A self-driven energy consumption tracking and energy-saving management system based on AI large models

Abstract

This invention relates to the field of industrial energy management technology and discloses an AI-driven large-scale model-based power consumption tracking and energy-saving control system. The method includes: acquiring multi-cycle power consumption sampling data from industrial equipment and performing phase normalization processing; extracting high-order statistical features at the same phase points to generate cross-cycle energy consumption feature vectors; calculating trend slope and phase synchronization to filter abnormal phase points; encoding abnormal information and inputting it into a large-scale model to generate semantic embedding vectors; using an attention mechanism to calculate semantic distance distribution and generate energy consumption anomaly feature labels; and generating cause diagnosis and energy-saving control instructions through large-scale model-driven inference. This invention achieves an automated closed loop from energy consumption tracking to energy-saving control.

Description

Technical Field

[0001] This invention relates to the field of industrial energy management technology, and more specifically, to an AI-driven large-scale model-based power consumption tracking and energy-saving control system. Background Technology

[0002] In industrial production parks, precision manufacturing equipment repeatedly performs processing operations according to strict process cycles, and the equipment's power consumption fluctuates regularly with these cycles. When equipment performance deteriorates or operating parameters deviate, the power consumption pattern changes, specifically manifested as minute power consumption disturbances at certain phase points of each process cycle. These disturbances are often masked by normal process fluctuations within a single cycle and are difficult to identify, but they exhibit a phase-synchronized cumulative growth trend over multiple process cycles, causing a continuous increase in power consumption.

[0003] Existing technologies typically employ threshold-based anomaly detection methods or long-term statistical analysis methods to monitor equipment energy consumption. Threshold-based anomaly detection methods set upper and lower limits for energy consumption, triggering an alarm when real-time energy consumption exceeds these limits. Long-term statistical analysis methods identify overall trends in energy consumption by performing statistical analyses such as mean and variance on historical energy consumption data.

[0004] However, existing technologies have the following problems: First, the energy consumption micro-disturbance amplitude in the early stage of equipment efficiency degradation is much smaller than the normal process fluctuation range. Conventional anomaly detection methods within a single cycle cannot identify this kind of hidden energy consumption growth because the detection threshold is higher than the micro-disturbance amplitude. Second, long-term statistical analysis methods mix and average signals of different phases, causing energy consumption anomaly patterns related to specific process phases to be submerged in the averaging process. Third, traditional energy-saving management relies on manual analysis and manual adjustment of equipment parameters, and cannot automatically generate and execute energy-saving strategies based on energy consumption tracking results, resulting in low energy-saving management efficiency. Summary of the Invention

[0005] This invention provides an AI large-scale model self-driven power consumption tracking and energy-saving management system, which solves the technical problems in related technologies such as the inability to identify hidden power consumption growth, the averaging of abnormal patterns, and the inability to automatically generate energy-saving strategies.

[0006] This invention provides a self-driven energy consumption tracking and energy-saving management method for large AI models, comprising: The system acquires energy consumption sampling data from multiple consecutive process cycles of industrial equipment, segments the energy consumption time sequence according to the process cycle boundary, performs phase normalization processing on each cycle segment, and generates a phase-aligned multi-cycle energy consumption segment sequence. High-order statistical features are extracted from the same phase point for each period segment after phase alignment, and a cross-cycle energy consumption feature vector sequence for each phase point is generated. The trend slope and phase synchronization index are calculated for the cross-cycle feature vector sequence of each phase point. Phase points with trend slope exceeding the trend threshold and phase synchronization higher than the synchronization threshold are selected to form an abnormal phase point set. The energy consumption feature vector sequence of the abnormal phase point set, along with the process stage identifier and equipment operating parameters, is encoded into structured descriptive text and input into the encoding layer of the large model to generate a semantic embedding vector of equipment energy consumption status. Load the current semantic embedding vector and the benchmark embedding vector set in the context window of the large model, calculate the semantic distance distribution using the attention mechanism, amplify the weights of the abnormal dimensions that deviate from the benchmark, and generate energy consumption anomaly feature labels. The energy consumption anomaly feature tags are associated and matched with the adjustable parameter range of the equipment and the energy-saving control rule library, and then input into the large model for self-driven reasoning to generate energy consumption anomaly cause diagnosis results. Based on the causal diagnosis results, a large model is used to autonomously generate equipment parameter adjustment schemes, and the adjustment schemes are encoded into an executable energy-saving control instruction sequence.

[0007] Furthermore, the higher-order statistical features include kurtosis, skewness, and autocorrelation decay coefficient; Kurtosis is obtained by calculating the ratio of the energy consumption value at the same phase point in each cycle to the square of the fourth central moment and the second central moment of the average energy consumption across cycles at that phase point. The skewness is obtained by calculating the ratio of the energy consumption value at the same phase point in each cycle to the 1.5 power of the third-order central moment and the second-order central moment of the average energy consumption across cycles at the phase point; The autocorrelation decay coefficient is obtained by calculating the autocorrelation function under different time delays on the cross-cycle energy consumption sequence at the phase point and fitting an exponential decay model.

