For classifying logits, The classification logits for the i-th state (16 state indices), , , , For training parameters, Let be the probability that the window belongs to the i-th state. The predicted power vector contains the predicted power of all nodes in a fixed order; It should be noted that the parameters of the parallel three-branch network, probability mapping, and power prediction are all obtained by training on the training data by calculating the cross-entropy.

[0057] S5. Trigger an early warning based on the state probability output by the parallel three-branch network; S5.1 Calculate the probability ratio statistic based on the window state probabilities output by the parallel three-branch network. The probability ratio is obtained by calculating the ratio of the probability of a completely normal state (no abnormal factors) to the sum of the probabilities of an abnormal state (one or more abnormal factors). This probability ratio statistic is then compared with a threshold. The comparison is performed, and if the value exceeds the threshold, an alarm is triggered and the abnormal state with the highest probability is output based on the state probability.

[0058] It should be noted that the threshold The Pf quantile method is used to determine the quantiles from at least 2000 windows.

[0059] S5.2 Obtain the power prediction calculation for the second-level single standard energy consumption, and aggregate and output the total energy consumption of the order to form an energy consumption monitoring list.

[0060] This embodiment also provides a computer device applicable to the label production energy consumption monitoring method based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the label production energy consumption monitoring method based on artificial intelligence as proposed in the above embodiment. The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0061] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the energy consumption monitoring method for tag production based on artificial intelligence as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0062] In summary, this invention solves the problem of controllable early warning and reliable monitoring of online energy consumption anomalies by defining the data collection calibration and energy consumption monitoring node set, screening abnormal factors and steady-state second points, constructing window residuals and hypothetical observation models, performing fusion analysis of multidimensional matrix features through parallel three-branch networks and outputting state probabilities and power predictions, and triggering early warnings based on state probabilities. This effectively improves the accuracy and timeliness of anomaly analysis in the tag production process.

[0063] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An artificial intelligence-based label production energy consumption monitoring method, characterized by: include, Collect label production data for calibration and define a set of energy consumption monitoring nodes; Define anomaly factors and filter the set of steady-state second points. Select energy consumption monitoring variables based on the anomaly factors, construct anomaly scores for each anomaly factor, and convert them into activation vectors to generate state indexes. Data is divided into windows and a window residual matrix is ​​established by obtaining a baseline sampling set for each node. The intensity of the anomaly source is calculated to establish a hypothetical observation model. The residuals of nodes within the window are processed and multi-dimensional matrix features are output. A parallel three-branch network is constructed to analyze the multi-dimensional matrix features, obtain fused features, and map them into state probabilities and power predictions. The warning is triggered based on the state probability output by the parallel three-branch network. 2.The artificial intelligence-based label production energy consumption monitoring method of claim 1, wherein: The collection of label production data for calibration and the definition of the energy consumption monitoring node set refer to the collection of production data from the PLC, MES, and sub-meters on the label production line. Images of finished label products are collected using visual sensors, and defects in the finished labels are analyzed using a pre-trained visual detection model to calculate the pass rate as a quality characteristic. By comparing with the master clock to obtain the correction timestamp, a 1-second master time grid is established, and production data and quality characteristics are combined into multi-source data and aligned to the grid. Select energy consumption monitoring nodes from the label production line and form a node set S.

3. The method for monitoring energy consumption in tag production based on artificial intelligence as described in claim 2, characterized in that: The definition of the abnormal factor and the screening of the steady-state second point set, the selection of the energy consumption monitoring variable according to the abnormal factor, the construction of the abnormal score for each abnormal factor and the conversion into an activation vector to generate a state index, and the definition of the abnormal factor according to the energy consumption monitoring node set ; Within each order, scan the second grid, filtering the set of steady second points ; Selecting energy consumption monitoring variables according to abnormal factors At steady state second point set Calculate steady state median With MAD; Robust Z-score of energy consumption monitoring variables is defined based on steady-state median and MAD. ; Anomaly scores are constructed for each anomaly factor based on the robust Z-score. ; For each anomalous factor, an activation threshold and a release threshold are set, and the corresponding anomalous score is binarized into an activation variable. ; Combining all activation variables into a vector And generate a state index through binary encoding. .

