Methods, model training methods, and systems for detecting abnormal energy consumption in industrial settings.

Abstract

This disclosure provides a method, model training method, and system for detecting energy consumption anomalies in industrial scenarios. The method includes: acquiring several sets of sample energy consumption training data for target energy-consuming objects in a target industrial scenario; the sample energy consumption training data includes sample energy consumption feature data and corresponding sample energy consumption anomaly annotation data; training a preset random projection model based on the sample energy consumption feature data in each set of sample energy consumption training data as input and the corresponding sample energy consumption anomaly annotation data as output to obtain an energy consumption anomaly detection model; wherein, the preset random projection model uses a random hyperplane for recursive segmentation. In this disclosure, by comprehensively considering multi-dimensional energy consumption correlation data in the scenario, fused features are obtained through a stacked autoencoder network, and then combined with the preset random projection model for joint modeling, to achieve efficient and accurate detection of energy efficiency anomalies in the target industrial scenario, enabling timely intervention and processing of energy consumption anomalies.

Description

Technical Field

[0001] This disclosure relates to the field of data analysis technology, and in particular to a method, model training method and system for detecting abnormal energy consumption in industrial scenarios. Background Technology

[0002] With the development of industrial intelligence, more and more enterprises are relying on advanced data analysis technologies to improve production efficiency and reduce energy consumption. In the energy efficiency management of workshop equipment, energy consumption anomaly assessment is particularly crucial. Energy consumption anomalies not only affect a company's operating costs but may also indicate equipment failures or systemic problems. Therefore, accurate identification and analysis of these anomalies are of great significance for production optimization and preventative maintenance.

[0003] Currently, energy consumption analysis of industrial equipment mostly relies on traditional monitoring methods based on simple statistical approaches and rules. While these methods can monitor equipment operation to some extent, they often cannot handle complex, multi-dimensional production data. When faced with multi-type, multi-parameter equipment operating data in modern workshops, the limitations of traditional methods become increasingly apparent, especially when energy efficiency fluctuations are small, making it difficult to guarantee the accuracy and stability of anomaly detection.

[0004] Equipment energy consumption is influenced by a variety of factors, including equipment operating status, production processes, environmental conditions, and operator behavior. Existing analytical methods often rely on single-dimensional monitoring or prediction, such as depending solely on historical energy consumption data or data from a fixed period. This approach struggles to comprehensively grasp the root causes of equipment energy efficiency fluctuations, potentially missing underlying systemic problems during energy efficiency anomaly detection. For instance, the actual energy efficiency of equipment may be closely related to the complex relationships between its operating status and production process parameters, which cannot be accurately identified by relying solely on energy consumption data.

[0005] Energy consumption data from industrial equipment typically exhibits strong time dependence. Short-term fluctuations can arise from changes in production plans, equipment status, or environmental conditions. Therefore, traditional static analysis methods often struggle to capture these dynamic changes over time. The processing and analysis of time-series data has become a key factor in improving the accuracy of energy efficiency assessments. Traditional methods cannot utilize trend information from historical data for prediction. While machine learning-based detection methods have made some progress, current energy anomaly detection technologies generally suffer from limitations in comprehensively and accurately assessing abnormal fluctuations in equipment energy consumption, resulting in low detection accuracy and failing to meet the needs of real-world scenarios. Summary of the Invention

[0006] The technical problem to be solved by this disclosure is to overcome the above-mentioned defects in the prior art, and the purpose is to provide a method, model training method and system for detecting abnormal energy consumption in industrial scenarios.

[0007] This disclosure aims to address the limitations of traditional detection methods in assessing equipment energy consumption anomalies. By comprehensively considering the equipment's operating status, energy efficiency indicators, time dimension, and analytical dimensions (such as production processes and environmental factors), and combining deep neural networks for multi-dimensional data joint modeling, it achieves efficient and accurate identification of equipment energy efficiency anomalies, enabling a comprehensive and accurate assessment of abnormal fluctuations in equipment energy consumption. The implementation scheme of this disclosure overcomes the shortcomings of traditional analysis methods that rely solely on single-dimensional data, revealing the root causes of equipment energy efficiency anomalies more comprehensively, improving the accuracy and stability of anomaly detection, and is applicable to fields such as intelligent manufacturing, equipment maintenance, and energy efficiency optimization, providing enterprises with more precise energy efficiency management and production optimization solutions. Specifically:

[0008] This disclosure solves the above-mentioned technical problems through the following technical solution:

[0009] This disclosure provides a training method for an energy consumption anomaly detection model, the training method comprising:

[0010] Acquire several sets of sample energy consumption training data of the target energy-consuming object in the target industrial scenario within a historical preset time period;

[0011] Each set of sample energy consumption training data includes sample energy consumption feature data and corresponding sample energy consumption anomaly annotation data;

[0012] The sample energy consumption feature data in each set of sample energy consumption training data is used as input, and the corresponding sample energy consumption anomaly labeling data is used as output. The preset random projection model is trained to obtain the energy consumption anomaly detection model.

[0013] In the preset random projection model, a random hyperplane is used for recursive segmentation.

[0014] Optionally, the step of training a preset random projection model based on the sample energy consumption feature data in each group of sample energy consumption training data as input and the corresponding sample energy consumption anomaly labeling data as output to obtain the energy consumption anomaly detection model includes:

[0015] A predetermined number of the sample energy consumption training data are randomly selected to form a training subset;

[0016] Using the training subset as the root node, recursively construct a binary tree;

[0017] A preset projection vector is randomly generated as the random hyperplane;

[0018] The projection value of the energy consumption training data of each sample at the current node onto the random hyperplane is calculated.

[0019] Within the projection range determined based on the projection value, a segmentation point is randomly selected;

[0020] Based on the projection range and the position of the segmentation point, the corresponding sample energy consumption training data is divided into the left child node or the right child node.

[0021] In response to the current node meeting the preset conditions, the current node is marked as a leaf node, and the recursive segmentation process is repeated until all nodes are marked as leaf nodes to obtain the energy consumption anomaly detection model.

[0022] Optionally, the preset conditions include at least one of the following: the data depth exceeds a predetermined depth, the current node contains only one sample energy consumption training data, and the sample energy consumption feature data of all the sample energy consumption training data in the current node are completely identical;

[0023] And / or,

[0024] The target industrial scenario comprises an N-layer four-dimensional industrial environment; where N ≥ 2 and is a positive integer;

[0025] And / or,

[0026] The target industrial scenario includes a three-layer, four-dimensional industrial environment. The target energy consumption object includes at least one of the workshop, production line, and equipment in the industrial environment. The different scenario levels in the industrial environment include at least two of the workshop layer, industry layer, and equipment layer.

[0027] Optionally, the step of obtaining several sets of sample energy consumption training data of the target energy-consuming object in the target industrial scenario within a historical preset time period includes:

[0028] Obtain several sets of first historical energy consumption association data of the target energy consumption object in the target industrial scenario within the historical preset time period;

[0029] Based on the first historical energy consumption correlation data, second historical energy consumption correlation data of different scenario levels in the target industrial scenario are obtained;

[0030] The second historical energy consumption correlation data is trained separately to obtain a trained single-layer autoencoder for each of the aforementioned scene levels;

[0031] The output of the previous single-layer autoencoder is used as the input of the next single-layer autoencoder. Multi-layer structure fusion training is performed on different scene levels to train a stacked autoencoder network.

[0032] Obtain the sample output results of the autoencoder network, and use the sample output results as the sample energy consumption feature data corresponding to the first historical energy consumption association data; wherein, the sample energy consumption feature data is the fusion feature corresponding to different scene levels.