[0008] Furthermore, the phase synchronization index is obtained by calculating the correlation coefficient of energy consumption values ​​at points of the same phase during adjacent cycles, specifically as follows: The energy consumption value sequence of the first N-1 cycles at the target phase point is compared with the energy consumption value sequence of the next N-1 cycles at the same phase point using Pearson correlation calculation. The resulting correlation coefficient is used as the phase synchronization index of the phase point, where N is the total number of process cycles.

[0009] Furthermore, the phase normalization process includes: The time coordinates of each period segment are mapped to a standard phase interval, so that each period segment has a unified phase coordinate system; A dynamic time warping algorithm is used to perform nonlinear correction on the time axis of each period segment, eliminating the difference in period length caused by fluctuations in equipment operating speed.

[0010] Furthermore, the method for determining the process cycle boundary includes: Calibration is performed based on the process start / stop signals output by the equipment control system; Based on the periodic characteristics of energy consumption time series data, the period length is identified through autocorrelation analysis and then segmented.

[0011] Furthermore, the calculation of semantic distance distribution using the attention mechanism includes: The attention weight matrix between the current embedded vector and each vector in the baseline set is calculated using the multi-head attention mechanism of the large model. The semantic distance between the current embedding vector and each vector in the benchmark set is calculated based on the attention weight matrix, forming a semantic distance distribution; Dimensions in the semantic distance distribution that deviate from the baseline mean by more than a preset standard deviation are weighted and amplified.

[0012] Furthermore, the diagnostic results for the causes of the energy consumption anomalies include: The inferred anomaly cause category, cause confidence score, list of associated parameter offsets, and offset direction of each parameter.

[0013] Furthermore, each instruction in the energy-saving control instruction sequence includes a target device identifier, a parameter adjustment instruction code, and an adjustment parameter value; The equipment parameter adjustment plan includes the name of the parameter to be adjusted, the target adjustment value, the timing of the adjustment, and the expected energy-saving effect.

[0014] Furthermore, it also includes: Send the energy-saving control command sequence to the equipment controller for execution; Collect energy consumption data after execution and calculate the deviation between the actual energy consumption change and the expected energy saving target; The deviation information is fed back into the context memory of the large model for adaptive optimization during the subsequent generation of control strategies.

[0015] This invention provides a self-driven energy consumption tracking and energy-saving management system for large AI models, comprising: The data acquisition and phase normalization module is used to acquire multi-cycle power consumption sampling data of industrial equipment, divide it according to the process cycle boundary and perform phase normalization processing to generate a phase-aligned multi-cycle power consumption segment sequence. The feature extraction module is used to extract high-order statistical features from each period segment at the same phase point and generate a cross-period energy consumption feature vector sequence. The abnormal phase point screening module is used to calculate the trend slope and phase synchronization index of each phase point and screen them to form a set of abnormal phase points; The semantic embedding generation module is used to encode the energy consumption characteristics and process information of abnormal phase points into structured descriptive text, and generate semantic embedding vectors through the large model encoding layer. The anomaly feature label generation module is used to calculate the semantic distance distribution using the large model attention mechanism and amplify the weight of the anomaly dimension to generate energy consumption anomaly feature labels. The cause diagnosis module is used to associate and match energy consumption anomaly feature tags with the energy-saving control rule base, and generate corresponding cause diagnosis results through large-scale model self-driven reasoning. The control instruction generation module is used to generate equipment parameter adjustment schemes based on the cause diagnosis results using a large model and encode them into an energy-saving control instruction sequence.

[0016] The beneficial effects of this invention are as follows: This invention maps the energy consumption time sequence of each process cycle to a unified phase coordinate system through phase normalization processing, so that the energy consumption values ​​of different cycles at the same process phase point can be directly compared and superimposed. This solves the technical problem that single-cycle detection cannot track the hidden energy consumption growth due to the disturbance amplitude being lower than the normal fluctuation range, and achieves the technical effect of being able to identify small energy consumption disturbances that are difficult to detect in a single cycle.

[0017] This invention refines the analytical granularity of anomaly detection to the level of a single phase point by extracting high-order statistical features at the same phase point and calculating the phase synchronization index. This solves the technical problem of anomaly patterns being submerged due to the mixing and averaging of different phase signals in long-term statistical analysis, and achieves the technical effect of accurately tracking and locating energy consumption anomaly patterns related to specific process phases.

[0018] This invention solves the technical problem of relying on manual analysis and manual parameter adjustment in traditional methods by autonomously generating cause diagnosis and energy-saving control instructions through large-scale model self-driven reasoning, and achieves the technical effect of automated closed-loop management from energy consumption tracking to energy-saving control.

[0019] This invention solves the technical problem that energy-saving management strategies cannot adaptively optimize by continuously collecting the strategy execution effect through a closed-loop mechanism of execution feedback and feeding it back to a large model, thus achieving the technical effect of continuous optimization of energy-saving management strategies. Attached Figure Description

[0020] Figure 1 This is a flowchart of the AI ​​large model self-driven power consumption tracking and energy-saving management method of the present invention; Figure 2 This is a time-series comparison diagram of multi-cycle power consumption of the present invention; Figure 3 This is the cross-period high-order statistical characteristic heatmap of the present invention; Figure 4 This is a scatter plot of abnormal phase point screening according to the present invention; Figure 5 This is the semantic distance distribution box-and-whisker diagram of the present invention; Figure 6This is a bar chart comparing the energy-saving control effects of the present invention; Figure 7 This is the multi-device energy consumption tracking relationship diagram of the present invention. Detailed Implementation

[0021] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0022] At least one embodiment of the present invention discloses a self-driven power consumption tracking and energy-saving management method for large AI models, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain multi-cycle power consumption sampling data of equipment in the industrial park, divide the power consumption time sequence according to the process cycle boundary, and generate a phase-aligned multi-cycle power consumption segment sequence.

