An energy consumption alarm threshold self-learning system and method for operating state of an electromechanical system and a storage medium thereof

Abstract

This invention discloses a self-learning system, method, and storage medium for energy consumption alarm thresholds in electromechanical systems, relating to the field of energy management technology. The system includes: a benchmark anchoring distribution module, which trains a probability distribution model based on a verified dataset of normal operating conditions of the electromechanical system, generates a benchmark anchoring distribution of normal energy consumption under the target operating condition, and freezes all model parameters. This benchmark distribution is used as the confidence score benchmark for energy consumption samples. A confidence-weighted incremental update module acquires newly added, non-alarm-triggered energy consumption operating samples, calculates the corresponding probability density values, and maps them to training weights to obtain a weighted sample set. A dynamically updated model is constructed using the model corresponding to the benchmark distribution as the initial state, and incremental fine-tuning is performed. The upper quantile value of the dynamic energy consumption distribution is taken as the alarm threshold. This invention can effectively alleviate the threshold offset problem caused by the accumulation of normal samples at the boundary, improving the accuracy and adaptability of energy consumption alarms.

Description

Technical Field

[0001] This invention relates to the field of energy management technology, and in particular to a self-learning system, method and storage medium for energy consumption alarm thresholds of electromechanical system operating status. Background Technology

[0002] In the operation and management of industrial electromechanical equipment, energy consumption monitoring and anomaly alarms are the core links to ensure the safe and stable operation of equipment and improve energy utilization efficiency. Reasonable alarm thresholds can identify abnormal deviations in the operating status of equipment in a timely manner, provide data support for operation and maintenance decisions, and are widely used in the daily operation and maintenance scenarios of various industrial electromechanical systems.

[0003] Existing energy consumption alarm solutions mostly use fixed threshold configurations or dynamically update alarm thresholds based on incremental learning using all newly added samples. Fixed thresholds are difficult to adapt to the energy consumption distribution shifts caused by changes in electromechanical system operating conditions and loads, easily leading to a large number of false alarms or missed alarms. On the other hand, incremental update solutions based on all samples will continue to participate in model iterations during long-term operation, gradually widening the distribution boundary of normal energy consumption. This causes a continuous positive shift in alarm thresholds, resulting in a gradual decrease in alarm sensitivity and an inability to maintain a stable and reliable alarm effect in the long term. Summary of the Invention

[0004] The purpose of this invention is to provide a self-learning system, method and storage medium for energy consumption alarm threshold of electromechanical system operating status, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a self-learning system for energy consumption alarm thresholds of electromechanical system operating status, comprising a benchmark anchoring distribution module and a confidence-weighted incremental update module;

[0006] The benchmark anchoring distribution module is used to train a probability distribution model based on a verified dataset of normal operation of the electromechanical system, and generate a benchmark anchoring distribution of normal energy consumption under the target operating conditions of the electromechanical system. After training, all parameters of the probability distribution model are frozen so that the benchmark anchoring distribution remains fixed during system operation and serves as the confidence score benchmark for energy consumption operation samples. The probability density value of the energy consumption operation sample under the benchmark anchoring distribution is the confidence score of the energy consumption operation sample.

[0007] The confidence-weighted incremental update module is used to acquire newly added energy consumption operation samples without alarms, input each energy consumption operation sample into the benchmark anchored distribution, calculate the probability density value corresponding to each energy consumption operation sample, map the probability density value to the training weight of the corresponding energy consumption operation sample to obtain a weighted sample set, construct a dynamic update model with the probability distribution model corresponding to the benchmark anchored distribution as the initial state, perform incremental fine-tuning on the dynamic update model based on the weighted sample set to obtain the dynamic energy consumption distribution at the current time, and take the preset upper quantile value of the dynamic energy consumption distribution as the energy consumption alarm threshold at the corresponding time.

