Large-scale equipment energy consumption anomaly detection method and storage medium

Abstract

The invention relates to the technical field of equipment energy consumption monitoring and fault diagnosis, and discloses a large-scale equipment energy consumption anomaly detection method and a storage medium. The method comprises the following steps: synchronously acquiring equipment energy consumption, state and environment data, fusing to generate multi-dimensional initial features, and performing multi-scale transformation. And performing preliminary detection on each scale feature subset to generate abnormal confidence and mode description. And screening the feature subsets according to the confidence coefficient, and dynamically selecting a corresponding detection strategy according to the mode description. And performing deep feature extraction on the feature subset based on the selected strategy, inputting the obtained high-dimensional feature vector into a corresponding anomaly evaluation network, calculating a standard state matching degree, comparing with a dynamic threshold to generate a fine-grained judgment result, and finally finishing positioning and attribution analysis in combination with anomaly description to form a detection report. According to the method, efficient and accurate large-scale equipment energy consumption anomaly detection is realized.

Description

Technical Field

[0001] This invention relates to the field of equipment energy consumption monitoring and fault diagnosis technology, specifically to a method and storage medium for detecting abnormal energy consumption in large-scale equipment. Background Technology

[0002] In the context of Industrial Internet of Things (IIoT) and smart manufacturing, real-time monitoring and anomaly detection of energy consumption in large-scale equipment clusters are crucial for achieving energy conservation, emission reduction, and predictive maintenance. Existing technical solutions mainly rely on single-dimensional analysis of time-series energy consumption data of equipment operation, or on simply overlaying equipment status and environmental parameters to fuse features and construct a unified anomaly detection model for processing. These methods attempt to use a fixed model architecture to handle all operating conditions and anomaly patterns, ignoring the differences in time scale and performance characteristics of different anomalies.

[0003] The current approach suffers from a drawback: when faced with high-dimensional, multi-source, and heterogeneous monitoring data, fixed models struggle to balance detection efficiency and accuracy. On one hand, performing uniform, in-depth computational analysis on all features incurs significant unnecessary computational overhead, impacting the real-time performance of large-scale applications. On the other hand, a single detection strategy cannot accurately adapt to the diverse anomaly patterns caused by different root causes such as mechanical wear, process deviations, and sudden environmental changes, resulting in coarse detection results with both false positives and false negatives, failing to provide effective guidance for subsequent fault localization and attribution. Achieving efficient and accurate detection, and identifying specific anomaly patterns, remains a challenge for current technology. Summary of the Invention

[0004] The purpose of this invention is to provide a method and storage medium for detecting abnormal energy consumption in large-scale equipment, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for detecting abnormal energy consumption in large-scale equipment, the method comprising:

[0006] Simultaneously collect the operating energy consumption time series sequence, equipment status time series waveform diagram and equipment operating environment parameter set of the target large-scale equipment, and perform fusion processing on the operating energy consumption time series sequence, equipment status time series waveform diagram and equipment operating environment parameter set to generate multi-dimensional initial features;

[0007] The multi-dimensional initial features are subjected to multi-scale transformation to obtain feature subsets of multiple scales. Preliminary anomaly detection is performed on the feature subsets of multiple scales respectively, and preliminary anomaly confidence and corresponding anomaly pattern description are output.

[0008] Based on the preliminary anomaly confidence, the feature subsets of the multiple scales are filtered to select the feature subsets that meet the conditions, and based on the anomaly pattern description, a detection strategy is dynamically selected from the multiple pre-established detection strategies.

[0009] Based on the dynamically selected detection strategy, deep feature extraction is performed on the feature subset that meets the conditions to obtain a high-dimensional deep feature vector;

[0010] The high-dimensional deep feature vector is input into the anomaly evaluation network corresponding to the detection strategy, the standard state matching degree is calculated, and the standard state matching degree is compared with a preset dynamic threshold to generate a fine-grained anomaly judgment result.

[0011] By combining the fine-grained anomaly determination results with the anomaly pattern description, the energy consumption anomaly status of the target large-scale equipment is located and the cause analysis is performed to form a complete energy consumption anomaly detection report.

[0012] Preferably, the operating energy consumption time-series sequence, equipment status time-series waveform diagram, and equipment operating environment parameter set of the target large-scale equipment are collected synchronously. The operating energy consumption time-series sequence, equipment status time-series waveform diagram, and equipment operating environment parameter set are then fused to generate multi-dimensional initial features, including:

[0013] By connecting energy consumption monitoring sensors to the target large-scale equipment, the power consumption value during the operation of the equipment is collected at fixed sampling intervals to generate an operating energy consumption time series.

[0014] By connecting vibration and temperature sensors to the target large-scale equipment, vibration waveforms and temperature change curves of key components of the equipment are collected synchronously and combined to generate a time-series waveform diagram of the equipment status.

[0015] Collect temperature, humidity, and dust concentration parameters of the space where the target large-scale equipment is located to form a set of equipment operating environment parameters;

[0016] The operating energy consumption time series is standardized, the device status time series waveform is filtered and normalized, and the device operating environment parameter set is vectorized and encoded.

[0017] The standardized operating energy consumption time series, the filtered and normalized device state time series waveform, and the vectorized encoded device operating environment parameter set are spliced ​​and fused together after being aligned in time sequence to generate the multi-dimensional initial features.

[0018] Preferably, the multi-dimensional initial features are subjected to multi-scale transformation to obtain feature subsets at multiple scales, and preliminary anomaly detection is performed on each of the feature subsets at multiple scales, outputting preliminary anomaly confidence and corresponding anomaly pattern descriptions, including:

[0019] By using sliding windows of different time lengths, the multi-dimensional initial features are segmented to obtain feature subsets at multiple scales corresponding to different time granularities;

[0020] The feature subsets at each scale are respectively input into a corresponding anomaly feature learning model for processing;

[0021] Each of the aforementioned anomaly feature learning models outputs the reconstruction error and feature distribution divergence of the corresponding feature subset;

[0022] Based on the reconstruction error and feature distribution divergence, calculate the preliminary anomaly confidence level corresponding to each feature subset;

[0023] Meanwhile, each of the aforementioned abnormal feature learning models outputs an abnormal pattern description that is closest to the current feature subset based on its internal pattern library. The abnormal pattern description includes energy consumption mutation mode, steady-state shift mode, and periodic distortion mode.