[0023] Specifically, high-frequency power consumption sampling data of equipment in the industrial park is collected over multiple consecutive process cycles, with the sampling frequency determined based on the equipment's process characteristics. The energy consumption time series is divided into a set of cycle segments according to the process cycle boundaries. Phase normalization is performed on each cycle segment to uniformly map the phase coordinates of each cycle segment to a standard phase interval. Generate a phase-aligned multi-cycle power consumption segment sequence ,in , , These represent the phases of the 1st, 2nd, and Nth periods, respectively. Energy consumption value at the location, This represents the total number of cycles.

[0024] It should be noted that the determination of the above-mentioned process cycle boundary can be based on the process start and stop signals output by the equipment control system, or it can be based on the periodic characteristics of energy consumption time series data, and the cycle length can be automatically identified and divided by autocorrelation analysis.

[0025] It should be noted that the input for the aforementioned autocorrelation analysis is an energy consumption time-series data sequence, and the output is an autocorrelation coefficient sequence under different time lags. By identifying the periodic peak positions in the autocorrelation coefficient sequence, the dominant period length of the energy consumption time series is determined, and the time series data is divided into periodic segments based on the dominant period length.

[0026] In this embodiment of the application, in order to improve the accuracy of phase alignment, when performing phase normalization processing on each period segment, a dynamic time warping algorithm can be used to perform nonlinear correction on the time axis of each period segment to eliminate the slight difference in period length caused by fluctuations in the operating speed of the equipment, thereby obtaining a more accurate phase alignment effect.

[0027] It should be noted that the aforementioned dynamic time warping algorithm takes as input the periodic segment sequence to be corrected and a reference periodic segment, and outputs the time-axis corrected periodic segment sequence. This algorithm calculates the cumulative distance matrix between the segment to be corrected and the reference segment, finds the time-axis mapping path that minimizes the cumulative distance, and performs a nonlinear scaling transformation on the time coordinates of the segment to be corrected along the time-axis mapping path, aligning the transformed segment with the reference segment on the time axis.

[0028] Figure 2 The data shows the multi-cycle power consumption timing data of the CNC-07 milling machine before and after the phase alignment process in step 1.

[0029] Step 2: Extract high-order statistical features from the same phase point for each period segment after phase alignment, and generate a cross-cycle energy consumption feature vector sequence for each phase point.

[0030] Specifically, for each periodic segment after phase normalization, at each discrete phase point Higher-order statistical features across periods are extracted. These higher-order statistical features include kurtosis. skewness and autocorrelation decay coefficient The calculation method is as follows: Kudo Used to describe the sharpness of energy consumption distribution:

[0031] in, Phase point Average energy consumption across cycles at a given location.

[0032] Skewness Used to describe the symmetry of energy consumption distribution:

[0033] Autocorrelation decay coefficient The decay rate used to describe the correlation of energy consumption between adjacent cycles is obtained by analyzing the phase points. Cross-period energy consumption sequence The autocorrelation function under different time delays was calculated and obtained by fitting an exponential decay model, where... , , These represent the phase points of the 1st, 2nd, and Nth cycles, respectively. Energy consumption value at the location.

[0034] It should be noted that the aforementioned exponential decay model is fitted with a sequence of autocorrelation function values ​​under different time delays as input, and outputs the decay coefficient. The fitting process of the exponential decay model assumes that the autocorrelation function decays exponentially with time delay, that is, the relationship between the autocorrelation function value and the time delay conforms to the form of an exponential function. The decay coefficient parameter of the exponential function is estimated by the least squares method or a nonlinear fitting algorithm.

[0035] The above features are combined to form the cross-cycle energy consumption feature vector for each phase point. Where T represents transpose, generating a sequence of cross-cycle energy consumption feature vectors for each phase point. ,in , , Let represent the cross-cycle energy consumption characteristic vectors of the 1st, 2nd, and Mth phase points, respectively. This represents the total number of discrete phase points.

[0036] Furthermore, discrete phase points In the standard phase interval The discrete phase points are obtained by sampling at equal intervals. The interval between adjacent phase points is determined by the time resolution requirements of the process cycle. It equals the number of sampling points within a single period.

[0037] It should be noted that the selection of the above-mentioned higher-order statistical features can be adjusted according to the specific equipment type and process characteristics, and may also include other indicators describing the pattern of energy consumption fluctuations, such as the spectral entropy of energy consumption fluctuations and multi-scale sample entropy.

[0038] Figure 3 The distribution of cross-cycle higher-order statistical characteristics (kurtosis, skewness, autocorrelation decay coefficient) at each phase point is shown.

[0039] Step 3: Calculate the trend slope and phase synchronization index of the cross-cycle feature vector sequence of each phase point, and screen out the set of abnormal phase points.

[0040] Specifically, for each phase point For the cross-period eigenvector sequence, calculate the slope of the trend of each eigencomponent as it changes with the period number. Taking kurtosis as an example, the trend slope is determined by analyzing the sequence. The result was obtained by linear regression fitting, where , , These represent the phase points of the 1st, 2nd, and Nth cycles, respectively. The kurtosis value at that location.