[0008] Preferably, the benchmark anchoring distribution module includes a sample receiving unit, a quantile regression training unit, and a parameter freezing unit;

[0009] The sample receiving unit is used to receive the verified steady-state normal operation dataset of the electromechanical system and output the training sample set;

[0010] The quantile regression training unit is used to train a quantile regression model with the training sample set as input and a quantile loss function, obtain the probability distribution of normal energy consumption through multi-quantile fitting, and output the benchmark anchoring distribution.

[0011] The parameter freezing unit is used to lock all model parameters of the quantile regression model after the training of the quantile regression model is completed.

[0012] Preferably, the confidence-weighted incremental update module includes a sample acquisition unit, a confidence calculation unit, a weight mapping unit, and an incremental update unit;

[0013] The sample acquisition unit is used to construct a sliding window according to a preset window length and step size, acquire all energy consumption operation samples within the sliding window that simultaneously meet the conditions of not triggering an alarm and not being marked as abnormal, and output a new sample set; during the initial operation phase of the system, the preset upper quantile value of the benchmark anchored distribution is used as the initial alarm threshold for filtering the alarm status of the energy consumption operation samples.

[0014] The confidence calculation unit is used to input each energy consumption operation sample in the newly added sample set into the benchmark anchoring distribution and calculate the probability density value corresponding to each energy consumption operation sample.

[0015] The weight mapping unit is used to map the probability density value of each energy consumption operation sample to the corresponding training weight, and generate and output a weighted sample set.

[0016] The incremental update unit is used to incrementally fine-tune the dynamic update model with the weighted sample set as input, and output the dynamic energy consumption distribution and corresponding energy consumption alarm threshold at the current time.

[0017] Preferably, the weight mapping unit uses a linear mapping rule to convert the probability density values ​​into training weights, and the mapping formula is:

[0018] ;

[0019] in, Let i be the training weights for the i-th energy consumption running sample. Let be the probability density value corresponding to the i-th energy consumption operation sample. The peak probability density value of the benchmark anchored distribution.

[0020] Preferably, the confidence-weighted incremental update module further includes a working condition determination unit;

[0021] The operating condition determination unit is used to calculate the average confidence level of all energy consumption operation samples within the sliding window; when the average confidence level of two consecutive sliding windows is lower than a preset confidence threshold, it is determined that the electromechanical system has experienced a continuous operating condition deviation, and the value of the global weight baseline is increased by a preset step size, and the increased global weight baseline does not exceed a preset upper limit value; when the average confidence level of the sliding window rises above a preset recovery threshold, the global weight baseline is restored to its initial value; the global weight baseline is a multiplicative coefficient used to correct the training weights;

[0022] The weight mapping unit is further configured to multiply the global weight baseline with the mapped training weights to obtain the final training weights for each energy consumption operation sample.

[0023] Preferably, both the probability distribution model corresponding to the baseline anchoring distribution and the dynamic update model are quantile regression models of a neural network architecture, including a backbone network and an output layer; when incrementally fine-tuning the dynamic update model, only the output layer parameters of the dynamic update model are updated, while the backbone network parameters of the dynamic update model remain unchanged.

[0024] This invention also provides a self-learning method for energy consumption alarm thresholds of electromechanical system operating status, comprising:

[0025] Step 100: Train a probability distribution model based on the verified normal operation dataset of the electromechanical system to generate a benchmark anchored distribution of normal energy consumption under the target operating conditions of the electromechanical system; after training, freeze all parameters of the probability distribution model so that the benchmark anchored distribution remains fixed during system operation and serves as the confidence score benchmark for energy consumption operation samples; wherein, the probability density value of the energy consumption operation sample under the benchmark anchored distribution is the confidence score of the energy consumption operation sample;

[0026] Step 200: Obtain newly added energy consumption operation samples without alarms, input each energy consumption operation sample into the benchmark anchoring distribution, and calculate the probability density value corresponding to each energy consumption operation sample.