[0033] Optionally, the step of obtaining several sets of first historical energy consumption association data of the target energy consumption object in the target industrial scenario within the historical preset time period includes:

[0034] Obtain the initial historical energy consumption association data of the target energy consumption object in the target industrial scenario within the preset historical time period;

[0035] Based on the defined operating conditions, intermediate historical energy consumption correlation data of the target energy consumption object in the target industrial scenario under each operating condition within the preset historical time period are obtained from the initial historical energy consumption correlation data.

[0036] The intermediate historical energy consumption correlation data under each operating condition is processed using a matching data standardization method to obtain several sets of the first historical energy consumption correlation data.

[0037] And / or,

[0038] Before the step of training the second historical energy consumption correlation data separately to obtain the trained single-layer autoencoder for each scene level, the method further includes:

[0039] Construct energy consumption-derived features;

[0040] The preset derived features include at least one of energy efficiency intensity features, transient features, and sequence features;

[0041] The step of training the second historical energy consumption correlation data separately to obtain a trained single-layer autoencoder for each scene level includes:

[0042] Based on the energy consumption derived features and the second historical energy consumption associated data, a single-layer autoencoder trained for each of the scene levels is obtained.

[0043] Optionally, after the step of obtaining the initial historical energy consumption association data of the target energy consumption object in the target industrial scenario within the historical preset time period, the method further includes:

[0044] Based on the time dimension, the missing values ​​in the initial historical energy consumption association data are filled using a matching preset filling method to obtain the processed initial historical energy consumption association data.

[0045] The preset filling method includes interpolation filling method or preset value filling method, wherein the preset value in the preset value filling method includes factory value or average value;

[0046] And / or,

[0047] The local anomaly factor algorithm is used to remove outliers from the initial historical energy consumption correlation data to obtain the processed initial historical energy consumption correlation data.

[0048] And / or,

[0049] Redundant fields are removed from the initial historical energy consumption association data to obtain the processed initial historical energy consumption association data.

[0050] A second aspect of this disclosure provides a method for detecting abnormal energy consumption in industrial settings, the detection method comprising:

[0051] Obtain preset energy consumption correlation data of the target energy consumption object in the target industrial scenario within the target preset time period;

[0052] Obtain the target energy consumption feature data corresponding to the preset energy consumption correlation data;

[0053] The target energy consumption feature data is input into a pre-trained energy consumption anomaly detection model to obtain the target energy consumption anomaly detection result;

[0054] The energy consumption anomaly detection model is obtained using the training method of the energy consumption anomaly detection model described in the first aspect;

[0055] The target energy consumption anomaly detection result is used to characterize whether there is an anomaly in the actual energy consumption status of the target industrial scenario within the target preset time period.

[0056] A third aspect of this disclosure provides a training system for an energy consumption anomaly detection model, the training system comprising:

[0057] The sample training data acquisition module is used to acquire several sets of sample energy consumption training data of the target energy consumption object in the target industrial scenario within a historical preset time period;

[0058] Each set of sample energy consumption training data includes sample energy consumption feature data and corresponding sample energy consumption anomaly annotation data;

[0059] The model training module is used to train a preset random projection model based on the sample energy consumption feature data in each group of sample energy consumption training data as input and the corresponding sample energy consumption anomaly labeling data as output, so as to obtain the energy consumption anomaly detection model.

[0060] In the preset random projection model, a random hyperplane is used for recursive segmentation.

[0061] Optionally, the model training module includes:

[0062] A random construction unit is used to randomly select a preset number of the sample energy consumption training data to form a training subset; and a binary tree is recursively constructed with the training subset as the root node.

[0063] A random hyperplane generation unit is used to randomly generate a preset projection vector as the random hyperplane.

[0064] The projection value calculation unit is used to calculate the projection value of each sample energy consumption training data at the current node on the random hyperplane.

[0065] The segmentation point selection unit is used to randomly select a segmentation point within a projection range determined based on the projection value;

[0066] A partitioning unit is used to determine whether the corresponding sample energy consumption training data is partitioned into a left child node or a right child node based on the projection range and the position of the partitioning point.

[0067] The model acquisition unit is used to mark the current node as a leaf node in response to the current node meeting the preset conditions, and to repeat the recursive segmentation process until all nodes are marked as leaf nodes, so as to obtain the energy consumption anomaly detection model.

[0068] Optionally, the preset conditions include at least one of the following: the data depth exceeds a predetermined depth, the current node contains only one sample energy consumption training data, and the sample energy consumption feature data of all the sample energy consumption training data in the current node are completely identical;

[0069] And / or,

[0070] The target industrial scenario comprises an N-layer four-dimensional industrial environment; where N ≥ 2 and is a positive integer;

[0071] And / or,

[0072] The target industrial scenario includes a three-layer, four-dimensional industrial environment. The target energy consumption object includes at least one of the workshop, production line, and equipment in the industrial environment. The different scenario levels in the industrial environment include at least two of the workshop layer, industry layer, and equipment layer.

[0073] Optionally, the sample training data acquisition module includes: a first association data acquisition unit, used to acquire several sets of first historical energy consumption association data of the target energy consumption object in the target industrial scenario within the historical preset time period;

[0074] The second associated data acquisition unit is used to acquire second historical energy consumption associated data at different scenario levels in the target industrial scenario based on the first historical energy consumption associated data;

[0075] A single-layer training unit is used to train the second historical energy consumption correlation data separately to obtain a trained single-layer autoencoder for each of the scene levels.

[0076] The stacked training unit is used to take the output of the previous single-layer autoencoder as the input of the next single-layer autoencoder, and perform multi-layer structure fusion training on different scene levels to train the stacked autoencoder network.

[0077] The energy consumption feature data acquisition module is used to acquire the sample output results of the autoencoder network and use the sample output results as the sample energy consumption feature data corresponding to the first historical energy consumption association data; wherein, the sample energy consumption feature data is a fusion feature corresponding to different scene levels.

[0078] Optionally, the first associated data acquisition unit is used to:

[0079] Obtain the initial historical energy consumption association data of the target energy consumption object in the target industrial scenario within the preset historical time period;

[0080] Based on the defined operating conditions, intermediate historical energy consumption correlation data of the target energy consumption object in the target industrial scenario under each operating condition within the preset historical time period are obtained from the initial historical energy consumption correlation data.

[0081] The intermediate historical energy consumption correlation data under each operating condition is processed using a matching data standardization method to obtain several sets of the first historical energy consumption correlation data.

[0082] And / or, the training system further includes:

[0083] The feature derivation module is used to construct energy consumption-derived features;

[0084] The preset derived features include at least one of energy efficiency intensity features, transient features, and sequence features;

[0085] The single-layer training unit is used to train independently based on the energy consumption derived features and the second historical energy consumption associated data to obtain the trained single-layer autoencoder for each scene level.

[0086] Optionally, the training system further includes:

[0087] The filling processing module is used to fill the missing values ​​of the data in the initial historical energy consumption association data sequentially based on the time dimension using a matching preset filling method to obtain the processed initial historical energy consumption association data.

[0088] The preset filling method includes interpolation filling method or preset value filling method, wherein the preset value in the preset value filling method includes factory value or average value;

[0089] And / or,

[0090] The outlier removal module is used to perform outlier removal processing on the initial historical energy consumption correlation data using the local outlier factor algorithm to obtain the processed initial historical energy consumption correlation data.

[0091] And / or,

[0092] The field processing module is used to delete redundant fields from the initial historical energy consumption association data to obtain the processed initial historical energy consumption association data.