[0041] It should be noted that the input for the aforementioned linear regression fitting is a periodic index vector. and the corresponding eigenvalue sequence The output is the slope of the linearly fitted line. And intercept. The linear regression fitting process uses the least squares method to determine the line parameters that minimize the squared error between the actual feature values ​​and the predicted values ​​of the fitted line, including the slope. This indicates the rate of change of the characteristic value as the period number increases.

[0042] Furthermore, for an eigenvector containing multiple feature components, the trend slope of each feature component is calculated separately, and the one with the largest absolute value is taken as the comprehensive trend slope of that phase point. This is to capture significant trends in any feature dimension.

[0043] Simultaneously, the correlation coefficient of eigenvalues ​​at points of the same phase during adjacent cycles is calculated as an indicator of phase synchronization. :

[0044] in, and They are respectively the front Each cycle and after Each cycle at the phase point The average energy consumption at that location.

[0045] Filter out trend slope Exceeding the preset trend threshold And phase synchronization Higher than the preset synchronization threshold The phase points form a set of anomalous phase points. This enables precise tracking and location of abnormal power consumption patterns.

[0046] It should be noted that the aforementioned trend threshold and synchronization threshold can be determined based on the statistical distribution of historical normal operation data, or they can be dynamically adjusted according to the current data characteristics using an adaptive threshold method.

[0047] Furthermore, trend threshold The synchronization threshold can be set to the average slope of the trend at each phase point during historical normal operation, plus a certain number of standard deviations. It can be set to a positive value to filter out phase points with positive correlation synchronization characteristics, phase synchronization degree As the correlation coefficient, its value range is: .

[0048] Figure 4 The two-dimensional distribution of the trend slope and phase synchronization index at each phase point is shown.

[0049] Step 4: Encode the energy consumption feature vector sequence of the abnormal phase point set, along with the process stage identifier and equipment operating parameters, into structured descriptive text, and input it into the encoding layer of the large model to generate the equipment energy consumption status semantic embedding vector.

[0050] Specifically, for the set of anomalous phase points Each phase point in Obtain the corresponding process stage identifier (e.g., machining, idling, standby) and current equipment operating parameters (e.g., spindle speed, feed rate, coolant flow rate). Then, use the energy consumption feature vector... Process stage identifiers and equipment operating parameters are encoded into structured descriptive text according to a predefined template format. .

[0051] Structured description text The input to the encoding layer of the pre-trained large model, after word embedding and position encoding, is processed by a multi-layer Transformer encoder to generate a semantic embedding vector of the device energy consumption state at the anomalous phase point. ,in For the embedded vector dimension.

[0052] It should be noted that the aforementioned pre-trained large model employs a Transformer encoder architecture. The input is a vector sequence of structured descriptive text transformed by a word embedding layer, and the output is a fixed-dimensional semantic embedding vector. The Transformer encoder includes a multi-head self-attention sub-layer and a feedforward neural network sub-layer. The multi-head self-attention sub-layer performs weighted aggregation on each position of the input sequence, while the feedforward neural network sub-layer performs a non-linear transformation on the aggregated representation. Residual connections and layer normalization operations are used between the sub-layers. This pre-trained large model does not require training in this application; its pre-trained language understanding capabilities are directly utilized to encode structured text in the energy consumption domain.

[0053] It should be noted that the template format of the above structured description text should include a textual expression of energy consumption numerical characteristics, a semantic description of the process stage, and normalized values ​​of equipment parameters, with specific delimiters used to distinguish between the parts.

[0054] In this embodiment of the application, in order to enhance the ability of semantic embedding vectors to represent abnormal energy consumption patterns, the structured description text can be preprocessed for domain adaptation before encoding the large model, mapping the professional terms in the energy consumption field to semantically similar general words in the large model vocabulary, or the ability of the large model to understand energy consumption data can be enhanced through domain adaptation fine-tuning.

[0055] Step 5: Load the current semantic embedding vector and the baseline embedding vector set in the context window of the large model, calculate the semantic distance distribution using the attention mechanism, amplify the weight of the abnormal dimension, and generate energy consumption anomaly feature labels.

[0056] Specifically, obtain the set of baseline embedding vectors from the device's historical high-efficiency operating states: ,in , , These represent the 1st, 2nd, and Rth reference embedding vectors, respectively. The baseline sample size is used. The semantic embedding vector of the current anomalous phase point is... The baseline embedding vector set is loaded into the context window of the large model at the same time.

[0057] By utilizing the multi-head attention mechanism of a large model, the attention weight matrix between the current embedding vector and each vector in the benchmark set is calculated, thereby obtaining the semantic distance distribution between the current embedding and the benchmark set. ,in , , These represent the semantic distances between the current embedding and the 1st, 2nd, and Rth reference embeddings, respectively.