[0027] Step 300: Map the probability density value to the training weights corresponding to the energy consumption operation sample to obtain a weighted sample set; construct a dynamic update model with the probability distribution model corresponding to the benchmark anchor distribution as the initial state; perform incremental fine-tuning on the dynamic update model based on the weighted sample set to obtain the dynamic energy consumption distribution at the current time; and take the preset upper quantile value of the dynamic energy consumption distribution as the energy consumption alarm threshold at the corresponding time.

[0028] The present invention also provides an electronic device, which is a physical device, comprising:

[0029] The processor and the memory are communicatively connected.

[0030] The memory is used to store at least one executable instruction executed by the processor, which executes the executable instruction to implement a self-learning method for energy consumption alarm threshold of electromechanical system operating status as described above.

[0031] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described self-learning method for energy consumption alarm thresholds of an electromechanical system operating state.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] The benchmark anchor distribution module trains a probability distribution model based on a verified dataset of normal operation of electromechanical systems. This generates a benchmark anchor distribution of normal energy consumption under the target operating conditions of the electromechanical system and freezes all parameters of the probability distribution model. The fixed benchmark anchor distribution is used as the confidence score benchmark for energy consumption operation samples. Then, the confidence weighted incremental update module obtains newly added non-alarm energy consumption operation samples, calculates the probability density value corresponding to each energy consumption operation sample, and maps it to the corresponding training weight to obtain a weighted sample set. A dynamic update model is constructed with the probability distribution model corresponding to the benchmark anchor distribution as the initial state. The dynamic update model is incrementally fine-tuned based on the weighted sample set. The upper quantile value of the dynamic energy consumption distribution is taken as the energy consumption alarm threshold. Under the premise of maintaining the stability of the benchmark statistical characteristics, the alarm threshold can be dynamically and adaptively adjusted according to the operating conditions. This effectively alleviates the threshold offset problem caused by the accumulation of normal samples at the boundary and improves the accuracy and adaptability of energy consumption alarms. Attached Figure Description

[0034] Figure 1This is a schematic diagram of the structure of a self-learning system for energy consumption alarm thresholds of electromechanical system operating status provided in an embodiment of the present invention;

[0035] Figure 2 This is a comparison curve of the threshold drift trend between the embodiments of the present invention and the traditional unweighted incremental learning scheme;

[0036] Figure 3 This is a main flowchart of a self-learning method for energy consumption alarm threshold of electromechanical system operating status provided in an embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0038] 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.

[0039] Please see Figure 1 The present invention provides a self-learning system for energy consumption alarm threshold of electromechanical system operation status, including a benchmark anchoring distribution module and a confidence weighted incremental update module.

[0040] The benchmark anchoring distribution module is used to train a probability distribution model based on a verified dataset of normal operation of the electromechanical system, and generate a benchmark anchoring distribution of normal energy consumption under the target operating conditions of the electromechanical system. After training, all parameters of the probability distribution model are frozen so that the benchmark anchoring distribution remains fixed during system operation and serves as the confidence score benchmark for energy consumption operation samples. The probability density value of the energy consumption operation sample under the benchmark anchoring distribution is the confidence score of the energy consumption operation sample.

[0041] The confidence-weighted incremental update module is used to acquire newly added energy consumption operation samples without alarms, input each energy consumption operation sample into the benchmark anchored distribution, calculate the probability density value corresponding to each energy consumption operation sample, map the probability density value to the training weight of the corresponding energy consumption operation sample to obtain a weighted sample set, construct a dynamic update model with the probability distribution model corresponding to the benchmark anchored distribution as the initial state, perform incremental fine-tuning on the dynamic update model based on the weighted sample set to obtain the dynamic energy consumption distribution at the current time, and take the preset upper quantile value of the dynamic energy consumption distribution as the energy consumption alarm threshold at the corresponding time.