[0024] Preferably, based on the preliminary anomaly confidence level, the feature subsets at the multiple scales are filtered to select a feature subset that meets the conditions, and based on the anomaly pattern description, a detection strategy is dynamically selected from a pre-established set of detection strategies, including:

[0025] Set a confidence level screening threshold;

[0026] The preliminary anomaly confidence scores corresponding to the feature subsets at the multiple scales are compared with the confidence score screening thresholds respectively, and all feature subsets whose preliminary anomaly confidence scores exceed the confidence score screening thresholds are selected as the feature subset set to be analyzed in depth.

[0027] Analyze the abnormal pattern descriptions output from the abnormal feature learning model that are associated with the feature subset set to be analyzed in depth;

[0028] Based on the main patterns in the abnormal pattern description obtained from statistics, a detection strategy that matches the pattern is dynamically selected from the strategy library, which includes mutation-sensitive detection strategies, offset quantization detection strategies, and periodic analysis detection strategies.

[0029] Preferably, according to the dynamically selected detection strategy, deep feature extraction is performed on the feature subset that meets the conditions to obtain a high-dimensional deep feature vector, including:

[0030] Based on the dynamically selected detection strategy, a corresponding combination of feature extraction operators is selected from multiple feature extraction operators;

[0031] Using the selected combination of feature extraction operators, nonlinear transformation and feature enhancement are performed on each feature subset in the set of feature subsets to be deeply analyzed;

[0032] For each feature subset that has undergone nonlinear transformation and feature enhancement, temporal feature aggregation is performed;

[0033] The representations of all aggregated feature subsets are concatenated and dimensionality reduced to finally compress and generate a unified high-dimensional deep feature vector.

[0034] Preferably, the high-dimensional deep feature vector is input into an anomaly evaluation network corresponding to the detection strategy, a standard state matching degree is calculated, and a fine-grained anomaly determination result is generated by comparing the standard state matching degree with a preset dynamic threshold, including:

[0035] For each detection strategy in the strategy library, a corresponding anomaly evaluation network is pre-trained, and each anomaly evaluation network is trained with a standard feature vector set under normal operating conditions.

[0036] Load the anomaly evaluation network corresponding to the detection strategy into memory;

[0037] The high-dimensional deep feature vector is input into the loaded anomaly evaluation network;

[0038] The anomaly assessment network calculates the distance between the high-dimensional deep feature vector and the internally stored normal pattern feature space, and converts the distance into a standard state matching degree through a mapping function;

[0039] Read the dynamic threshold associated with the current target large-scale equipment operating condition label;

[0040] The calculated standard state matching degree is compared with the dynamic threshold. If the standard state matching degree is lower than the dynamic threshold, a fine-grained anomaly judgment result containing abnormal components and the degree of deviation is generated; otherwise, a judgment result indicating that the equipment is in normal condition is generated.

[0041] Preferably, by combining the fine-grained anomaly determination results with the anomaly pattern description, the energy consumption anomaly status of the target large-scale equipment is located and attributed to generate a complete energy consumption anomaly detection report, including:

[0042] When the fine-grained anomaly determination result indicates that the equipment is in normal condition, the anomaly pattern description is used as an early warning prompt and combined with the fine-grained anomaly determination result to generate an energy consumption anomaly detection report containing early warning information.

[0043] When the fine-grained anomaly determination result contains anomaly components and deviation degree, the anomaly pattern description is associated and matched with the fine-grained anomaly determination result;

[0044] Based on the association matching results, a preset fault knowledge graph is queried, which stores the mapping relationship between abnormal patterns, feature deviations and specific component failures of the equipment;

[0045] Based on the mapping relationship, identify one or more suspected faulty components that cause abnormal energy consumption, and speculate on possible causes of the fault;

[0046] The fine-grained anomaly determination results, associated anomaly pattern descriptions, identified suspected faulty components, and speculated causes of the faults are integrated to generate a structured energy consumption anomaly detection report.

[0047] Preferably, the method for determining the operating condition label includes:

[0048] Continuously monitor the load rate, operating mode signals, and production scheduling instructions of the target large-scale equipment;

[0049] According to preset rules, the combination of load rate, working mode signal and production scheduling instruction is mapped to a specific operating condition label.

[0050] The dynamic threshold is obtained by querying a threshold table bound to the operating condition label, and different operating condition labels correspond to different dynamic thresholds.

[0051] Preferably, the method further includes a step of optimizing the detection strategy:

[0052] Collect the complete energy consumption anomaly detection report generated during the historical detection process and its corresponding multi-dimensional initial features;

[0053] Annotate confirmed anomaly reports and their corresponding multi-dimensional initial features to construct an optimized training sample set;

[0054] The optimized training sample set is used to perform periodic incremental training and parameter updates on the detection strategies in the policy library and their corresponding anomaly evaluation networks.

[0055] Preferably, when the computer program is executed by the processor, it implements the large-scale equipment energy consumption anomaly detection method as described in any one of the above descriptions.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] After performing multi-scale transformations on multi-dimensional features, independent preliminary detection and evaluation are performed on feature subsets at each scale. Based on the output preliminary anomaly confidence scores, subsets requiring focused attention are selected. This avoids investing subsequent deep computational resources in scale features initially judged to be normal, concentrating main computing power on high-risk feature intervals. In practical applications with high data dimensionality and limited computing resources, this can significantly reduce the overall computational load of the system, improve the real-time processing and online detection efficiency of energy consumption data for large-scale device clusters, and make the system more agile in response.

[0058] In the initial detection phase, anomaly pattern descriptions are generated synchronously, and based on these descriptions, a specific detection strategy that best matches the pattern is dynamically selected from a pre-defined strategy library. Subsequent deep feature extraction and anomaly evaluation correspond to this selected strategy. The deep feature extraction module, guided by the strategy, mines the most relevant high-dimensional representations of specific anomaly patterns from a selected feature subset. The anomaly evaluation network utilizes discrimination logic customized for this type of pattern to calculate a more accurate standard state matching degree. This allows the system to adaptively invoke the most professional analysis process based on the suspected anomaly type, thereby achieving more refined and accurate identification and judgment of anomalies with different mechanisms and forms, improving the targeting of detection. Attached Figure Description

[0059] Figure 1 This is a schematic diagram illustrating the working principle of the large-scale equipment energy consumption anomaly detection method described in this invention.