[0058] It should be noted that the aforementioned multi-head attention mechanism includes a query transformation layer, a key transformation layer, and a value transformation layer. The inputs are the current embedding vector and the set of reference embedding vectors, and the outputs are an attention weight matrix and a weighted semantic representation. The query transformation layer maps the current embedding vector to a query vector, while the key transformation layer and the value transformation layer map the reference embedding vectors to key vectors and value vectors, respectively. The attention weights of the current embedding to each reference embedding are obtained by calculating and normalizing the dot product of the query vector and each key vector. The value vectors are then weighted and summed based on these attention weights to obtain a semantic representation that incorporates reference information. The semantic distance is calculated using the Euclidean distance or cosine distance between the current embedding vector and the weighted semantic representation. This multi-head attention mechanism does not require training in this application; it directly utilizes the parameters already present in the pre-trained large model for inference computation.

[0059] Anomaly dimensions in the semantic distance distribution that deviate from the baseline mean by more than a preset standard deviation are weighted and amplified to generate weighted anomalous feature vectors. These anomalous feature vectors are then matched with predefined anomaly type templates to generate phase-correlated energy consumption anomaly feature labels. The tag content includes the anomaly type, anomaly severity level, and associated feature dimension index.

[0060] Furthermore, the aforementioned weight amplification process is achieved by multiplying the feature values ​​of the outlier dimension by an amplification factor. The amplification factor is positively correlated with the degree to which the dimension deviates from the baseline mean; the greater the deviation, the greater the amplification factor, thus enabling the outlier dimension to have higher discriminative power in the subsequent matching process.

[0061] Furthermore, the above-mentioned anomaly type template matching is achieved by calculating the cosine similarity between the weighted anomalous feature vector and each predefined template vector. The anomaly type corresponding to the template with the highest similarity is selected as the matching result, and the similarity value is also used to determine the anomaly degree level.

[0062] Furthermore, the aforementioned preset standard deviation multiple is used to define the judgment boundary of abnormal dimensions. When the difference between the semantic distance value of a certain dimension and the mean semantic distance of the benchmark set exceeds the preset standard deviation multiple of the standard deviation of the semantic distance of the benchmark set, the dimension is judged as an abnormal dimension and its weight is amplified.

[0063] It should be noted that the above-mentioned set of baseline embedding vectors should be updated regularly to reflect the slow drift of the energy consumption baseline during the normal aging process of the equipment, so as to avoid misjudging normal aging as abnormal.

[0064] Figure 5 This demonstrates the semantic distance distribution between the semantic embedding vector of the current anomalous phase point and the set of baseline embedding vectors.

[0065] Step 6: Associate and match the energy consumption anomaly feature tags with the adjustable parameter range of the equipment and the energy-saving control rule library, input them into the large model for self-driven reasoning, and generate the energy consumption anomaly cause diagnosis results.

[0066] Specifically, obtain a list of adjustable parameter ranges for the equipment (including the upper and lower limits and adjustment steps for each parameter) and an energy-saving control rule base (containing historically accumulated parameter adjustment rules and their applicable conditions). Label abnormal energy consumption characteristics. The adjustable parameter range and energy-saving control rule base are encoded as prompt text, which is then input into the large model for self-driven reasoning.

[0067] Based on its pre-trained knowledge and input context information, the large model autonomously analyzes the causes of energy consumption growth corresponding to abnormal phase points, identifies the device parameter offsets associated with these causes, and generates diagnostic results for the causes of energy consumption anomalies. The diagnostic results include the inferred category of the abnormality, the confidence score of the cause, a list of associated parameter offsets, and the offset direction of each parameter.

[0068] It should be noted that the aforementioned energy-saving control rule base can be organized in the form of a knowledge graph, which can structurally express the relationship between anomaly types, equipment parameters and adjustment strategies, making it easier for large models to perform correlation reasoning.

[0069] Step 7: Based on the cause diagnosis results, use the large model to autonomously generate equipment parameter adjustment schemes, and encode the adjustment schemes into an executable energy-saving control instruction sequence.

[0070] Specifically, the diagnostic results of abnormal energy consumption will be used to... As input, and combining the current operating status of the equipment with the constraints of adjustable parameters, the large model autonomously generates equipment parameter adjustment schemes for each abnormal phase point. The adjustment scheme includes the name of the parameter to be adjusted, the target adjustment value, the adjustment timing (corresponding process phase interval), and the expected energy-saving effect.

[0071] Furthermore, the process phase range corresponding to the above adjustment timing is based on the abnormal phase point within the standard phase range. Once the position within is determined, a certain phase range is extended forward and backward from the abnormal phase point to form the phase interval where the adjustment takes effect. The parameter adjustment command is triggered and executed when the device enters this phase interval.

[0072] The equipment parameter adjustment schemes output by the large model are subjected to structured parsing. Using predefined parameter name matching rules and numerical extraction modes, the target adjustment values ​​of each parameter are identified and extracted from the adjustment scheme text. The extracted results are then converted into parameter-value key-value pairs. The parsed equipment parameter adjustment schemes are then encoded into an executable energy-saving control command sequence according to the equipment control interface protocol. ,in , , They represent Article 1, Article 2, and Article 3 respectively. Energy conservation management instructions The total number of instructions, each instruction It includes the target device identifier, parameter adjustment command code, and adjustment parameter value.

[0073] It should be noted that after the above-mentioned energy-saving control instruction sequence is generated, it can be checked by the safety valve verification module to verify whether the parameter values ​​of each instruction are within the safe range allowed by the equipment, so as to avoid generating instructions that may cause equipment damage or safety accidents.