[0042] In this embodiment, the energy consumption alarm threshold self-learning system for the electromechanical system is deployed on the local monitoring controller or cloud monitoring platform of the electromechanical system. It is used to perform real-time analysis and dynamic threshold self-learning of the time-series data of the operating energy consumption of industrial electromechanical equipment such as fans, pump sets, and machine tools. This improves the problems of poor adaptability to fixed alarm thresholds, threshold creep caused by the accumulation of normal samples at the boundary during long-term incremental learning, and decreased alarm sensitivity. The system mainly includes a benchmark anchoring distribution module and a confidence-weighted incremental update module. The two modules achieve interactive transmission of distribution parameters and sample data through a standardized data interface.

[0043] Understandably, when the system is applied to the daily operation and maintenance of industrial electromechanical equipment, it can continuously ensure the rationality of alarm thresholds during long-term continuous operation of the equipment, reduce missed alarms and false alarms caused by threshold drift, reduce the ineffective troubleshooting workload of operation and maintenance personnel, assist in timely identification of abnormal deviations in equipment operating status, and ensure the stable operation of electromechanical systems.

[0044] Furthermore, the benchmark anchoring distribution module is used to construct a stable confidence benchmark throughout the entire lifecycle. Once the model training is complete, the parameters remain fixed and are not updated during subsequent system operation. This benchmark serves as the confidence score metric for all newly added samples. The probability density value of an energy consumption operation sample under the benchmark anchoring distribution represents the confidence level of that sample. A higher probability density value indicates that the sample is closer to the central region of normal operating conditions, and the higher its confidence level as a normal sample. A lower probability density value indicates that the sample is closer to the boundary of the normal distribution.

[0045] The confidence-weighted incremental update module is used to complete the weighted training and dynamic threshold update of newly added samples based on the benchmark distribution. The execution process is as follows: First, acquire the energy consumption operation samples that are newly added during system operation and have not triggered alarms. Substitute each sample into a fixed benchmark anchor distribution to calculate the corresponding probability density value. Then, map the probability density value to the training weight of the sample, so that the sample closer to the center of the distribution has a higher weight and the boundary sample has a lower weight, and summarize to obtain a weighted sample set. Construct a dynamic update model with the probability distribution model corresponding to the benchmark anchor distribution as the initial state. Perform incremental fine-tuning on the dynamic update model based on the weighted sample set to obtain the dynamic energy consumption distribution that adapts to the latest operating state at the current time. Finally, take the preset upper quantile value of the dynamic energy consumption distribution as the energy consumption alarm threshold at the corresponding time.

[0046] In this embodiment, the preset upper percentile value can be 95%, 99%, etc., corresponding to the alarm logic that the higher the energy consumption value, the more abnormal it is, ensuring that the vast majority of normal energy consumption data are below the alarm threshold.

[0047] In an optional embodiment, the benchmark anchoring distribution module includes a sample receiving unit, a quantile regression training unit, and a parameter freezing unit;

[0048] The sample receiving unit is used to receive the verified steady-state normal operation dataset of the electromechanical system and output the training sample set;

[0049] The quantile regression training unit is used to train a quantile regression model with the training sample set as input and a quantile loss function, obtain the probability distribution of normal energy consumption through multi-quantile fitting, and output the benchmark anchoring distribution.

[0050] The parameter freezing unit is used to lock all model parameters of the quantile regression model after the training of the quantile regression model is completed.

[0051] Furthermore, the sample receiving unit is used to receive the steady-state normal operation dataset of the electromechanical system verified by on-site maintenance personnel. This dataset needs to cover the complete operating cycle data under typical operating conditions of the electromechanical system, remove known fault, start-up, shutdown, and abnormal operating condition data, and output a standardized training sample set for subsequent model training.

[0052] The quantile regression training unit takes the training sample set as input and uses the quantile loss function to train the quantile regression model. Unlike the traditional regression method of fitting the mean, quantile regression can directly fit the energy consumption quantile values ​​at different confidence levels, adapting to the non-Gaussian and skewed distribution characteristics of industrial energy consumption data.