[0060] Figure 2 Flowchart generated for multi-dimensional initial features;

[0061] Figure 3 This is a flowchart for preliminary detection of multi-scale feature anomalies;

[0062] Figure 4 A trend chart showing the standard state matching degree and dynamic threshold under different working conditions;

[0063] Figure 5 This is a comparison chart showing the impact of traditional and improved incremental training on the decrease in the model's standard state matching accuracy. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Please see Figure 1This invention provides a method for detecting energy consumption anomalies in large-scale equipment. The method includes: simultaneously acquiring the operating energy consumption time-series sequence, equipment state time-series waveform diagram, and equipment operating environment parameter set of the target large-scale equipment; fusing the operating energy consumption time-series sequence, equipment state time-series waveform diagram, and equipment operating environment parameter set to generate multi-dimensional initial features; performing multi-scale transformation on the multi-dimensional initial features to obtain feature subsets at multiple scales; performing preliminary anomaly detection on each of the feature subsets at multiple scales, outputting preliminary anomaly confidence and corresponding anomaly pattern descriptions; filtering the feature subsets at multiple scales based on the preliminary anomaly confidence, selecting feature subsets that meet certain conditions, and dynamically selecting a detection strategy from multiple pre-established detection strategies based on the anomaly pattern description; performing deep feature extraction on the feature subsets that meet the conditions according to the dynamically selected detection strategy to obtain high-dimensional deep feature vectors; inputting the high-dimensional deep feature vectors into an anomaly evaluation network corresponding to the detection strategy, calculating the standard state matching degree, and comparing the standard state matching degree with a preset dynamic threshold to generate a fine-grained anomaly judgment result. By combining the fine-grained anomaly determination results with the anomaly pattern description, the energy consumption anomaly status of the target large-scale equipment is located and the cause analysis is performed to form a complete energy consumption anomaly detection report.

[0066] Example 1: See Figure 2 Energy consumption monitoring sensors connected to the target large-scale equipment are used to collect power consumption values ​​during equipment operation at fixed sampling intervals, generating an operational energy consumption time-series sequence. Vibration and temperature sensors connected to the target large-scale equipment are used to simultaneously collect vibration waveforms and temperature change curves of key components of the equipment, combining them to generate an equipment status time-series waveform diagram. Temperature, humidity, and dust concentration parameters of the space where the target large-scale equipment is located are collected to form a set of equipment operating environment parameters. The operational energy consumption time-series sequence is standardized, the equipment status time-series waveform diagram is filtered and normalized, and the set of equipment operating environment parameters is vectorized and encoded. The standardized operational energy consumption time-series sequence, the filtered and normalized equipment status time-series waveform diagram, and the vectorized set of equipment operating environment parameters are aligned in time and then spliced ​​and fused to generate the multi-dimensional initial features.

[0067] In a specific implementation, taking the energy consumption anomaly detection of an industrial centrifugal compressor as an example, the process of generating multi-dimensional initial features is illustrated. Energy consumption monitoring sensors connected to the industrial centrifugal compressor collect power consumption values ​​during equipment operation at fixed sampling intervals, generating an operating energy consumption time series. The fixed sampling interval is set to 1 second. The collected raw power consumption values ​​constitute time series data. In some embodiments, vibration and temperature sensors connected to the main shaft and motor windings of the industrial centrifugal compressor synchronously collect vibration waveforms and temperature change curves of key components of the equipment. The vibration sensor collects triaxial acceleration data, and the temperature sensor collects temperature data, combining them to generate a time series waveform diagram of the equipment status. This time series waveform diagram includes curves of vibration amplitude changing over time and curves of temperature changing over time. In a specific implementation, temperature, humidity, and dust concentration parameters of the machine room where the industrial centrifugal compressor is located are collected and recorded using environmental sensors at the same sampling interval, forming a set of equipment operating environment parameters. Each parameter in the set of equipment operating environment parameters is a time series.

[0068] In practice, the operating energy consumption time series is standardized using the Z-score method, which transforms the original operating energy consumption time series into a series with a mean of 0 and a standard deviation of 1. The formula is as follows:

[0069]

[0070] in: This represents the original operating energy consumption time series. Represents the original operating energy consumption time series. The mean, Represents the original operating energy consumption time series. standard deviation The standardized operating energy consumption time series is represented by filtering and normalization of the equipment status time series waveform. The filtering process uses a low-pass filter to remove high-frequency noise, and the normalization process linearly transforms the vibration amplitude and temperature values ​​to the [0,1] interval. The set of equipment operating environment parameters is vectorized and encoded, and the values ​​of the three parameters, temperature, humidity and dust concentration, at the same time are combined into a three-dimensional vector to form an environmental parameter vector sequence.

[0071] In some embodiments, the standardized operating energy consumption time series, the filtered and normalized device status time series waveform, and the vectorized encoded device operating environment parameter set are spliced ​​and fused after being aligned in time sequence. The time sequence alignment is based on a unified timestamp to ensure that data from different data sources correspond at the same point in time. The splicing operation connects the standardized energy consumption value, normalized vibration amplitude value, normalized temperature value, ambient temperature value, ambient humidity value, and ambient dust concentration value at each time point into a multi-dimensional feature vector. The fusion operation generates multi-dimensional initial features through concatenation. The multi-dimensional initial features are a multi-dimensional time series matrix. First, the... The standardized operating energy consumption time series, the filtered and normalized equipment status time series waveform, and the vectorized equipment operating environment parameter set are strictly aligned in time to ensure that each data point on each sampling timestamp corresponds one-to-one. Then, at each aligned time point, the standardized energy consumption value, the normalized vibration amplitude and normalized temperature value in the equipment status waveform, and the ambient temperature, ambient humidity, and ambient dust concentration values ​​in the environmental parameter vector are directly connected in a predetermined order to form a unified multidimensional feature vector. Finally, these multidimensional feature vectors at all time points are arranged in time series to form a multidimensional time series matrix, i.e., multidimensional initial features. Optionally, data comparison shows that after standardization, the numerical distribution of the original operating energy consumption time series is centered; after filtering, the high-frequency noise component of the equipment status time series waveform is reduced, and after normalization, the amplitude range is unified; after vectorization encoding, the set of equipment operating environment parameters is integrated from scattered parameters into structured vectors; time alignment ensures the time synchronization of multi-source data; splicing and fusion form a composite feature representation containing energy consumption, status, and environmental information. It can be understood that the fixed sampling interval can be adjusted according to the dynamic characteristics of the equipment. For slow-changing processes, the sampling interval can be extended, and for fast-changing processes, the sampling interval can be shortened. In specific implementation, the generation of multi-dimensional initial features is the input basis for subsequent multi-scale transformation and anomaly detection.