[0074] In this embodiment of the application, in order to improve the execution effect of control commands, the large model can generate operation mode optimization suggestions at the same time when generating adjustment schemes, including system-level optimization measures such as equipment start-up and shutdown timing adjustment and load balancing distribution, forming a multi-level energy-saving control strategy with parameter-level adjustments.

[0075] In addition to step 7, the following steps are also included: Step 8: Send the energy-saving control command sequence to the equipment controller for execution, collect the power consumption data after execution, calculate the energy consumption deviation and feed it back to the large model for adaptive optimization of the control strategy.

[0076] Specifically, the energy-saving control instruction sequence generated in step 7 The command is sent to the corresponding device controller for execution via an industrial communication protocol. During the execution of the command and within the observation window after execution, the device's power consumption data is continuously collected.

[0077] Furthermore, the duration of the aforementioned observation window is determined based on the length of the equipment's process cycle. The observation window should cover at least one complete process cycle to ensure that the assessment of energy consumption changes includes the complete process and avoids assessment bias caused by incomplete observation periods.

[0078] Calculate the deviation between the actual change in energy consumption and the expected energy saving target. ,in This represents the actual change in energy consumption after the instruction is executed. The expected energy savings predicted by the large model. Incorporating bias information. The corresponding instruction execution records are fed back into the context memory of the large model for reference and adaptive optimization when generating subsequent control strategies.

[0079] Furthermore, the aforementioned changes in actual energy consumption The energy consumption is obtained by comparing the cumulative energy consumption within the same time period before and after the instruction is executed. The cumulative energy consumption within the observation window before the instruction is executed is used as the baseline value, and the cumulative energy consumption within the same length observation window after the instruction is executed is used as the control value. The difference between the two is the actual change in energy consumption.

[0080] It should be noted that the above feedback information can be aggregated according to time windows to form a strategy effectiveness evaluation report, including statistical indicators such as the average energy saving effect of various adjustment strategies and the success rate of strategies, providing historical experience reference for the generation of strategies for large models.

[0081] In this embodiment of the application, in order to accelerate the optimization and convergence of the control strategy, a strategy effect prediction model can be established to pre-evaluate the expected effect of the newly generated strategy based on historical feedback data, select the strategy with better expected effect and execute it first, thereby reducing the trial and error cost of inefficient strategies.

[0082] It should be noted that the aforementioned strategy effect prediction model employs a feedforward neural network architecture. The input layer receives the concatenated result of the strategy feature vector and the device state feature vector, while the output layer outputs a single scalar value representing the expected energy-saving effect score. The strategy feature vector includes information such as the magnitude of change of each adjustment parameter and the timing of adjustment, while the device state feature vector includes information such as the current device operating parameters and historical energy consumption trends. Since the components in the strategy feature vector and the device state feature vector have different physical dimensions and numerical ranges, Z-score standardization is performed on each component before inputting it into the neural network to eliminate the impact of dimensional differences on model training and prediction. The strategy effect prediction model is trained using supervised learning, with training samples derived from historically executed strategies and their corresponding actual energy-saving effects. The loss function is the mean squared error function, and the model parameters are optimized using a gradient descent algorithm to minimize the error between the predicted energy-saving effect score and the actual effect.

[0083] Figure 6 The study demonstrates the energy consumption comparison at different process stages before and after the implementation of energy-saving control instructions.

[0084] This implementation maps the energy consumption time sequence of each process cycle to a unified phase coordinate system through phase normalization processing, allowing direct comparison and superposition of energy consumption values ​​at the same process phase point in different cycles. Because phase normalization eliminates the time axis differences between cycles, micro-disturbance signals that were originally too small to be detected in a single cycle exhibit a recognizable cumulative trend after being superimposed across multiple cycles. This overcomes the problem that single-cycle detection cannot track the growth of hidden energy consumption due to disturbance amplitudes being below the normal fluctuation range.

[0085] This implementation refines the analytical granularity of anomaly detection to the single-phase-point level by extracting higher-order statistical features at the same phase point and calculating the phase synchronization index. Because higher-order statistical features can capture subtle changes in energy consumption distribution patterns, and phase synchronization can identify systematic correlations over multiple cycles, it avoids the problem of anomaly patterns being overwhelmed by mixing and averaging different phase signals in long-term statistical analysis.

[0086] This implementation utilizes the attention mechanism of a large model to selectively amplify the weights of anomalous dimensions in the semantic distance distribution. Because the attention mechanism can adaptively focus on feature dimensions that deviate from the baseline distribution, it achieves accurate tracking and localization of energy consumption anomaly patterns, improving the accuracy of anomalous feature labels.

[0087] This implementation method autonomously generates cause diagnosis and energy-saving control instructions through large-scale model self-driven reasoning. Because the large-scale model can comprehensively utilize pre-trained knowledge and contextual information to make autonomous reasoning decisions, it overcomes the limitations of traditional methods that rely on manual analysis and manual parameter adjustment, and realizes an automated closed loop from energy consumption tracking to energy-saving control.

[0088] This implementation method continuously collects the strategy execution effect through a closed-loop feedback mechanism and feeds it back to the large model. Because the feedback information provides a reference for the actual effect of subsequent strategy generation, it realizes adaptive optimization of energy-saving management strategy, thereby completing the self-driven closed-loop management of power consumption tracking and energy-saving control.