[0093] In a fourth aspect, this disclosure provides a detection system for abnormal energy consumption in industrial settings, the detection system comprising:

[0094] The preset data acquisition module is used to acquire preset energy consumption associated data of the target energy consumption object in the target industrial scenario within the target preset time period;

[0095] The target feature acquisition module is used to acquire target energy consumption feature data corresponding to the preset energy consumption correlation data;

[0096] The target detection result acquisition module is used to input the target energy consumption feature data into a pre-trained energy consumption anomaly detection model to obtain the target energy consumption anomaly detection result;

[0097] The energy consumption anomaly detection model is obtained using the training system for the energy consumption anomaly detection model as described in the third aspect.

[0098] The target energy consumption anomaly detection result is used to characterize whether there is an anomaly in the actual energy consumption status of the target industrial scenario within the target preset time period.

[0099] A fifth aspect of this disclosure provides an electronic device, including a memory, a processor, and a computer program stored in the memory and for running on the processor, characterized in that the processor executes the computer program to implement a training method for an energy consumption anomaly detection model as described in the first aspect; or a method for detecting energy consumption anomalies in an industrial scenario as described in the second aspect.

[0100] A sixth aspect of this disclosure provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a training method for an energy consumption anomaly detection model as described in the first aspect; or a method for detecting energy consumption anomalies in an industrial scenario as described in the second aspect.

[0101] A seventh aspect of this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements a training method for an energy consumption anomaly detection model as described in the first aspect; or a method for detecting energy consumption anomalies in an industrial scenario as described in the second aspect.

[0102] Based on common knowledge in the field, the above-mentioned preferred conditions can be combined arbitrarily to obtain various preferred embodiments of this disclosure.

[0103] The positive and progressive effects of this disclosure are as follows:

[0104] This disclosure achieves efficient and accurate detection of energy efficiency anomalies in target industrial scenarios by comprehensively considering multi-dimensional energy consumption correlation data and combining it with a pre-designed, optimized random projection model for joint modeling. This enables timely intervention and processing of energy consumption anomalies, comprehensively and accurately assessing abnormal fluctuations in equipment energy consumption. It effectively overcomes the shortcomings of traditional detection methods that rely on only single-dimensional data, and can more comprehensively reveal the root causes of equipment energy efficiency anomalies, improving the accuracy and stability of energy consumption anomaly detection. This makes it well-suited for fields such as intelligent manufacturing, equipment maintenance, and energy efficiency optimization, providing enterprises with more precise energy efficiency management and production optimization solutions. Attached Figure Description

[0105] Figure 1 This is a flowchart of the training method for the energy consumption anomaly detection model in Embodiment 1 of this disclosure;

[0106] Figure 2 This is a schematic diagram of the hierarchical architecture of a three-layer, four-dimensional workshop scene according to Embodiment 1 of this disclosure;

[0107] Figure 3 This is a schematic diagram showing the distribution of energy consumption in different dimensions of a three-layer, four-dimensional workshop scene according to Embodiment 1 of this disclosure;

[0108] Figure 4 This is a first flowchart of the training method for the energy consumption anomaly detection model of Embodiment 2 of this disclosure;

[0109] Figure 5 This is a second flowchart of the training method for the energy consumption anomaly detection model in Embodiment 2 of this disclosure;

[0110] Figure 6 This is the third flowchart of the training method for the energy consumption anomaly detection model in Embodiment 2 of this disclosure;

[0111] Figure 7 This is a flowchart of the method for detecting abnormal energy consumption in an industrial setting according to Embodiment 3 of this disclosure;

[0112] Figure 8 This is a schematic diagram of the modules of the training system for the energy consumption anomaly detection model of Embodiment 4 of this disclosure;

[0113] Figure 9 This is a schematic diagram of the modules of the training system for the energy consumption anomaly detection model in Embodiment 5 of this disclosure;

[0114] Figure 10 This is a schematic diagram of the first module of the energy consumption anomaly detection system in an industrial scenario according to Embodiment 6 of this disclosure;

[0115] Figure 11 This is a schematic diagram of the second module of the energy consumption anomaly detection system in an industrial scenario according to Embodiment 6 of this disclosure;

[0116] Figure 12 This is a schematic diagram of the structure of the electronic device according to Embodiment 7 of this disclosure. Detailed Implementation

[0117] The present disclosure is further illustrated below by way of embodiments, but the present disclosure is not limited to the scope of the embodiments described herein.

[0118] The prefixes such as "first" and "second" used in this disclosure are merely for distinguishing different descriptive objects and do not limit the position, order, priority, quantity, or content of the described objects. The use of ordinal numbers and other prefixes used to distinguish descriptive objects in this disclosure does not constitute a limitation on the described objects. The description of the described objects is given in the claims or the context of the embodiments, and should not be construed as an unnecessary limitation. Furthermore, in the description of this embodiment, unless otherwise stated, "multiple" means two or more.

[0119] Example 1

[0120] like Figure 1 As shown, the training method for the energy consumption anomaly detection model in this embodiment includes:

[0121] S101. Obtain several sets of sample energy consumption training data of the target energy consumption object in the target industrial scene within a historical preset time period;

[0122] Each set of sample energy consumption training data includes sample energy consumption feature data and corresponding sample energy consumption anomaly annotation data;

[0123] In a feasible solution, the target industrial scenario comprises an N-layer, four-dimensional industrial environment; where N ≥ 2 and is a positive integer;

[0124] For example, the target industrial scenario is a three-layer, four-dimensional industrial environment. The target energy consumption objects include workshops, production lines, and equipment within the industrial environment. Different scenario layers within this industrial environment are the workshop layer, the industrial layer, and the equipment layer, etc. Figure 2 The diagram shows a three-layer architecture of a three-dimensional workshop scenario; the workshop layer corresponds to the workshop; the production line layer corresponds to production line A and production line B; and the equipment layer corresponds to equipment A, equipment B, equipment C, and equipment D.

[0125] Among them, the sample energy consumption characteristic data are the characteristic data corresponding to the sample energy consumption correlation data. The sample energy consumption correlation data includes data on the equipment's operating status, energy efficiency indicators, time dimension, and analysis dimension (such as production process and environmental factors).

[0126] Specifically, energy consumption-related data from various production lines and equipment in the factory workshop are obtained, such as key energy consumption data, production information, and auxiliary information. Among them, auxiliary information includes historical operating status data, monitoring data (such as voltage, current, vibration, etc.), equipment parameters (such as rated power, energy efficiency, etc.), and environmental information (such as temperature, humidity, etc.).

[0127] Specifically, such as Figure 3 As shown, some key energy consumption data in a three-layer, four-dimensional workshop scenario include state-level data (such as energy consumption in shutdown state, maintenance / fault state, standby state, and operating state), energy efficiency index-level data (such as energy consumption per unit product, energy consumption on non-production days, and energy consumption per unit of working hours), and analysis-level data (such as month-on-month, year-on-year, and trend data).

[0128] S102. Based on the sample energy consumption feature data in the energy consumption training data of each group of samples as input and the corresponding sample energy consumption anomaly labeling data as output, train the preset random projection model to obtain the energy consumption anomaly detection model.

[0129] In the preset random projection model, a random hyperplane is used for recursive segmentation.

[0130] Specifically, for situations where energy consumption data from different dimensions are highly correlated and there is information coupling between features, the iForest (isolated forest) random projection model is used to model and identify energy consumption anomalies. During training, the axis parallel segmentation of the original iForest is replaced by a random hyperplane to accurately capture the complex relationships between features, thereby effectively improving the accuracy and reliability of the energy consumption anomaly detection results output by the model.