[0072] Example 2: See Figure 3The multi-dimensional initial features are segmented using sliding windows of different time lengths to obtain feature subsets at multiple scales corresponding to different time granularities. Each feature subset at each scale is input into a corresponding anomaly feature learning model for processing. Each anomaly feature learning model outputs the reconstruction error and feature distribution divergence for its corresponding feature subset. Based on the reconstruction error and feature distribution divergence, a preliminary anomaly confidence score is calculated for each feature subset. Simultaneously, each anomaly feature learning model outputs the anomaly pattern description closest to the current feature subset based on its internal pattern library. These anomaly pattern descriptions include energy consumption mutation patterns, steady-state shift patterns, and periodic distortion patterns. A confidence score screening threshold is set. The preliminary anomaly confidence scores corresponding to the feature subsets at each of the multiple scales are compared with the confidence score screening threshold, and all feature subsets with preliminary anomaly confidence scores exceeding the threshold are selected as the feature subset set to be analyzed in depth. The anomaly pattern descriptions output from the anomaly feature learning models and associated with the feature subset set to be analyzed in depth are analyzed. Based on the main patterns in the abnormal pattern description obtained from statistics, a detection strategy that matches the pattern is dynamically selected from the strategy library, which includes mutation-sensitive detection strategies, offset quantization detection strategies, and periodic analysis detection strategies.

[0073] In specific implementation, sliding windows of different time lengths are used to segment the multi-dimensional initial features, obtaining multiple feature subsets corresponding to different time granularities. The time lengths of the sliding windows include 1 minute, 5 minutes, and 30 minutes, capturing instantaneous fluctuations, short-term trends, and long-term steady-state features, respectively. The sliding window at each scale slides on the time axis with a 50% overlap rate, thereby extracting multiple fixed-length feature subsets from the multi-dimensional initial features. Each feature subset is a three-dimensional tensor, with dimensions representing the time step, feature channels, and sequence batches. In some embodiments, the feature subsets at each scale are input into a corresponding anomaly feature learning model for processing. The anomaly feature learning model is a neural network model based on an autoencoder architecture. Each model is trained independently for the data distribution at a specific time scale. Each anomaly feature learning model outputs the reconstruction error and feature distribution divergence of the corresponding feature subset. The reconstruction error is obtained by calculating the mean square error between the input feature subset and the model reconstruction output. The feature distribution divergence is obtained by calculating the Mahalanobis distance between the feature vector of the input feature subset in the model's hidden space and the feature vector center of the normal dataset.

[0074] In practice, based on the reconstruction error and the feature distribution divergence, the preliminary anomaly confidence level corresponding to each feature subset is calculated. The calculation formula is:

[0075]

[0076] in: Indicates the initial confidence level of the anomaly. Indicates reconstruction error, Represents the divergence of the characteristic distribution. and These represent the minimum and maximum reconstruction errors obtained during the training phase, respectively. and These represent the minimum and maximum values ​​of the feature distribution divergence obtained during the training phase, respectively. and It is a preset weighting coefficient and satisfies Meanwhile, each anomaly feature learning model outputs the anomaly pattern description closest to the current feature subset based on its internal pattern library. The internal pattern library stores feature templates for various predefined anomaly patterns. By calculating the similarity between the current feature subset and each template, the description corresponding to the template with the highest similarity is used as the output. The anomaly pattern descriptions include energy consumption mutation patterns, steady-state shift patterns, and periodic distortion patterns. It can be understood that, based on data comparison, the 1-minute scale feature subset is sensitive to instantaneous power consumption spikes, and its anomaly pattern description is often an energy consumption mutation pattern; the 5-minute scale feature subset is sensitive to trend changes, and its anomaly pattern description is often a steady-state shift pattern; and the 30-minute scale feature subset is sensitive to the disruption of periodic patterns, and its anomaly pattern description is often a periodic distortion pattern.

[0077] In some embodiments, a confidence screening threshold is set, which is determined based on the statistical percentile of the preliminary anomaly confidence calculated during the historical normal data processing. The preliminary anomaly confidence corresponding to feature subsets at multiple scales is compared with the confidence screening threshold. Feature subsets whose preliminary anomaly confidence exceeds the confidence screening threshold are selected as the feature subset set to be analyzed in depth. If the preliminary anomaly confidence of all feature subsets at a certain time scale does not exceed the confidence screening threshold, the feature subset at that scale is not included in the subsequent deep analysis process. The anomaly pattern descriptions output from the anomaly feature learning model and associated with the feature subset set to be analyzed in depth are analyzed, and the anomaly pattern descriptions corresponding to all selected feature subsets are statistically analyzed. The frequency of occurrence of the described anomaly pattern can be selected. Optionally, based on the main patterns in the statistically obtained anomaly pattern descriptions, a matching detection strategy can be dynamically selected from the strategy library. The strategy library includes mutation-sensitive detection strategies, offset quantization detection strategies, and periodic analysis detection strategies. If the most frequent anomaly pattern description is the energy consumption mutation pattern, then the mutation-sensitive detection strategy is dynamically selected. This strategy is configured with feature extraction operators that focus on gradient and mutation point detection. It can be understood that the dynamic selection mechanism ensures that the subsequent deep feature extraction and evaluation stages are optimized for the most significant anomaly pattern, avoiding the limitations of fixed strategies. Optionally, if multiple anomaly pattern descriptions have the same frequency of occurrence, the mutation-sensitive detection strategy is selected by default because it corresponds to the most urgent anomaly type.

[0078] Example 3: Based on the dynamically selected detection strategy, a corresponding feature extraction operator combination is selected from multiple feature extraction operators. Using the selected feature extraction operator combination, nonlinear transformation and feature enhancement are performed on each feature subset in the set of feature subsets to be deeply analyzed. Temporal feature aggregation is performed on each feature subset after nonlinear transformation and feature enhancement. The representations of all aggregated feature subsets are concatenated and dimensionality reduced, ultimately compressed to generate a unified high-dimensional deep feature vector. A corresponding anomaly evaluation network is pre-trained for each detection strategy in the strategy library, and each anomaly evaluation network is trained using a standard feature vector set under normal operating conditions. The anomaly evaluation network corresponding to the detection strategy is loaded into memory. The high-dimensional deep feature vector is input into the loaded anomaly evaluation network. The anomaly evaluation network calculates the distance between the high-dimensional deep feature vector and the internally stored normal mode feature space, and converts the distance into a standard state matching degree through a mapping function. A dynamic threshold associated with the current target large-scale equipment operating condition label is read. The calculated standard state matching degree is compared with the dynamic threshold. If the standard state matching degree is lower than the dynamic threshold, a fine-grained anomaly judgment result containing abnormal components and the degree of deviation is generated; otherwise, a judgment result indicating that the equipment is in normal condition is generated.