[0089] Figure 7 The study demonstrated the relationship between energy consumption anomaly detection and control strategies for 12 CNC milling machines within the industrial park.

[0090] A precision machining park has deployed 12 five-axis CNC milling machines for the batch processing of aerospace parts. Each machine follows a fixed process flow, executing a machining cycle of "rapid traverse positioning → rough milling → finish milling → tool retraction and reset," with a single cycle time of approximately 180 seconds. The park's energy management system continuously collects the power consumption data of each machine at a sampling frequency of 100Hz.

[0091] In step 1, taking the CNC milling machine numbered CNC-07 as an example, the system collects the power consumption data of the machine for 50 consecutive process cycles. After determining the cycle boundaries through the process start / stop signals output by the equipment control system, the power consumption time sequence is divided into 50 cycle segments, each containing 18,000 sampling points. After phase normalization processing of each cycle segment, a multi-cycle power consumption segment sequence with phase alignment is generated. The power consumption sampling values ​​at key phase points of some cycle segments are shown in the table below: Table 1. Multi-cycle power consumption samples after phase alignment (unit: kW):

[0092] In step 2, for each periodic segment after phase normalization, at the discrete phase point Extract high-order statistical features across cycles at the midpoint of the fine milling stage.

[0093] The average energy consumption across cycles at this phase point was calculated. kW, peak skewness Autocorrelation attenuation coefficient The above features are combined to form the cross-cycle energy consumption feature vector for this phase point. .

[0094] In step 3, for the phase point The trend slope and phase synchronization index are calculated from the cross-cycle eigenvector sequence. After linear regression fitting, the comprehensive trend slope at this phase point is... Phase synchronization Set a trend threshold. Synchronization threshold ,because and Phase point The candidate phase points were included in the set of anomalous phase points. The selection results for each candidate phase point are shown in the table below: Table 2. Results of screening candidate phase points based on trend slope and phase synchronization:

[0095] In step 5, for abnormal phase points The semantic distance distribution is calculated between the semantic embedding vector and the baseline embedding vector set. The baseline set contains 20 baseline samples from the device's historical high-efficiency operating conditions, with a semantic distance distribution mean of 0.12 and a standard deviation of 0.03. The current embedding vector has a semantic distance value of 0.28 in the 47th dimension, which deviates from the baseline mean by 5.3 times the standard deviation. This dimension is identified as an anomaly and its weight is amplified. Anomaly type template matching is then used to generate energy consumption anomaly feature labels. The tag content is "Abnormal type: Spindle transmission efficiency decrease; Abnormality level: moderate; Related feature dimension index: 47".

[0096] In step 6, the energy consumption anomaly feature tags are associated and matched with the adjustable parameter range of the equipment and the energy-saving control rule base, and then input into the large model for self-driven inference. Based on the input context information, the large model generates the energy consumption anomaly cause diagnosis result: the anomaly cause category is "spindle bearing preload offset", the cause confidence score is 0.82, the associated parameter offset item is "spindle speed compensation coefficient", and the offset direction is positive and high.

[0097] In step 7, the large model autonomously generates a device parameter adjustment plan based on the causal diagnosis results.

[0098] The adjustment plan is as follows: the spindle speed compensation coefficient will be reduced from the current value of 1.05 to 0.98, and the adjustment will be made during the phase interval. (Corresponding to the precision milling stage), the expected energy saving effect is 0.35 kWh per cycle. The adjustment plan will be coded as an energy-saving control instruction. The instructions are shown in the table below: Table 3 Energy Saving Control Instructions:

[0099] In step 8, the energy-saving control command will be... The command is sent to the CNC-07 equipment controller for execution, and power consumption data is continuously collected within the observation window after execution (covering 5 complete process cycles). The energy consumption comparison before and after execution is shown in the table below: Table 4 compares energy consumption before and after instruction execution:

[0100] Calculate the actual energy consumption change kWh, expected energy saving kWh, deviation kWh. Deviation information and instruction execution records are fed back into the context memory of the large model for adaptive optimization during subsequent control strategy generation.

[0101] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for self-driven power consumption tracking and energy-saving management using a large AI model, characterized in that, Includes the following steps: The system acquires energy consumption sampling data from multiple consecutive process cycles of industrial equipment, segments the energy consumption time sequence according to the process cycle boundary, performs phase normalization processing on each cycle segment, and generates a phase-aligned multi-cycle energy consumption segment sequence. High-order statistical features are extracted from the same phase point for each period segment after phase alignment, and a cross-cycle energy consumption feature vector sequence for each phase point is generated. Calculate the trend slope and phase synchronization index for the cross-cycle feature vector sequence of each phase point, and screen out the phase points whose trend slope exceeds the trend threshold and whose phase synchronization is higher than the synchronization threshold to form an abnormal phase point set. The energy consumption feature vector sequence of the abnormal phase point set, along with the process stage identifier and equipment operating parameters, is encoded into structured descriptive text and input into the encoding layer of the large model to generate a semantic embedding vector of equipment energy consumption status. Load the current semantic embedding vector and the benchmark embedding vector set in the context window of the large model, calculate the semantic distance distribution using the attention mechanism, amplify the weights of the abnormal dimensions that deviate from the benchmark, and generate energy consumption anomaly feature labels. The energy consumption anomaly feature tags are associated and matched with the adjustable parameter range of the equipment and the energy-saving control rule library, and then input into the large model for self-driven reasoning to generate energy consumption anomaly cause diagnosis results. Based on the causal diagnosis results, a large model is used to autonomously generate equipment parameter adjustment schemes, and the adjustment schemes are encoded into an executable energy-saving control instruction sequence.