[0131] In this embodiment, by comprehensively considering multi-dimensional energy consumption correlation data in the target industrial scenario and combining it with a pre-designed optimized random projection model for joint modeling, efficient and accurate detection of energy efficiency anomalies in the target industrial scenario is achieved. This allows for timely intervention and processing of energy consumption anomalies, and can comprehensively and accurately assess abnormal fluctuations in equipment energy consumption. This effectively overcomes the shortcomings of traditional detection methods that rely on only single-dimensional data, and can more comprehensively reveal the root causes of equipment energy efficiency anomalies, improving the accuracy and stability of energy consumption anomaly detection. As a result, it is well-suited for fields such as intelligent manufacturing, equipment maintenance, and energy efficiency optimization, providing enterprises with more precise energy efficiency management and production optimization solutions.

[0132] Example 2

[0133] The training method of the energy consumption anomaly detection model in this embodiment is a further improvement on Embodiment 1, specifically:

[0134] In a feasible solution, such as Figure 4 As shown, step S101 includes:

[0135] S1011. Obtain several sets of first historical energy consumption association data of the target energy consumption object in the target industrial scenario within a historical preset time period;

[0136] S1012. Based on the first historical energy consumption correlation data, obtain the second historical energy consumption correlation data of different scenario levels in the target industrial scenario;

[0137] S1013. Train the second historical energy consumption correlation data separately to obtain a trained single-layer autoencoder for each scene level.

[0138] S1014. The output of the previous single-layer autoencoder is used as the input of the next single-layer autoencoder. Multi-layer structure fusion training is performed on different scene levels to train the stacked autoencoder network.

[0139] S1015. Obtain the sample output results of the autoencoder network and use the sample output results as the sample energy consumption feature data corresponding to the first historical energy consumption association data; wherein, the sample energy consumption feature data is the fusion feature corresponding to different scene levels.

[0140] For example, by using historical energy consumption parameters such as equipment operating status, energy efficiency indicators, timestamps, year-on-year and month-on-month analysis values ​​as input parameters, an SAE (Sparse Autoencoder) model can be constructed, which is a stacked autoencoder network. Specifically:

[0141] The stacked autoencoder network is a stacked autoencoder, which consists of multiple autoencoder layers stacked together. Each layer is trained separately, and the training process follows an unsupervised learning strategy. Initialize the encoder and decoder: initialize the encoder and decoder weights of each layer, where a restricted Boltzmann machine is used to initialize the weights.

[0142] Specifically, in combination Figure 2 The energy consumption structure of the equipment in the workshop is trained layer by layer to obtain a single-layer autoencoder. The number of stacked autoencoder layers is set according to the scene level. Assuming that there is an energy consumption unit N1 at the equipment level, an energy consumption unit N2 at the production line level, and an energy consumption unit N3 at the workshop level, the constructed stacked autoencoder has three layers, and the number of neurons in each layer corresponds to N1, N2, and N3, respectively. During the layer-by-layer training process, the labels of these neurons correspond to the actual energy consumption values ​​of the corresponding energy consumption units.

[0143] Encoding stage: Input data is mapped by the encoder to generate hidden layer representations; Decoding stage: The hidden layer representations are reconstructed by the decoder to reconstruct the input data; Loss function: The reconstruction error is calculated, usually using MSE (mean squared error) as the loss function; Gradient descent update: The parameters (weights and biases) of the encoder and decoder are updated through the backpropagation algorithm; Training each layer: The autoencoder is trained layer by layer, and each layer uses the output of the previous layer as input for training;

[0144] The encoder parts of each trained single-layer autoencoder are stacked to form a deeper network structure, and the encoder output of each layer is used as the input of the next layer; training multi-layer structure: after stacking the single-layer autoencoders layer by layer, a multi-layer autoencoder network is obtained, and the output of the last layer of its encoding process is the feature vector set obtained in the feature fusion stage.

[0145] In this disclosure, we address the problem that various factors influence equipment energy consumption, and the lack of prior knowledge often makes it difficult to extract key features and indicators for identifying anomalies. This paper proposes a stacked autoencoder network to obtain fused features. Each sample contains energy consumption data, production information, and operating status data, which are input into the stacked autoencoder network to output fused features after multi-dimensional feature fusion. This allows for dimensionality reduction and automatic feature extraction of the samples, enabling more accurate correlation between the extracted features and equipment energy consumption anomalies. This ensures the reliability of subsequent model training and further reduces the model's misjudgment rate.

[0146] In one feasible implementation, prior to step S1012, the training method further includes:

[0147] Construct energy consumption-derived features;

[0148] Among them, the pre-defined derived features include energy efficiency intensity features, transient features, and sequence features;

[0149] Specifically, the energy efficiency intensity characteristics, transient characteristics, and sequence characteristics correspond to the following:

[0150] Energy efficiency intensity characteristics (energy efficiency index): Calculates the energy consumption per unit of output (energy intensity), which is one of the core indicators for measuring efficiency;

[0151] Transient characteristics: Extracting startup peaks, startup time, and energy consumption curve slope during equipment startup and shutdown; among them, it is mainly considered that many early faults are not obvious under steady state, but will be exposed during transient processes.

[0152] Sequence features: Statistical features (mean, variance, slope, kurtosis, etc.) within the window are calculated using a sliding window, as well as frequency domain features (the main frequency energy is extracted through Fast Fourier Transform (FFT)). These features can better characterize the operating mode of the device.

[0153] like Figure 5 As shown, step S1013 includes:

[0154] S10131. Based on energy consumption derived features and second historical energy consumption correlation data, train separately to obtain a trained single-layer autoencoder for each scene level.

[0155] In this scheme, considering the limited information of the original energy consumption sequence, which makes it difficult to fully learn and fuse features through machine learning, more discriminative energy consumption derivative features are calculated and constructed based on industry knowledge and other factors. These constructed energy consumption derivative features are then combined with other second historical energy consumption correlation data to train each scene level separately to obtain a single-layer autoencoder. This ensures the accuracy and reliability of the single-layer autoencoder, thereby ensuring the accuracy and reliability of the stacked autoencoder network after subsequent fusion training. This enables efficient and high-quality acquisition of sample energy consumption feature data corresponding to different scene levels.

[0156] In a feasible solution, such as Figure 5 As shown, step S1011 includes:

[0157] S10111. Obtain the initial historical energy consumption association data of the target energy consumption object in the target industrial scenario within a preset historical time period;

[0158] S10112. Based on the defined operating conditions, obtain the intermediate historical energy consumption correlation data of the target energy consumption object in the target industrial scenario under each operating condition within the historical preset time period from the initial historical energy consumption correlation data.

[0159] For example, based on information such as equipment speed and output, the data can be divided into operating conditions such as "no load", "half load", and "full load" to obtain the data corresponding to each operating condition.

[0160] S10113. The intermediate historical energy consumption correlation data under each operating condition is processed using a matching data standardization method to obtain several sets of first historical energy consumption correlation data.

[0161] In this solution, considering the strong correlation between equipment energy consumption and its load rate, directly normalizing the entire sequence would obscure the energy consumption patterns under different operating conditions, thereby reducing the accuracy of the subsequently trained model. Based on this, this problem is solved by constructing a data standardization method for operating condition awareness. By standardizing the data in each operating condition separately, the model can learn the normal energy consumption benchmark under each steady state during subsequent model training, making it more sensitive to anomalies that deviate from the benchmark, thereby further reducing the model's misjudgment rate.