[0079] In practical implementation, based on the dynamically selected detection strategy, a corresponding combination of feature extraction operators is chosen from multiple feature extraction operators. These operators include gradient calculation operators, sliding statistics operators, and frequency domain transformation operators. For dynamically selected mutation-sensitive detection strategies, the feature extraction operator combination includes gradient calculation operators and peak detection operators. Using the selected feature extraction operator combination, each feature subset in the feature subset set to be deeply analyzed undergoes nonlinear transformation and feature enhancement. The nonlinear transformation is implemented using the ReLU activation function, and feature enhancement reweights different time steps of the feature subset using a learnable attention weight matrix to highlight temporal segments related to mutation patterns. In some embodiments, temporal feature aggregation is performed on each feature subset after nonlinear transformation and feature enhancement. For mutation-sensitive strategies, the feature aggregation operation uses max pooling to capture features at different time scales. The most significant mutation response is concentrated, and the representations of all aggregated feature subsets are concatenated and dimensionality reduced. The concatenation operation connects the aggregated feature vectors at 1-minute, 5-minute, and 30-minute scales into a high-dimensional intermediate vector. The dimensionality reduction operation compresses the high-dimensional intermediate vector through a fully connected neural network layer to generate a unified high-dimensional deep feature vector. The dimension of the high-dimensional deep feature vector is fixed at 128 dimensions. Optionally, data comparison shows that for the same set of feature subsets to be analyzed in depth, if the offset quantization detection strategy is dynamically selected, the feature extraction operator combination includes the moving average and variance operators, and the feature aggregation uses average pooling. The generated high-dimensional deep feature vector has prominent values ​​in the dimension reflecting trend statistics. If the periodic analysis detection strategy is dynamically selected, the feature extraction operator combination includes the fast Fourier transform operator, and the feature aggregation uses spectral energy summation. The generated high-dimensional deep feature vector has concentrated energy in the dimensions corresponding to the fundamental frequency and harmonics.

[0080] In practice, a corresponding anomaly evaluation network is pre-trained for each detection strategy in the strategy library. Each anomaly evaluation network is trained using a standard feature vector set under normal operating conditions. The standard feature vector set is a set of high-dimensional deep feature vectors extracted and stored during the device's historical normal operation through the aforementioned steps. For mutation-sensitive detection strategies, the corresponding anomaly evaluation network is a trained single-class support vector machine model. This model learns the distribution boundary of high-dimensional deep feature vectors under normal operating conditions during the training phase. The anomaly evaluation network corresponding to the dynamically selected detection strategy is loaded into memory, and the high-dimensional deep feature vectors are input into the loaded anomaly evaluation network. The anomaly evaluation network calculates the relative distance between the high-dimensional deep feature vectors and the internally learned normal pattern boundary, and maps this relative distance to a standardized state matching degree. The calculation formula is:

[0081]

[0082] in: This represents the standard state matching degree, and its value range is (0,1). This represents the relative distance calculated by the anomaly assessment network. This represents the steepness coefficient of the sigmoid function, used to control the sensitivity of the matching degree as distance changes. This represents the offset parameter, and the standard state matching degree. The closer the value is to 1, the higher the degree of matching with the standard state.

[0083] In some embodiments, a dynamic threshold associated with the current target large-scale equipment operating condition label is read. The operating condition label is provided in real time by an external monitoring system, such as the "rated load operation" label. The dynamic threshold is obtained by querying a threshold table bound to the operating condition label. Under the "rated load operation" condition, the corresponding dynamic threshold may be set to 0.70. The calculated standard state matching degree is compared with the dynamic threshold. If the standard state matching degree is lower than the dynamic threshold, a fine-grained anomaly judgment result containing abnormal components and the degree of deviation is generated. The abnormal components are identified by analyzing the high-dimensional deep feature vectors that lead to distance. The increase in the main dimensional contribution is used to determine the degree of deviation, which is quantified by calculating the arithmetic difference between the standard state matching degree and the dynamic threshold. Otherwise, a judgment result indicating that the device state is normal is generated. Optionally, data comparison shows that when the input high-dimensional deep feature vector is highly similar to the normal vector in the training set, the distance... The standard state matching degree, calculated using the formula, is close to 0. A value close to 1, far exceeding the dynamic threshold of 0.70, is considered normal. However, when the input high-dimensional deep feature vector exhibits a clear abnormal mutation pattern, the distance... The calculated standard state matching degree is significantly increased. It may drop to 0.50, below the dynamic threshold, thereby triggering a fine-grained anomaly judgment result that includes specific anomaly dimensions and a deviation degree of 0.20. It can be understood that by calculating the standard state matching degree through the sigmoid function and comparing the result with the dynamic threshold that depends on the operating conditions, an adaptive fine-grained evaluation of the equipment state under different operating conditions is achieved.

[0084] Example 4: When the fine-grained anomaly determination result indicates that the equipment is in normal condition, the anomaly pattern description is used as a warning and merged with the fine-grained anomaly determination result to generate an energy consumption anomaly detection report containing warning information. When the fine-grained anomaly determination result contains anomaly components and deviation degrees, the anomaly pattern description is associated and matched with the fine-grained anomaly determination result. Based on the association matching result, a preset fault knowledge graph is queried. The fault knowledge graph stores the mapping relationship between anomaly patterns, feature deviations, and specific component failures of the equipment. Based on the mapping relationship, one or more suspected faulty components causing energy consumption anomalies are identified, and possible causes of failure are inferred. The fine-grained anomaly determination result, the associated anomaly pattern description, the identified suspected faulty components, and the inferred causes of failure are integrated to generate a structured energy consumption anomaly detection report. The method for determining the operating condition label includes: continuously monitoring the load rate, working mode signal, and production scheduling instructions of the target large-scale equipment. According to preset rules, the combination of the load rate, working mode signal, and production scheduling instructions is mapped to an operating condition label. The dynamic threshold is obtained by querying a threshold table bound to the operating condition label, and different operating condition labels correspond to different dynamic thresholds.