[0053] In this embodiment, the bouncing loss is used as the quantile loss function, and the expression is:

[0054] ;

[0055] Among them, For quantile levels, This represents the actual energy consumption value. The values ​​represent the quantiles predicted by the model. Training uses the Adam optimizer with an initial learning rate of 0.001 and a batch size of 32. Convergence is achieved when the validation set loss decreases by less than 1e-5 for five consecutive rounds. During training, multiple sets of continuous quantile levels with a step size of 0.01 are set within the range of 0.01 to 0.99 to obtain multiple sets of corresponding quantile values. Then, cubic spline interpolation is used to fit the discrete quantile points to obtain a continuous probability density function. The output is the baseline anchored distribution.

[0056] Alternatively, if the energy consumption data conforms to an approximately normal distribution, other probability distribution modeling methods such as Gaussian mixture models and kernel density estimation can be used to replace quantile regression, all of which can achieve the construction of a benchmark anchored distribution.

[0057] Furthermore, the parameter freezing unit is used to lock all model parameters of the quantile regression model after training is completed, ensuring that the shape, position, and scale of the benchmark anchor distribution remain unchanged during subsequent system operation, providing a stable and unified reference scale for confidence calculation throughout the entire cycle. In this embodiment, parameter locking can be achieved by setting the `requires_grad=False` attribute of the model parameters in the deep learning framework. After locking, the model can only be used for forward inference to calculate the probability density and cannot be modified by gradient updates.

[0058] In an optional embodiment, the confidence-weighted incremental update module includes a sample acquisition unit, a confidence calculation unit, a weight mapping unit, and an incremental update unit;

[0059] The sample acquisition unit is used to construct a sliding window according to a preset window length and step size, acquire all energy consumption operation samples within the sliding window that simultaneously meet the conditions of not triggering an alarm and not being marked as abnormal, and output a new sample set; during the initial operation phase of the system, the preset upper quantile value of the benchmark anchored distribution is used as the initial alarm threshold for filtering the alarm status of the energy consumption operation samples.

[0060] The confidence calculation unit is used to input each energy consumption operation sample in the newly added sample set into the benchmark anchoring distribution and calculate the probability density value corresponding to each energy consumption operation sample.

[0061] The weight mapping unit is used to map the probability density value of each energy consumption operation sample to the corresponding training weight, and generate and output a weighted sample set.

[0062] The incremental update unit is used to incrementally fine-tune the dynamic update model with the weighted sample set as input, and output the dynamic energy consumption distribution and corresponding energy consumption alarm threshold at the current time.

[0063] Furthermore, the sample acquisition unit constructs a sliding window with a preset window length and step size, and collects all energy consumption operation samples within the sliding window that simultaneously meet the conditions of not triggering an alarm and not being marked as abnormal, outputting a new sample set. In this embodiment, the window length can be set to 24 hours and the step size to 1 hour, adapting to the rolling update requirements of the daily operation cycle of industrial equipment. During the initial operation phase of the system, the dynamic update model has not yet completed its first incremental training. At this time, the preset upper quantile value of the benchmark anchored distribution is used as the initial alarm threshold to perform real-time alarm judgment on the collected energy consumption data, and samples that have not triggered alarms are selected for inclusion in subsequent incremental training, completing the sample selection logic of the system cold start phase.

[0064] The confidence calculation unit sequentially inputs each energy consumption operation sample in the newly added sample set into the benchmark anchor distribution, calculates the probability density value corresponding to each energy consumption operation sample through the probability density function, and completes the quantitative scoring of the sample confidence.

[0065] The weight mapping unit maps the probability density value of each energy consumption operation sample to the corresponding training weight, generates and outputs a weighted sample set, so that the training weight of samples closer to the center of the normal distribution is higher, and the training weight of samples closer to the boundary is lower, thereby suppressing the stretching effect of boundary samples on the model distribution from the data level.

[0066] The incremental update unit takes the weighted sample set as input and performs incremental fine-tuning on the dynamic update model. It only updates the model parameters slightly based on the new data in the window, without the need for full retraining, and outputs the dynamic energy consumption distribution and corresponding energy consumption alarm threshold at the current moment. The incremental update method can reduce the consumption of computing resources and adapt to the real-time requirements of industrial sites.