[0162] In one feasible solution, the initial historical energy consumption correlation data can be divided by a preset time window (e.g., 2s) to obtain the first historical energy consumption correlation data corresponding to each time window. For example, taking 2s as a time window, the data can be sliced ​​using this time window. Each time window corresponds to one sample, and a total of 3762 non-overlapping samples are obtained.

[0163] In one feasible implementation, after step S10111, the training method further includes:

[0164] Based on the time dimension, the missing values ​​in the initial historical energy consumption correlation data are filled using a matching preset filling method to obtain the processed initial historical energy consumption correlation data.

[0165] Among them, the preset filling method includes interpolation filling method, preset value filling method, etc. The preset value in the preset value filling method includes factory value or average value.

[0166] For example, missing values ​​in energy consumption and production data can be interpolated and filled sequentially along the time dimension. The corresponding filling formula is as follows:

[0167] ;

[0168] in, Let t be the energy consumption or output value at time t. This represents the energy consumption or output value at time t+1. This represents the energy consumption or output value at time t-1. This is the specific timestamp at time t. This is the specific timestamp at time t+1. This is the specific timestamp at time t-1.

[0169] For time series data (voltage, current, vibration, temperature and humidity, etc.), the above-mentioned interpolation filling method can also be used; or other filling methods can be used, as long as they can be effectively implemented, there are no restrictions on this.

[0170] For relatively stable data such as equipment parameters, use factory default values ​​or average values ​​for filling; the corresponding filling formula is as follows:

[0171] ;

[0172] ;

[0173] in, Let be the device parameters at time t. The average value of the equipment parameters within a set time period. These are the factory default values ​​corresponding to the equipment parameters. In this embodiment, the initial historical energy consumption correlation data is preprocessed by filling it in, and different types of data are processed using matching filling methods. This ensures the rationality and quality of the data, thereby guaranteeing the accuracy and reliability of the subsequent model and reducing the false alarm rate of the model.

[0174] In one feasible implementation, after step S10111, the training method further includes:

[0175] The local anomaly factor algorithm is used to remove outliers from the initial historical energy consumption correlation data to obtain the processed initial historical energy consumption correlation data.

[0176] The calculation formula for the local anomaly factor algorithm is as follows:

[0177] ;

[0178] in, For data points Local anomalous factors, For data points of Neighbor, For data points Locally achievable density, For data points The locally achievable density.

[0179] In this solution, considering that energy consumption data is often subject to increased metering errors due to load changes and voltage fluctuations, directly using such data for model training could easily lead to unreliable models and misjudgments. This problem is addressed by employing the Local Outlier Factor (LOF) algorithm to remove outliers. Specifically, the LOF is calculated for each data point, measuring its sparsity relative to its neighborhood. Each data point is assigned an outlier factor dependent on neighborhood density to determine if it is an outlier. This ensures the accuracy and reliability of the data, improves data quality, and consequently guarantees the accuracy of the subsequent model, thus ensuring the reliability of energy consumption anomaly detection results and reducing the false alarm rate, better meeting the needs of real-world scenarios.

[0180] In one feasible implementation, after step S10111, the training method further includes:

[0181] Redundant fields are removed from the initial historical energy consumption correlation data to obtain processed initial historical energy consumption correlation data. In this solution, by removing redundant fields (such as redundant fixed-value fields) from the initial historical energy consumption correlation data, unnecessary redundant data is automatically removed, improving data quality, ensuring the accuracy of subsequent model training data, and reducing the false positive rate of the model. This reduces the amount of training data required, thereby reducing the computational load of the model training process and improving its efficiency. It should be noted that one or more of the above-mentioned data preprocessing methods, such as padding, outlier removal, and redundant field removal, can be used. Of course, other preprocessing operations can also be used according to actual needs, as long as they improve data quality; there are no specific restrictions.

[0182] In a feasible solution, such as Figure 6 As shown, step S102 includes:

[0183] S1021. Randomly select a preset number of sample energy consumption training data to form a training subset;

[0184] S1022. Recursively construct a binary tree with the training subset as the root node;

[0185] S1023. Randomly generate a preset projection vector as a random hyperplane;

[0186] S1024. Calculate the projection value of the energy consumption training data of each sample on the random hyperplane for the current node.

[0187] S1025. Within the projection range determined based on the projection value, randomly select a segmentation point;

[0188] S1026. Based on the projection range and the position of the segmentation point, determine whether the corresponding sample energy consumption training data is divided into the left child node or the right child node.

[0189] S1027. In response to the current node meeting the preset conditions, the current node is marked as a leaf node, and the recursive segmentation process is repeated until all nodes are marked as leaf nodes to obtain the energy consumption anomaly detection model.

[0190] The preset conditions include the following: the data depth exceeds the predetermined depth, the current node contains only one sample energy consumption training data, and the sample energy consumption feature data of all sample energy consumption training data in the current node are completely identical.

[0191] In this scheme, for situations where energy consumption data of different dimensions are highly correlated and there is information coupling between features, the random projection model iForest is used to model and identify energy consumption anomalies. Specifically, the fused features output by the stacked autoencoder network are used to train the random preset projection model. During the training process, the axis parallel segmentation of the original iForest is replaced by a random hyperplane to accurately capture the complex relationships between features, thereby effectively improving the accuracy and reliability of the energy consumption anomaly detection results output by the model.

[0192] Specifically, the process of constructing a pre-defined random projection model, determining parameters such as the number of trees, the number of samples for training each tree, and the dimension of the random projection, and the specific training process to obtain the pre-defined random projection model is as follows:

[0193] For each iTree in the forest, perform the following operations:

[0194] (1) Subsampling: ψ samples are randomly selected from the training set X without replacement to form a subset X';

[0195] (2) Constructing tree nodes: Starting with subset X' as the root node, recursively construct the binary tree;

[0196] (3) Recursive partitioning process: The termination condition is that if the current node meets any of the following conditions, it is marked as a leaf node and the partitioning stops: The data depth exceeds the predetermined height limit. (usually set as) The current node contains only one sample, and all samples in the current node have the same feature values.

[0197] On random projection and segmentation

[0198] Generate a random hyperplane: Randomly generate a d-dimensional projection vector v; for example, each component is independently drawn from a standard normal distribution N(0,1); wherein, optionally, the projection vector v is normalized to make it a unit vector.

[0199] Calculate the projection value: For each sample x at the current node, calculate its projection value on the random hyperplane: ;

[0200] Selecting a split point: Randomly select a split point s within the range of the projected value p; for example, randomly select between min(p) and max(p);

[0201] Split nodes: Project values The samples are assigned to the left child node; the projected values ​​are... The samples are assigned to the right child node;

[0202] Repeated recursion: Repeat step (3) recursively for the left or right child node until all nodes become leaf nodes to obtain the preset random projection model;

[0203] In real-world energy consumption anomaly detection scenarios, a pre-trained, pre-defined random projection model is deployed to the real-time data stream to detect anomalies in the collected energy consumption data.

[0204] For a new sample x, calculate its anomaly score:

[0205] Calculate the path length: Let sample x traverse every iTree in the forest; on each tree, starting from the root node, decide whether to go to the left subtree or the right subtree based on the random hyperplane v and the split point s on the node, until the leaf node is reached; record the number of edges that sample x traverses on the tree, which is called the path length h(x);

[0206] Standardized path length: Calculates the average path length of sample x across all trees. ; Use the following standardized formula to convert it into anomaly scores s(x,ψ):

[0207] ;

[0208] in, ψ is the average path length under the sample, used for standardization, and its approximation is... H is the harmonic number.

[0209] Interpretation of results: A score close to 1 indicates a high probability of an anomaly; a score much less than 0.5 indicates a high probability of a normal range; a score close to 0.5 indicates no obvious abnormal characteristics.