[0085] In specific implementation, when the fine-grained anomaly determination result indicates that the equipment status is normal, the anomaly pattern description is used as an early warning prompt and merged with the fine-grained anomaly determination result to generate an energy consumption anomaly detection report containing early warning information. The early warning information clearly lists the anomaly pattern description identified in the preliminary detection stage, such as "a sudden change in instantaneous performance consumption pattern feature was detected, but the comprehensive assessment did not exceed the dynamic threshold." The report also includes the standard state matching degree value and the dynamic threshold value used in the fine-grained anomaly determination result. In some embodiments, when the fine-grained anomaly determination result contains anomaly components and deviation degrees, the anomaly pattern description is associated and matched with the fine-grained anomaly determination result. The association and matching operation is based on time window alignment, which associates the original time period corresponding to the high-dimensional deep feature vector that triggers the fine-grained anomaly determination with the anomaly feature learning model that outputs the anomaly pattern description within that time period, establishing a mapping relationship between the anomaly pattern description and the specific anomaly feature dimension.

[0086] In practical implementation, based on the association matching results, a preset fault knowledge graph is queried. The fault knowledge graph stores the mapping relationship between abnormal patterns, feature deviations, and specific component failures of the equipment. The fault knowledge graph is stored in the form of a graph database. Nodes include abnormal pattern types, feature dimensions, and equipment components. Edges define the causal relationships and probability weights between them. Based on the mapping relationship, one or more suspected faulty components causing energy consumption abnormalities are identified, and possible fault causes are inferred. For example, if the association matching result is "energy consumption mutation pattern" and "significant deviation in high-frequency vibration energy feature dimension", then querying the fault knowledge graph may map to suspected faulty components and causes such as "bearing lubrication failure" or "minor rubbing of rotor components". The fine-grained abnormality judgment results, associated abnormal pattern descriptions, identified suspected faulty components, and inferred fault causes are integrated to generate a structured energy consumption abnormality detection report. The structured report includes timestamps, standard state matching degree, dynamic thresholds, abnormal pattern descriptions, main abnormal feature dimensions, a list of suspected faulty components, fault cause inferences, and report generation time.

[0087] The method for determining the operating condition label includes: continuously monitoring the load rate, working mode signals, and production scheduling instructions of the target large-scale equipment. The load rate is read as a real-time power percentage from the equipment control system. Working mode signals such as "rated operation", "variable frequency operation", and "no load" are parsed from the status words on the equipment control panel. Production scheduling instructions come from the upper-level manufacturing execution system. According to preset rules, the combination of load rate, working mode signals, and production scheduling instructions is mapped to the operating condition label. The mapping rules are implemented through a preset lookup table. Dynamic thresholds are obtained by querying the threshold table bound to the operating condition label. Different operating condition labels correspond to different dynamic thresholds. See Table 1. The threshold table is loaded from the database during system initialization.

[0088] Table 1: Correspondence between Operating Condition Labels and Dynamic Thresholds

[0089]

[0090] Understandably, the fine-grained classification of operating condition labels allows the dynamic threshold to more accurately match the normal fluctuation range of the equipment under different operating conditions. Data comparison shows that under the "full load production" condition, the baseline fluctuation of equipment vibration and energy consumption is relatively large, so the dynamic threshold is set lower to reduce false alarms; under the "low load standby" condition, the equipment operates smoothly, and the normal state matching degree is concentrated and the value is high, so the dynamic threshold is set higher to maintain detection sensitivity. Optional, the degree of deviation... It can be calculated using the formula:

[0091]

[0092] in: Indicates the degree of deviation. This represents the dynamic threshold obtained from the threshold table. This represents the calculated standard state matching degree, or the degree of deviation. A positive value indicates an anomaly, and its magnitude reflects the severity of the anomaly. In some embodiments, this deviation is... The values ​​are recorded in the fine-grained anomaly judgment results and used as one of the reference bases for prioritizing suspected fault causes. Optionally, the query results of the fault knowledge graph may return multiple possible fault components. These components are sorted according to the probability weight of the edges in the graph and listed in the report. It can be understood that by combining the anomaly pattern description from the preliminary detection and the fine-grained judgment results from the in-depth assessment, the localization and attribution analysis can be cross-validated from both the macro-pattern and micro-feature levels, improving the pertinence of fault inference. The final structured energy consumption anomaly detection report provides equipment maintenance personnel with a comprehensive view from abnormal phenomena, quantitative indicators to possible causes.

[0093] See Figure 4 This study presents the temporal relationship between the standard state matching degree (blue curve) and the corresponding dynamic threshold (red dashed line) of a target large-scale equipment under three operating conditions: "full-load production," "half-load commissioning," and "low-load standby," and marks outliers (black dots). Specifically, with time as the horizontal axis and the standard state matching degree / dynamic threshold as the vertical axis, the three operating conditions are distinguished by background color blocks: the light yellow area represents "full-load production," corresponding to a dynamic threshold of 0.82; the light green area represents "half-load commissioning," corresponding to a dynamic threshold of 0.88; and the light purple area represents "low-load standby," corresponding to a dynamic threshold of 0.94. The standard state matching degree curve fluctuates over time, and its fluctuation amplitude is consistent with the operating condition characteristics: the fluctuation amplitude is largest under the "full-load production" condition, and the fluctuation amplitude is smallest under the "low-load standby" condition. The logic for marking outliers is: when the standard state matching degree is lower than the dynamic threshold of the corresponding operating condition, it is judged as an anomaly. For example, the standard state matching degree (0.75) of the anomaly point in the "full load production" condition (around 04:00) is significantly lower than the dynamic threshold (0.82) of the condition; the matching degree (0.85) of the anomaly point in the "half load debugging" condition (around 12:00) is close to the threshold (0.88) of the condition; and the matching degree (0.92) of the anomaly point in the "low load standby" condition (around 20:00) is lower than the threshold (0.94) of the condition.