[0067] In an optional embodiment, the weight mapping unit uses a linear mapping rule to convert the probability density values ​​into training weights, and the mapping formula is:

[0068] ;

[0069] in, Let i be the training weights for the i-th energy consumption running sample. Let be the probability density value corresponding to the i-th energy consumption operation sample. The peak probability density value of the benchmark anchored distribution is the maximum probability density corresponding to the distribution center point.

[0070] Furthermore, this linear mapping rule normalizes the peak probability density of the baseline distribution, assigning a weight of 1 to samples at the distribution center. The further a sample is from the distribution center, the lower its probability density and the smaller its training weight. In the loss calculation for incremental training, the pull of boundary samples on model parameter updates is correspondingly weakened, thus mitigating the distribution widening and threshold increase problems caused by the continuous accumulation of boundary samples. Simultaneously, the linear mapping has low computational complexity and does not increase the system's computational burden.

[0071] Optionally, for operating scenarios with severe sample distribution skewness, nonlinear mapping rules such as exponential mapping and piecewise linear mapping can also be used. By adjusting the steepness of the mapping curve, the weight decay rate of boundary samples can be flexibly controlled to adapt to different operating stability requirements.

[0072] In an optional embodiment, the confidence-weighted incremental update module further includes a working condition determination unit;

[0073] The operating condition determination unit is used to calculate the average confidence level of all energy consumption operation samples within the sliding window; when the average confidence level of two consecutive sliding windows is lower than a preset confidence threshold, it is determined that the electromechanical system has experienced a continuous operating condition deviation, and the value of the global weight baseline is increased by a preset step size, and the increased global weight baseline does not exceed a preset upper limit value; when the average confidence level of the sliding window rises above a preset recovery threshold, the global weight baseline is restored to its initial value; the global weight baseline is a multiplicative coefficient used to correct the training weights;

[0074] The weight mapping unit is further configured to multiply the global weight baseline with the mapped training weights to obtain the final training weights for each energy consumption operation sample.

[0075] In this embodiment, the global weight baseline is a multiplicative coefficient used to correct the training weights. The initial default value is 1, at which time the training weights output by the weight mapping unit are not scaled. The specific way it participates in training is as follows: the original training weights output by the weight mapping unit are multiplied by the global weight baseline to obtain the final training weights of the samples, and then the final weights are substituted into the quantile loss function to participate in weighted training.

[0076] Specifically, the operating condition determination unit first calculates the average confidence level of all energy consumption operation samples within the sliding window. When the average confidence level of two consecutive sliding windows is lower than a preset confidence threshold, it determines that the electromechanical system has experienced a continuous operating condition shift. At this time, the value of the global weight baseline is increased by a preset step size, and the overall sample training weights are amplified to accelerate the adaptation speed of the dynamic update model to the new operating condition. At the same time, a preset upper limit is set to ensure that the adjusted global weight baseline does not exceed the upper limit, avoiding excessive weight amplification that could cause a significant threshold drift. When the average confidence level of subsequent sliding windows rises above the preset recovery threshold, it indicates that the operating condition has returned to the baseline state, and the global weight baseline is restored to its initial value.

[0077] In this embodiment, the preset confidence threshold is set to 0.3 times the average probability density of the baseline distribution, the preset recovery threshold is set to 0.8 times the average probability density of the baseline distribution, the typical step size for adjustment is 0.2, and the typical upper limit value is 2.0. The continuous dual-window judgment can filter out false judgments caused by instantaneous data fluctuations and random noise, and only trigger weight adjustment for continuous operating condition deviations.

[0078] In an optional embodiment, both the probability distribution model corresponding to the baseline anchoring distribution and the dynamic update model are quantile regression models of a neural network architecture, including a backbone network and an output layer; when incrementally fine-tuning the dynamic update model, only the output layer parameters of the dynamic update model are updated, while the backbone network parameters of the dynamic update model remain unchanged.