[0210] The following example (such as a key energy-consuming device in a factory's blade workshop) illustrates the implementation principle of the training method for the energy consumption anomaly detection model in this embodiment:

[0211] (1) Obtain energy consumption related data (key data) from various workshops, production lines and equipment in the factory, including key energy consumption data, production information, auxiliary information, etc.; among which, auxiliary information includes historical data of operating status, monitoring data (such as energy consumption, voltage, current, vibration), equipment parameters (such as rated power, energy efficiency), environmental information (such as temperature, humidity) and historical abnormal records, etc.

[0212] (2) Perform preprocessing and cleaning of key data. The main preprocessing steps include filling missing values, removing a small number of outliers by using the local outlier factor LOF, and deleting redundant fixed value fields.

[0213] (3) A 2s time window is used to divide the preprocessed data to obtain several different overlapping sample historical energy correlation data;

[0214] Alternatively, based on the different operating conditions determined by the division, data under different operating conditions can be obtained, and the data under each operating condition can be standardized to obtain several samples of historical energy consumption correlation data.

[0215] Among them, some pre-defined derived features can be constructed according to actual needs, such as energy efficiency indicators, transient features, and sequence features. The data corresponding to these constructed pre-defined derived features can be used as historical energy consumption correlation data of the sample.

[0216] (4) Construct a stacked autoencoder network, and input the energy consumption data, production information, and operating status data contained in the historical energy consumption association data of each sample into the stacked autoencoder network, and use the fused features after multi-dimensional feature fusion, i.e., sample energy consumption feature data.

[0217] (5) A preset random projection model is trained based on the sample energy consumption feature data. During the training process, the axis parallel segmentation of the original iForest is replaced by a random hyperplane.

[0218] Example 3

[0219] The method for detecting abnormal energy consumption in industrial scenarios in this embodiment is based on the above embodiment 1 or 2.

[0220] like Figure 7 As shown, the method for detecting abnormal energy consumption in an industrial scenario in this embodiment includes:

[0221] S201. Obtain the preset energy consumption associated data of the target energy consumption object in the target industrial scenario within the target preset time period;

[0222] S202. Obtain the target energy consumption characteristic data corresponding to the preset energy consumption correlation data;

[0223] The process involves preprocessing the preset energy consumption-related data (such as filling in missing values ​​and removing outliers) and then inputting it into an autoencoder network with stacked values ​​for multi-dimensional feature fusion processing to obtain the target energy consumption feature data.

[0224] S203. Input the target energy consumption feature data into the pre-trained energy consumption anomaly detection model to obtain the target energy consumption anomaly detection result;

[0225] The energy consumption anomaly detection model is obtained using the training method of the energy consumption anomaly detection model in Embodiment 1 or 2 above;

[0226] The target energy consumption anomaly detection results are used to characterize whether there are any anomalies in the actual energy consumption status of the target industrial scenario within the target preset time period.

[0227] In this embodiment, in a real-world energy consumption anomaly detection scenario, a pre-trained, pre-defined random projection model is deployed into the real-time data stream. This model is used to detect anomalies in the real-time collected energy consumption data, ensuring the accuracy and efficiency of energy consumption anomaly detection in industrial scenarios. It effectively overcomes the shortcomings of traditional detection methods that rely solely on single-dimensional data, revealing the root causes of equipment energy efficiency anomalies more comprehensively and improving the accuracy and stability of energy consumption anomaly detection. This makes it well-suited for fields such as intelligent manufacturing, equipment maintenance, and energy efficiency optimization, providing enterprises with more precise energy efficiency management and production optimization solutions.

[0228] In one feasible solution, step S203 is followed by:

[0229] In response to the target energy consumption anomaly detection result characterization based on the currently input preset energy consumption correlation data, if an energy consumption anomaly is detected, alarm information is generated to promptly remind relevant personnel to check, maintain or repair it in a timely manner.

[0230] This solution enables relevant personnel to be promptly aware of any abnormal energy consumption, allowing for timely detection and resolution of these anomalies, thereby ensuring the stability and reliability of the entire target industrial environment.

[0231] In one feasible solution, step S203 is followed by:

[0232] In response to the target energy consumption anomaly detection result representation based on the currently input preset energy consumption correlation data, if an energy consumption anomaly is detected, an alarm message is generated, and a matching energy consumption anomaly handling strategy is generated.

[0233] Among them, the energy consumption anomaly handling strategy can be automatically determined based on the pre-built mapping relationship, or it can be implemented based on the pre-trained model, including large models; there is no restriction on which method is used, as long as it can be implemented.

[0234] This solution enables relevant personnel to be promptly aware of any abnormal energy consumption, and facilitates their intervention based on automatically generated energy consumption anomaly handling strategies. This allows energy consumption anomalies to be detected and resolved in a timely manner, further optimizing the energy consumption anomaly detection process in the entire industrial scenario and ensuring the stability and reliability of the entire target industrial environment.

[0235] Example 4

[0236] like Figure 8 As shown, the training system for the energy consumption anomaly detection model in this embodiment includes:

[0237] Sample training data acquisition module 1 is used to acquire several sets of sample energy consumption training data of the target energy consumption object in the target industrial scenario within a historical preset time period.

[0238] Each set of sample energy consumption training data includes sample energy consumption feature data and corresponding sample energy consumption anomaly annotation data;

[0239] Model training module 2 is used to train a preset random projection model based on the sample energy consumption feature data in each group of sample energy consumption training data as input and the corresponding sample energy consumption anomaly labeling data as output, so as to obtain an energy consumption anomaly detection model.

[0240] In the preset random projection model, a random hyperplane is used for recursive segmentation.

[0241] In this embodiment, by comprehensively considering multi-dimensional energy consumption correlation data in the target industrial scenario and combining it with a pre-designed optimized random projection model for joint modeling, efficient and accurate detection of energy efficiency anomalies in the target industrial scenario is achieved. This allows for timely intervention and processing of energy consumption anomalies, and can comprehensively and accurately assess abnormal fluctuations in equipment energy consumption. This effectively overcomes the shortcomings of traditional detection methods that rely on only single-dimensional data, and can more comprehensively reveal the root causes of equipment energy efficiency anomalies, improving the accuracy and stability of energy consumption anomaly detection. As a result, it is well-suited for fields such as intelligent manufacturing, equipment maintenance, and energy efficiency optimization, providing enterprises with more precise energy efficiency management and production optimization solutions.

[0242] Example 5

[0243] like Figure 9 As shown, the training system for the energy consumption anomaly detection model in this embodiment is a further improvement on embodiment 4, specifically:

[0244] In one feasible embodiment, the model training module 2 includes:

[0245] A randomized construction unit is used to randomly select a preset number of sample energy consumption training data to form a training subset; a binary tree is recursively constructed with the training subset as the root node.

[0246] The random hyperplane generation unit is used to randomly generate a preset projection vector as a random hyperplane.

[0247] The projection value calculation unit is used to calculate the projection value of the energy consumption training data of each sample on the current node onto the random hyperplane.

[0248] The segmentation point selection unit is used to randomly select a segmentation point within a projection range determined based on the projection value;

[0249] The partitioning unit is used to determine whether the corresponding sample energy consumption training data is partitioned into the left or right child node based on the projection range and the position of the partitioning point.

[0250] The model acquisition unit is used to mark the current node as a leaf node in response to the current node meeting the preset conditions, and to repeat the recursive segmentation process until all nodes are marked as leaf nodes, so as to obtain the energy consumption anomaly detection model.