4. The method for monitoring energy consumption in tag production based on artificial intelligence as described in claim 3, characterized in that: The process of dividing data by window and obtaining a baseline sampling set for each node to establish a window residual matrix, calculating the anomaly source intensity, and establishing a hypothetical observation model refers to the second-grid time point. Based on the time taken from the start of the order, the time is 1 second. Ending seconds Generate window start sequence ; For each window's starting point, divide the window set into a fixed set of N windows and define the baseline replay time. Obtain the baseline sample set for each node s ; Compute node s in time Constructing a window residual matrix from residual power ; Use MAD to estimate the noise scale in the baseline sample set; Based on anomaly scoring Calculate the strength of the anomaly source ; A hypothetical observation model is established by combining the intensity of the anomaly source.

5. The method for monitoring energy consumption in tag production based on artificial intelligence as described in claim 4, characterized in that: The process of processing the residuals of nodes within the window and outputting multidimensional matrix features includes: After noise normalization of the residual at each node within the window, symmetrical amplitude limiting is applied. For each window w and node s, the node energy characteristics are obtained by statistically averaging the node residuals after symmetric clipping. Logarithmic compression of node energy features yields node features. And form the output matrix ; The node similarity weights are calculated using the RBF kernel, and a continuous adjacency matrix is ​​constructed. ; Continuous adjacency matrix Binarize into a binary adjacency matrix ; After removing the mean from the node residuals based on the node energy characteristics, symmetrical amplitude limiting is applied. Calculate the sample covariance of the nodal residuals after symmetric clipping. And construct the sample covariance matrix The sample covariance matrix is ​​shrunk. The shrunken sample covariance matrix is ​​normalized to the correlation coefficient matrix. And constitute a tensor ; Read the PLC status in each window and convert it into a numerical value. ; Constructing the additional vector at each time step ; The original input vector is obtained by concatenating the node residual vector and the additional vector. After noise normalization and symmetric clipping of the original input vector, linear embedding is performed and a linear embedding matrix is ​​constructed. .

6. The method for monitoring energy consumption in tag production based on artificial intelligence as described in claim 5, characterized in that: The construction of a parallel three-branch network to analyze multidimensional matrix features, obtain fused features, and map them to state probabilities and power predictions refers to the construction of a parallel three-branch network, including a GAT network, a CNN network, and a Transformer network. Output matrix and binary adjacency matrix Inputting a two-layer GAT network yields node embeddings, which are then processed using global max pooling to obtain graph-level vectors. ; tensor After inputting into the CM-CNN convolutional network, the vector is flattened and reduced to one dimension to obtain a vector. ; Linear embedding matrix Inputting the Transformer temporal branch yields a contextualized sequence representation, which is then flattened into a one-dimensional vector. ; The concatenation feature is obtained by concatenating the output vectors of the three-branch network. And integrate them; Based on the fusion results, a probability mapping is performed, and the state probability and power prediction are output.

7. The method for monitoring energy consumption in tag production based on artificial intelligence as described in claim 6, characterized in that: The step of determining the trigger warning based on the state probability output by the parallel three-branch network refers to calculating the probability ratio statistic based on the window state probability output by the parallel three-branch network. The probability ratio statistic is compared with the threshold. The comparison is performed, and if the value exceeds the threshold, an alarm is triggered and the abnormal state with the highest probability is output based on the state probability.

8. The method for monitoring energy consumption in tag production based on artificial intelligence as described in claim 6, characterized in that: The obtained power prediction calculation is performed on a second-level single-standard energy consumption basis, and the total energy consumption of the order is aggregated and output within the order to form an energy consumption monitoring list.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the artificial intelligence-based tag production energy consumption monitoring method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the AI-based tag production energy consumption monitoring method according to any one of claims 1 to 8.

Images