[0079] In this embodiment, the probability distribution model corresponding to the baseline anchored distribution and the dynamic update model adopt the same network structure. Both are quantile regression models with a neural network architecture, consisting of a backbone network and an output layer. The dynamic update model directly loads all parameters of the baseline model for initialization, and the two architectures are perfectly matched, enabling seamless parameter transfer. Specifically, it is a three-layer fully connected network structure. The backbone network contains two hidden layers, each with 64 neurons and using the ReLU activation function. The input dimension corresponds to univariate energy consumption time series data. The output layer is a fully connected layer with the same number of neurons as the number of quantile levels and uses a linear activation function to output the predicted energy consumption value at the corresponding quantile level. Finally, they are combined to form a complete dynamic energy consumption distribution.

[0080] Furthermore, when incrementally fine-tuning the dynamically updated model, only the output layer parameters are updated, while the backbone network parameters remain unchanged. During training, only the output layer parameters are set to an updatable state, while the parameters of the two hidden layers of the backbone network are frozen. The optimizer uses SGD with a learning rate of 0.0001, and the single-window sample training iterations are performed in 3 rounds.

[0081] In this setting, the backbone network inherits from the benchmark model, retaining the feature extraction capability and statistical properties of the benchmark distribution, which can avoid the loss of benchmark features caused by full parameter updates; at the same time, fewer parameters can be updated, which can reduce the computational cost of incremental training, and the parameter update range is limited, which can control the drift range of the threshold.

[0082] In this embodiment, a benchmark anchoring distribution module trains a probability distribution model based on a verified dataset of normal operation of the electromechanical system. This generates a benchmark anchoring distribution of normal energy consumption under the target operating condition of the electromechanical system and freezes all parameters of the probability distribution model. The fixed benchmark anchoring distribution is used as the confidence score benchmark for energy consumption operation samples. Then, a confidence-weighted incremental update module obtains newly added non-alarm energy consumption operation samples, calculates the probability density value corresponding to each energy consumption operation sample, and maps it to the corresponding training weight to obtain a weighted sample set. A dynamic update model is constructed with the probability distribution model corresponding to the benchmark anchoring distribution as the initial state. The dynamic update model is incrementally fine-tuned based on the weighted sample set. The upper quantile value of the dynamic energy consumption distribution is taken as the energy consumption alarm threshold. This allows for dynamic adaptive adjustment of the alarm threshold according to the operating condition while maintaining the stability of the benchmark statistical characteristics. This effectively alleviates the threshold offset problem caused by the accumulation of normal samples at the boundary and improves the accuracy and adaptability of energy consumption alarms.

[0083] In addition, when the system is applied to energy consumption monitoring and operation and maintenance scenarios of industrial electromechanical equipment, it can maintain the sensitivity and reliability of the alarm mechanism for a long time, reduce the occurrence of false alarms and missed alarms, assist operation and maintenance personnel in timely detection of equipment operation abnormalities, and improve the stability of electromechanical system operation and operation and maintenance management efficiency.

[0084] To verify the threshold drift suppression effect of this scheme, a simulation comparison was conducted based on a dataset of steady-state energy consumption of industrial fans. The simulation parameters were consistent with those of the aforementioned embodiments: the sliding window length was 24 hours, the step size was 1 hour, the alarm threshold was taken as the 95th percentile of the energy consumption distribution, and the simulation period was equivalent to 30 days of continuous equipment operation. The comparison scheme adopted the traditional unweighted incremental learning method, that is, all non-alarm samples participated in the incremental update of the model with the same weight.

[0085] Simulation results are as follows Figure 2 As shown, the horizontal axis represents the number of days the equipment has been running continuously, and the vertical axis represents the energy consumption alarm threshold (unit: kW) at the corresponding time. The alarm threshold of the traditional unweighted incremental learning scheme continuously increases positively with operating time, exhibiting significant cumulative drift over 30 days and displaying typical threshold creep characteristics. However, when using the scheme described in this embodiment, the alarm threshold fluctuates slightly around the initial baseline value, without a continuous positive drift trend. It only undergoes adaptive fine-tuning when there are slight fluctuations in operating conditions before quickly returning to a stable level, maintaining stable alarm sensitivity over a long period.