[0251] In an feasible solution, the preset conditions include at least one of the following: the data depth exceeds a predetermined depth, the current node contains only one sample energy consumption training data, and the sample energy consumption feature data of all sample energy consumption training data in the current node are completely identical.

[0252] In a feasible solution, the target industrial scenario comprises an N-layer, four-dimensional industrial environment; where N ≥ 2 and is a positive integer;

[0253] In an feasible solution, the target industrial scenario includes a three-layer, four-dimensional industrial environment, and the target energy consumption objects include at least one of workshops, production lines, and equipment in the industrial environment. The different scenario levels in the industrial environment include at least two of the workshop layer, industry layer, and equipment layer.

[0254] In one feasible solution, the sample training data acquisition module 1 includes: a first associated data acquisition unit, used to acquire several sets of first historical energy consumption associated data of the target energy consumption object in the target industrial scenario within a historical preset time period;

[0255] The second associated data acquisition unit is used to acquire second historical energy consumption associated data at different scenario levels in the target industrial scenario based on the first historical energy consumption associated data;

[0256] A single-layer training unit is used to train the second historical energy consumption correlation data separately to obtain a trained single-layer autoencoder for each scene level.

[0257] Stacked training units are used to take the output of the previous single-layer autoencoder as the input of the next single-layer autoencoder, and perform multi-layer structure fusion training on different scene levels to train the stacked autoencoder network.

[0258] The energy consumption feature data acquisition module is used to acquire the sample output results of the autoencoder network and use the sample output results as the sample energy consumption feature data corresponding to the first historical energy consumption association data; wherein, the sample energy consumption feature data are fusion features corresponding to different scene levels.

[0259] In one feasible solution, the first associated data acquisition unit is used for:

[0260] Obtain the initial historical energy consumption correlation data of the target energy consumption object in the target industrial scenario within a preset historical time period;

[0261] Based on the defined operating conditions, intermediate historical energy consumption correlation data of the target energy consumption object in the target industrial scenario under each operating condition within the historical preset time period are obtained from the initial historical energy consumption correlation data.

[0262] The intermediate historical energy consumption correlation data under each operating condition are processed using a matching data standardization method to obtain several sets of first historical energy consumption correlation data.

[0263] In one feasible embodiment, the training system also includes:

[0264] Feature Derivation Module 3 is used to construct energy consumption derived features;

[0265] Among them, the preset derived features include at least one of energy efficiency intensity features, transient features, and sequence features;

[0266] Single-layer training units are used to train individually based on energy consumption-derived features and second historical energy consumption correlation data to obtain a trained single-layer autoencoder for each scene level.

[0267] In one feasible embodiment, the training system also includes:

[0268] The filling processing module 4 is used to fill the missing values ​​in the initial historical energy consumption association data according to the time dimension, using a matching preset filling method, so as to obtain the processed initial historical energy consumption association data.

[0269] Among them, the preset filling method includes interpolation filling method or preset value filling method, and the preset value in the preset value filling method includes factory value or average value;

[0270] In one feasible embodiment, the training system also includes:

[0271] The outlier removal processing module 5 is used to perform outlier removal processing on the initial historical energy consumption correlation data using the local outlier factor algorithm to obtain the processed initial historical energy consumption correlation data.

[0272] In one feasible embodiment, the training system also includes:

[0273] Field processing module 6 is used to delete redundant fields from the initial historical energy consumption association data to obtain the processed initial historical energy consumption association data.

[0274] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs.

[0275] Example 6

[0276] like Figure 10 As shown, the energy consumption anomaly detection system in the industrial scenario of this embodiment includes:

[0277] The preset data acquisition module 7 is used to acquire preset energy consumption associated data of the target energy consumption object in the target industrial scenario within the target preset time period;

[0278] Target feature acquisition module 8 is used to acquire target energy consumption feature data corresponding to preset energy consumption correlation data;

[0279] The target detection result acquisition module 9 is used to input the target energy consumption feature data into the pre-trained energy consumption anomaly detection model to obtain the target energy consumption anomaly detection result;

[0280] The energy consumption anomaly detection model is obtained using the training system of the energy consumption anomaly detection model as described in Example 4 or 5.

[0281] The target energy consumption anomaly detection results are used to characterize whether there are any anomalies in the actual energy consumption status of the target industrial scenario within the target preset time period.

[0282] In this embodiment, in a real-world energy consumption anomaly detection scenario, a pre-trained, pre-defined random projection model is deployed into the real-time data stream. This model is used to detect anomalies in the real-time collected energy consumption data, ensuring the accuracy and efficiency of energy consumption anomaly detection in industrial scenarios. It effectively overcomes the shortcomings of traditional detection methods that rely solely on single-dimensional data, revealing the root causes of equipment energy efficiency anomalies more comprehensively and improving the accuracy and stability of energy consumption anomaly detection. This makes it well-suited for fields such as intelligent manufacturing, equipment maintenance, and energy efficiency optimization, providing enterprises with more precise energy efficiency management and production optimization solutions.

[0283] In a feasible solution, such as Figure 11 As shown, the detection system also includes:

[0284] The first anomaly handling module 10 is used to respond to the target energy consumption anomaly detection result characterization based on the currently input preset energy consumption correlation data, detect the existence of energy consumption anomalies, generate alarm information, and promptly remind relevant personnel to check, maintain or repair in a timely manner.

[0285] In a feasible solution, such as Figure 11 As shown, the detection system also includes:

[0286] The second anomaly handling module 11 is used to respond to the target energy consumption anomaly detection result characterization based on the currently input preset energy consumption correlation data, detect the existence of energy consumption anomalies, generate alarm information, and generate a matching energy consumption anomaly handling strategy.

[0287] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs.

[0288] Example 7

[0289] Figure 12 This is a schematic diagram of the structure of an electronic device according to an example embodiment of the present disclosure. The electronic device includes a memory, a processor, and a computer program stored in the memory and used to run on the processor. When the processor executes the computer program, it implements the method described in any of the above embodiments. Figure 12 The electronic device 120 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0290] like Figure 12 As shown, the electronic device 120 can be manifested as a general-purpose computing device, such as a server device. The components of the electronic device 120 may include, but are not limited to: at least one processor 121, at least one memory 122, and a bus 123 connecting different system components (including memory 122 and processor 121).

[0291] Bus 123 includes a data bus, an address bus, and a control bus.

[0292] The memory 122 may include volatile memory, such as random access memory (RAM) 1221 and / or cache memory 1222, and may further include read-only memory (ROM) 1223.

[0293] The memory 122 may also include a program tool 1225 (or utility) having a set (at least one) program module 1224, such program module 1224 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0294] The processor 121 performs various functional applications and data processing, such as the methods provided in any of the above embodiments, by running computer programs stored in the memory 122.

[0295] Electronic device 120 can also communicate with one or more external devices 124 (e.g., keyboard, pointing device, etc.). This communication can be performed via input / output (I / O) interface 125. Furthermore, electronic device 120 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public network, such as the Internet) via network adapter 126. As shown, network adapter 126 communicates with other modules of electronic device 120 via bus 123. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with electronic device 120, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (disk array) systems, tape drives, and data backup storage systems.

[0296] It should be noted that although several units / modules or sub-units / modules of the electronic device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.

[0297] Example 8

[0298] This disclosure also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method provided in any of the above embodiments.

[0299] The readable storage medium may be more specifically adopted, including but not limited to: portable disk, hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.

[0300] Example 9

[0301] This disclosure also provides a computer program product, including a computer program that, when executed by a processor, implements the methods described in any of the above embodiments.