[0094] Example 5: Collect the complete energy consumption anomaly detection reports and their corresponding multi-dimensional initial features generated during historical detection processes. Label the confirmed anomaly reports and their corresponding multi-dimensional initial features to construct an optimized training sample set. Use the optimized training sample set to periodically perform incremental training and parameter updates on the detection strategies in the strategy library and their corresponding anomaly evaluation networks. In practice, complete energy consumption anomaly detection reports and their corresponding multi-dimensional initial features generated during historical detection processes are collected. The complete energy consumption anomaly detection reports and the multi-dimensional initial features that triggered the reports are associated and stored through a unified detection task identifier. In an example scenario, for a monitoring system of an industrial centrifugal compressor, multiple detection task instances are generated every day. Each instance contains end-to-end data from synchronous acquisition to report generation. This data is persistently stored in a time-series database and a relational database. Anomaly reports confirmed as genuine and their corresponding multi-dimensional initial features are labeled to construct an optimized training sample set. The confirmation of the report's authenticity depends on subsequent equipment maintenance records and expert review. For example, an energy consumption anomaly detection report labeled "bearing lubrication failure" can be labeled as a genuine anomaly sample if the maintenance record shows that the compressor's bearing was indeed lubricated during the same time period. Reports labeled as "false positives" and their corresponding multi-dimensional initial features are added to the optimized training sample set as negative samples.

[0095] In some embodiments, the detection strategies in the policy library and their corresponding anomaly evaluation networks are periodically incrementally trained and their parameters are updated using an optimized training sample set. Periodic incremental training is automatically triggered during periods of low system load, such as early Sunday morning. The incremental training process includes loading newly labeled samples from the optimized training sample set, performing feature extraction and transformation on the samples in accordance with the online process, generating corresponding high-dimensional deep feature vectors and labels, and then using these new data to fine-tune the anomaly evaluation network corresponding to a specific detection strategy. Optionally, the fine-tuning adopts the gradient descent algorithm, and the loss function adds a penalty term for new samples to the original reconstruction error or single-classification loss to adjust the network decision boundary. The parameter update is not limited to the weights of the anomaly evaluation network, but may also involve the internal parameters of the feature extraction operator combination associated with the detection strategy in the policy library. For example, the sensitivity coefficient of the gradient operator in the mutation-sensitive detection strategy can be adaptively adjusted by learning the mutation characteristics of real anomaly samples.

[0096] In practice, incremental training involves iteratively updating the model parameters. Its core steps are calculating the loss and backpropagation. The loss function considers a balance between retaining historical knowledge and learning new knowledge. Optionally, a formula for weight adjustment can be expressed as:

[0097]

[0098] in: This represents the updated model parameters. This represents the model parameters before the update. Indicates the incremental learning rate. This represents the loss gradient calculated on the newly optimized training sample set. The elasticity coefficient is used to control the strength of retention of old knowledge. This represents the reference model parameters saved after the last incremental training. It can be understood that the purpose of introducing the elasticity coefficient and the reference parameter term is to prevent catastrophic forgetting of the learned old normal pattern features when adapting to new abnormal patterns. Data comparison shows that after traditional incremental training without introducing the elasticity term, the model's standard state matching degree to historical normal data decreased by an average of about 0.15, while after using the above method, the matching degree decrease was controlled within 0.03.

[0099] Optionally, optimizing the construction of the training sample set also includes the allocation of sample weights. For real anomaly samples that are repeatedly identified as having the same fault, their sample weights will be appropriately increased as the number of occurrences increases, so as to ensure that the model can reinforce the learning of such stable and reproducible fault modes. In some embodiments, after incremental training is completed, the updated anomaly evaluation network parameters and detection strategy configuration will be packaged into a new model version. After a simplified offline verification process, it will be seamlessly deployed to the online detection system to replace the old version, thereby completing the optimization of the detection strategy. It can be understood that this periodic optimization mechanism enables the entire detection system to continuously improve itself as the device running time accumulates and the potential fault modes evolve, without the need for manual redesign of the model.

[0100] See Figure 5During the periodic incremental training of the detection strategy, the difference in the impact of traditional incremental training and improved incremental training (with elasticity coefficient) on the decline value of the model's standard state matching degree is visually presented. Specifically, the decline value of the standard state matching degree is used to quantify the degree to which the model forgets historical normal data after incremental training: the higher the curve value, the more significant the model's deviation from the features of old normal patterns. The red curve in the figure corresponds to traditional incremental training, whose matching degree decline value continuously increases with the number of training rounds, and exceeds the maximum decline threshold of traditional training (0.15) in the later stages of training; the blue curve corresponds to improved incremental training (with elasticity coefficient), whose matching degree decline value remains at a low level and does not exceed the maximum decline threshold of improved training (0.03). In the parameter configuration process, improved incremental training introduces an elasticity coefficient to control the strength of retaining old knowledge, and adds a new sample penalty term to the loss function to balance the learning of historical and new knowledge. With 20 training rounds set, by comparing the curve trends of the two training methods, it can be clearly seen that improved incremental training effectively suppresses the model's "catastrophic forgetting" and ensures the stability of the model's recognition of historical normal patterns after incremental training.

[0101] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0102] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A large-scale equipment energy consumption anomaly detection method, characterized in that, The method comprises: Synchronously collecting a running energy consumption time sequence, a device state time sequence waveform diagram and a device running environment parameter set of a target large-scale device, performing fusion processing on the running energy consumption time sequence, the device state time sequence waveform diagram and the device running environment parameter set, and generating multi-dimensional initial features; Performing multi-scale transformation on the multi-dimensional initial features, obtaining feature subsets of multiple scales, and performing preliminary anomaly detection on the feature subsets of multiple scales respectively, and outputting preliminary anomaly confidence and corresponding anomaly mode description; According to the preliminary anomaly confidence, screening the feature subsets of multiple scales, selecting feature subsets meeting conditions, and dynamically selecting a detection strategy from a plurality of detection strategies pre-established based on the anomaly mode description; According to the dynamically selected detection strategy, performing deep feature extraction on the feature subsets meeting conditions, and obtaining high-dimensional deep feature vectors; Inputting the high-dimensional deep feature vectors into an anomaly evaluation network corresponding to the detection strategy, calculating a standard state matching degree, and comparing the standard state matching degree with a pre-set dynamic threshold to generate a fine-grained anomaly judgment result; Combining the fine-grained anomaly judgment result and the anomaly mode description, positioning and attributing the energy consumption anomaly state of the target large-scale device to form a complete energy consumption anomaly detection report.