2. The AI ​​large-scale model self-driven power consumption tracking and energy-saving management method according to claim 1, characterized in that, The higher-order statistical features include kurtosis, skewness, and autocorrelation decay coefficient; Kurtosis is obtained by calculating the ratio of the energy consumption value at the same phase point in each cycle to the square of the fourth central moment and the second central moment of the average energy consumption across cycles at that phase point. The skewness is obtained by calculating the ratio of the energy consumption value at the same phase point in each cycle to the 1.5 power of the third-order central moment and the second-order central moment of the average energy consumption across cycles at the phase point; The autocorrelation decay coefficient is obtained by calculating the autocorrelation function under different time delays on the cross-cycle energy consumption sequence at the phase point and fitting an exponential decay model.

3. The AI ​​large-scale model self-driven power consumption tracking and energy-saving management method according to claim 1, characterized in that, The phase synchronization index is obtained by calculating the correlation coefficient of energy consumption values ​​at points of the same phase during adjacent cycles, specifically: The energy consumption value sequence of the first N-1 cycles at the target phase point is compared with the energy consumption value sequence of the next N-1 cycles at the same phase point using Pearson correlation calculation. The resulting correlation coefficient is used as the phase synchronization index of the phase point, where N is the total number of process cycles.

4. The AI ​​large-scale model self-driven power consumption tracking and energy-saving management method according to claim 1, characterized in that, The phase normalization process includes: The time coordinates of each period segment are mapped to a standard phase interval, so that each period segment has a unified phase coordinate system; A dynamic time warping algorithm is used to perform nonlinear correction on the time axis of each period segment, eliminating the difference in period length caused by fluctuations in equipment operating speed.

5. The AI ​​large-scale model self-driven power consumption tracking and energy-saving management method according to claim 1, characterized in that, The methods for determining the process cycle boundary include: Calibration is performed based on the process start / stop signals output by the equipment control system; Based on the periodic characteristics of energy consumption time series data, the period length is identified through autocorrelation analysis and then segmented.

6. The AI ​​large-scale model self-driven power consumption tracking and energy-saving management method according to claim 1, characterized in that, The calculation of semantic distance distribution using the attention mechanism includes: The attention weight matrix between the current embedded vector and each vector in the baseline set is calculated using the multi-head attention mechanism of the large model. The semantic distance between the current embedding vector and each vector in the benchmark set is calculated based on the attention weight matrix, forming a semantic distance distribution; Dimensions in the semantic distance distribution that deviate from the baseline mean by more than a preset standard deviation are weighted and amplified.

7. The AI ​​large-scale model self-driven power consumption tracking and energy-saving management method according to claim 1, characterized in that, The diagnostic results for the causes of abnormal energy consumption include: The inferred anomaly cause category, cause confidence score, list of associated parameter offsets, and offset direction of each parameter.

8. The AI ​​large-scale model self-driven power consumption tracking and energy-saving management method according to claim 1, characterized in that, Each instruction in the energy-saving control instruction sequence includes a target device identifier, a parameter adjustment instruction code, and an adjustment parameter value; The equipment parameter adjustment plan includes the name of the parameter to be adjusted, the target adjustment value, the timing of the adjustment, and the expected energy-saving effect.

9. The AI ​​large-scale model self-driven power consumption tracking and energy-saving management method according to any one of claims 1 to 8, characterized in that, Also includes: Send the energy-saving control command sequence to the equipment controller for execution; Collect energy consumption data after execution and calculate the deviation between the actual energy consumption change and the expected energy saving target; The deviation information is fed back into the context memory of the large model for adaptive optimization during the subsequent generation of control strategies.

10. A self-driven power consumption tracking and energy-saving management system for large AI models, used to execute the self-driven power consumption tracking and energy-saving management method for large AI models as described in any one of claims 1 to 9, characterized in that, include: The data acquisition and phase normalization module is used to acquire multi-cycle power consumption sampling data of industrial equipment, divide it according to the process cycle boundary and perform phase normalization processing to generate a phase-aligned multi-cycle power consumption segment sequence. The feature extraction module is used to extract high-order statistical features from each period segment at the same phase point and generate a cross-period energy consumption feature vector sequence. The abnormal phase point screening module is used to calculate the trend slope and phase synchronization index of each phase point and screen them to form a set of abnormal phase points; The semantic embedding generation module is used to encode the energy consumption characteristics and process information of abnormal phase points into structured descriptive text, and generate semantic embedding vectors through the large model encoding layer. The anomaly feature label generation module is used to calculate the semantic distance distribution using the large model attention mechanism and amplify the weight of the anomaly dimension to generate energy consumption anomaly feature labels. The cause diagnosis module is used to associate and match energy consumption anomaly feature tags with the energy-saving control rule base, and generate corresponding cause diagnosis results through large-scale model self-driven reasoning. The control instruction generation module is used to generate equipment parameter adjustment schemes based on the cause diagnosis results using a large model and encode them into an energy-saving control instruction sequence.

Images