[0302] The program code for executing the computer program product of this disclosure can be written in any combination of one or more programming languages, and the program code can be executed entirely on a user device, partially on a user device, as a stand-alone software package, partially on a user device and partially on a remote device, or entirely on a remote device.

[0303] While specific embodiments of this disclosure have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of this disclosure is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of this disclosure, but all such changes and modifications fall within the scope of protection of this disclosure.

Claims

1. A training method for an energy consumption anomaly detection model, characterized in that, The training method includes: Acquire several sets of sample energy consumption training data of the target energy-consuming object in the target industrial scenario within a historical preset time period; Each set of sample energy consumption training data includes sample energy consumption feature data and corresponding sample energy consumption anomaly annotation data; The sample energy consumption feature data in each set of sample energy consumption training data is used as input, and the corresponding sample energy consumption anomaly labeling data is used as output. The preset random projection model is trained to obtain the energy consumption anomaly detection model. In the preset random projection model, a random hyperplane is used for recursive segmentation.

2. The training method for the energy consumption anomaly detection model as described in claim 1, characterized in that, The step of training a preset random projection model to obtain the energy consumption anomaly detection model based on the sample energy consumption feature data in each group of sample energy consumption training data as input and the corresponding sample energy consumption anomaly labeling data as output includes: A predetermined number of the sample energy consumption training data are randomly selected to form a training subset; Using the training subset as the root node, recursively construct a binary tree; A preset projection vector is randomly generated as the random hyperplane; The projection value of the energy consumption training data of each sample at the current node onto the random hyperplane is calculated. Within the projection range determined based on the projection value, a segmentation point is randomly selected; Based on the projection range and the position of the segmentation point, the corresponding sample energy consumption training data is divided into the left child node or the right child node. In response to the current node meeting the preset conditions, the current node is marked as a leaf node, and the recursive segmentation process is repeated until all nodes are marked as leaf nodes to obtain the energy consumption anomaly detection model.

3. The training method for the energy consumption anomaly detection model as described in claim 2, characterized in that, The preset conditions include at least one of the following: the data depth exceeds a predetermined depth, the current node contains only one sample energy consumption training data, and the sample energy consumption feature data of all the sample energy consumption training data in the current node are completely identical; And / or, The target industrial scenario comprises an N-layer four-dimensional industrial environment; where N ≥ 2 and is a positive integer; And / or, The target industrial scenario includes a three-layer, four-dimensional industrial environment. The target energy consumption object includes at least one of the workshop, production line, and equipment in the industrial environment. The different scenario levels in the industrial environment include at least two of the workshop layer, industry layer, and equipment layer.

4. The training method for the energy consumption anomaly detection model as described in any one of claims 1-3, characterized in that, The step of acquiring several sets of sample energy consumption training data of the target energy consumption object in the target industrial scenario within a historical preset time period includes: Obtain several sets of first historical energy consumption association data of the target energy consumption object in the target industrial scenario within the historical preset time period; Based on the first historical energy consumption correlation data, second historical energy consumption correlation data of different scenario levels in the target industrial scenario are obtained; The second historical energy consumption correlation data is trained separately to obtain a trained single-layer autoencoder for each of the aforementioned scene levels; The output of the previous single-layer autoencoder is used as the input of the next single-layer autoencoder. Multi-layer structure fusion training is performed on different scene levels to train a stacked autoencoder network. Obtain the sample output results of the autoencoder network, and use the sample output results as the sample energy consumption feature data corresponding to the first historical energy consumption association data; wherein, the sample energy consumption feature data is the fusion feature corresponding to different scene levels.

5. The training method for the energy consumption anomaly detection model as described in claim 4, characterized in that, The step of obtaining several sets of first historical energy consumption association data of the target energy consumption object in the target industrial scenario within the historical preset time period includes: Obtain the initial historical energy consumption association data of the target energy consumption object in the target industrial scenario within the preset historical time period; Based on the defined operating conditions, intermediate historical energy consumption correlation data of the target energy consumption object in the target industrial scenario under each operating condition within the preset historical time period are obtained from the initial historical energy consumption correlation data. The intermediate historical energy consumption correlation data under each operating condition is processed using a matching data standardization method to obtain several sets of the first historical energy consumption correlation data. And / or, Before the step of training the second historical energy consumption correlation data separately to obtain the trained single-layer autoencoder for each scene level, the method further includes: Construct energy consumption-derived features; The preset derived features include at least one of energy efficiency intensity features, transient features, and sequence features; The step of training the second historical energy consumption correlation data separately to obtain a trained single-layer autoencoder for each scene level includes: Based on the energy consumption derived features and the second historical energy consumption associated data, a single-layer autoencoder trained for each of the scene levels is obtained.

6. The training method for the energy consumption anomaly detection model as described in claim 5, characterized in that, After the step of obtaining the initial historical energy consumption association data of the target energy consumption object in the target industrial scenario within the preset historical time period, the method further includes: Based on the time dimension, the missing values ​​in the initial historical energy consumption association data are filled using a matching preset filling method to obtain the processed initial historical energy consumption association data. The preset filling method includes interpolation filling method or preset value filling method, wherein the preset value in the preset value filling method includes factory value or average value; And / or, The local anomaly factor algorithm is used to remove outliers from the initial historical energy consumption correlation data to obtain the processed initial historical energy consumption correlation data. And / or, Redundant fields are removed from the initial historical energy consumption association data to obtain the processed initial historical energy consumption association data.

7. A method for detecting abnormal energy consumption in an industrial setting, characterized in that, The detection method includes: Obtain preset energy consumption correlation data of the target energy consumption object in the target industrial scenario within the target preset time period; Obtain the target energy consumption feature data corresponding to the preset energy consumption correlation data; The target energy consumption feature data is input into a pre-trained energy consumption anomaly detection model to obtain the target energy consumption anomaly detection result; The energy consumption anomaly detection model is obtained using the training method for the energy consumption anomaly detection model as described in any one of claims 1-6; The target energy consumption anomaly detection result is used to characterize whether there is an anomaly in the actual energy consumption status of the target industrial scenario within the target preset time period.

8. A training system for an energy consumption anomaly detection model, characterized in that, The training system includes: The sample training data acquisition module is used to acquire several sets of sample energy consumption training data of the target energy consumption object in the target industrial scenario within a historical preset time period; Each set of sample energy consumption training data includes sample energy consumption feature data and corresponding sample energy consumption anomaly annotation data; The model training module is used to train a preset random projection model based on the sample energy consumption feature data in each group of sample energy consumption training data as input and the corresponding sample energy consumption anomaly labeling data as output, so as to obtain the energy consumption anomaly detection model. In the preset random projection model, a random hyperplane is used for recursive segmentation.

9. A detection system for abnormal energy consumption in an industrial setting, characterized in that, The detection system includes: The preset data acquisition module is used to acquire preset energy consumption associated data of the target energy consumption object in the target industrial scenario within the target preset time period; The target feature acquisition module is used to acquire target energy consumption feature data corresponding to the preset energy consumption correlation data; The target detection result acquisition module is used to input the target energy consumption feature data into a pre-trained energy consumption anomaly detection model to obtain the target energy consumption anomaly detection result; The energy consumption anomaly detection model is obtained using the training system for the energy consumption anomaly detection model as described in claim 8; The target energy consumption anomaly detection result is used to characterize whether there is an anomaly in the actual energy consumption status of the target industrial scenario within the target preset time period.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and for running on the processor, characterized in that, When the processor executes the computer program, it implements the training method for the energy consumption anomaly detection model as described in any one of claims 1 to 6; or the energy consumption anomaly detection method in an industrial scenario as described in claim 7.

Images