2. The method of claim 1, wherein, Synchronously collecting a running energy consumption time sequence, a device state time sequence waveform diagram and a device running environment parameter set of a target large-scale device, performing fusion processing on the running energy consumption time sequence, the device state time sequence waveform diagram and the device running environment parameter set, and generating multi-dimensional initial features, comprising: Collecting power consumption values in the device running process at a fixed sampling interval through an energy consumption monitoring sensor connected to the target large-scale device to generate a running energy consumption time sequence; Synchronously collecting vibration waveforms and temperature change curves of key components of the device through vibration sensors and temperature sensors connected to the target large-scale device to generate a device state time sequence waveform diagram; Collecting temperature, humidity and dust concentration parameters of the space where the target large-scale device is located to form a device running environment parameter set; Performing standardization processing on the running energy consumption time sequence, filtering and normalization processing on the device state time sequence waveform diagram, and vectorization encoding on the device running environment parameter set; After aligning the running energy consumption time sequence after standardization processing, the device state time sequence waveform diagram after filtering and normalization processing and the device running environment parameter set after vectorization encoding in time sequence, splicing and fusing to generate the multi-dimensional initial features.

3. The method of claim 2, wherein, Performing multi-scale transformation on the multi-dimensional initial features, obtaining feature subsets of multiple scales, and performing preliminary anomaly detection on the feature subsets of multiple scales respectively, and outputting preliminary anomaly confidence and corresponding anomaly mode description, comprising: Using sliding windows of different time lengths to segment the multi-dimensional initial features to obtain feature subsets of multiple scales corresponding to different time granularities; inputting the feature subsets of each scale into corresponding anomaly feature learning models respectively for processing; each anomaly feature learning model outputs a reconstruction error and a feature distribution divergence of the corresponding feature subset; a preliminary anomaly confidence of each feature subset is calculated according to the reconstruction error and the feature distribution divergence; Meanwhile, each anomaly feature learning model outputs an anomaly pattern description closest to the current feature subset according to an internal pattern library, the anomaly pattern description including an energy consumption mutation pattern, a steady-state deviation pattern and a periodic distortion pattern.

4. The method of claim 3, wherein, According to the preliminary anomaly confidence, the feature subsets of multiple scales are screened to select feature subsets meeting the conditions, and a detection strategy is dynamically selected from a plurality of detection strategies pre-established based on the anomaly pattern description, including: setting a confidence screening threshold; comparing the preliminary anomaly confidence corresponding to the feature subsets of multiple scales with the confidence screening threshold respectively, screening feature subsets whose preliminary anomaly confidence exceeds the confidence screening threshold as a feature subset set to be deeply analyzed; analyzing the anomaly pattern description associated with the feature subset set to be deeply analyzed output from the anomaly feature learning model; dynamically selecting a detection strategy matching the main pattern in the anomaly pattern description from a strategy library, the strategy library including a mutation sensitive detection strategy, a deviation quantification detection strategy and a periodic analysis detection strategy.

5. The method of claim 4, wherein, According to the dynamically selected detection strategy, deep feature extraction is performed on the feature subsets meeting the conditions to obtain a high-dimensional deep feature vector, including: According to the dynamically selected detection strategy, a corresponding feature extraction operator combination is selected from a plurality of feature extraction operators; using the selected feature extraction operator combination, performing nonlinear transformation and feature enhancement on each feature subset in the feature subset set to be deeply analyzed respectively; performing feature aggregation on each feature subset after nonlinear transformation and feature enhancement in time sequence; concatenating and dimensionality reducing the representations of all aggregated feature subsets to finally compress and generate a unified high-dimensional deep feature vector.

6. The method of claim 5, wherein, The high-dimensional deep feature vector is input into an anomaly evaluation network corresponding to the detection strategy to calculate a standard state matching degree, and the standard state matching degree is compared with a preset dynamic threshold to generate a fine-grained anomaly judgment result, including: pre-training a corresponding anomaly evaluation network for each detection strategy in the strategy library, each anomaly evaluation network being trained with a standard feature vector set under normal working conditions; loading the anomaly evaluation network corresponding to the detection strategy into the memory; inputting the high-dimensional deep feature vector into the loaded anomaly evaluation network; the anomaly evaluation network calculates the distance between the high-dimensional deep feature vector and the normal mode feature space stored internally, and converts the distance into a standard state matching degree through a mapping function; reading a dynamic threshold associated with the current target large-scale equipment running condition label; The standard state matching degree obtained by calculation is compared with the dynamic threshold value, if the standard state matching degree is lower than the dynamic threshold value, a fine-grained abnormality judgment result containing abnormal components and deviation degree is generated, otherwise a judgment result of normal device state is generated.

7. The method of claim 6, wherein the method further comprises: The fine-grained abnormality judgment result is combined with the abnormal mode description, and the energy consumption abnormal state of the target large-scale device is located and attributed to analyze, forming a complete energy consumption abnormality detection report, including: When the fine-grained abnormality judgment result is normal device state, the abnormal mode description is taken as a pre-warning prompt, combined with the fine-grained abnormality judgment result, an energy consumption abnormality detection report containing pre-warning information is generated; When the fine-grained abnormality judgment result contains abnormal components and deviation degree, the abnormal mode description is associated and matched with the fine-grained abnormality judgment result; According to the association matching result, a pre-set fault knowledge graph is queried, the fault knowledge graph stores the mapping relationship between abnormal mode, feature deviation and device specific component fault; According to the mapping relationship, one or more suspected fault components causing energy consumption abnormality are determined, and possible fault reasons are speculated; The fine-grained abnormality judgment result, the associated abnormal mode description, the determined suspected fault component and the speculated fault reason are integrated to generate a structured energy consumption abnormality detection report. 8.The method of claim 7, wherein, The determination method of the running condition label includes: Continuously monitoring the load rate, working mode signal and production scheduling instruction of the target large-scale device; According to a pre-set rule, the combination of the load rate, working mode signal and production scheduling instruction is mapped to a specific running condition label; The dynamic threshold value is obtained by querying the threshold value table bound to the running condition label, different running condition labels correspond to different dynamic threshold values. 9.The method of claim 1, wherein, The method further includes the step of optimizing the detection strategy: Collect the complete energy consumption abnormality detection report and its corresponding multi-dimensional initial features generated in the historical detection process; Label the abnormal reports confirmed to be real and their corresponding multi-dimensional initial features to build an optimization training sample set; Use the optimization training sample set to periodically incrementally train and update the parameters of the detection strategy and its corresponding abnormality evaluation network in the strategy library.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the large-scale device energy consumption abnormality detection method according to any one of claims 1 to 9. The computer program is executed by the processor to realize the large-scale device energy consumption abnormality detection method according to any one of claims 1 to 9.

Images