[0086] Based on the above embodiments, such as Figure 3 As shown, the present invention also provides a self-learning method for energy consumption alarm thresholds of electromechanical system operating status, the method comprising:

[0087] Step 100: Train a probability distribution model based on the verified normal operation dataset of the electromechanical system to generate a benchmark anchored distribution of normal energy consumption under the target operating conditions of the electromechanical system; after training, freeze all parameters of the probability distribution model so that the benchmark anchored distribution remains fixed during system operation and serves as the confidence score benchmark for energy consumption operation samples; wherein, the probability density value of the energy consumption operation sample under the benchmark anchored distribution is the confidence score of the energy consumption operation sample;

[0088] Step 200: Obtain newly added energy consumption operation samples without alarms, input each energy consumption operation sample into the benchmark anchoring distribution, and calculate the probability density value corresponding to each energy consumption operation sample.

[0089] Step 300: Map the probability density value to the training weights corresponding to the energy consumption operation sample to obtain a weighted sample set; construct a dynamic update model with the probability distribution model corresponding to the benchmark anchor distribution as the initial state; perform incremental fine-tuning on the dynamic update model based on the weighted sample set to obtain the dynamic energy consumption distribution at the current time; and take the preset upper quantile value of the dynamic energy consumption distribution as the energy consumption alarm threshold at the corresponding time.

[0090] Furthermore, step 100 specifically includes:

[0091] Receive the verified steady-state normal operation dataset of the electromechanical system and construct a training sample set;

[0092] Using the training sample set as input, a quantile regression model of a neural network architecture is trained using the quantile loss function. The probability distribution of normal energy consumption is obtained by fitting multiple quantiles, and the benchmark anchoring distribution is generated.

[0093] After the quantile regression model is trained, all model parameters of the quantile regression model are locked.

[0094] Furthermore, step 200 specifically includes:

[0095] A sliding window is constructed according to a preset window length and step size. All energy consumption operation samples within the sliding window that simultaneously meet the conditions of not triggering alarms and not being marked as abnormal are collected to obtain a new sample set.

[0096] In the initial operation phase before the dynamic update model completes its first incremental fine-tuning, the preset upper quantile value of the benchmark anchored distribution is used as the initial alarm threshold for screening the alarm status of the energy consumption operation sample.

[0097] Each energy consumption operation sample in the newly added sample set is input into the benchmark anchored distribution, and the probability density value corresponding to each energy consumption operation sample is calculated.

[0098] Furthermore, in step 300, the probability density values ​​are converted into training weights using a linear mapping rule, and the mapping formula is as follows:

[0099] ;

[0100] in, Let i be the training weights for the i-th energy consumption running sample. Let be the probability density value corresponding to the i-th energy consumption operation sample. The peak probability density value of the benchmark anchored distribution.

[0101] Furthermore, step 300 also includes:

[0102] Calculate the average confidence level of all energy consumption operation samples within the sliding window;

[0103] When the average confidence level of two consecutive sliding windows is lower than a preset confidence threshold, it is determined that the electromechanical system has experienced a continuous operating condition deviation. The value of the global weight baseline is increased by a preset step size, and the increased global weight baseline does not exceed a preset upper limit value. When the average confidence level of the sliding window rises above a preset recovery threshold, the global weight baseline is restored to its initial value. The global weight baseline is a multiplicative coefficient used to correct the training weights.

[0104] The final training weights of each energy consumption operation sample are obtained by multiplying the global weight baseline with the mapped training weights.

[0105] Furthermore, the dynamic update model is a quantile regression model with a neural network architecture, including a backbone network and an output layer;

[0106] In step 300, when incrementally fine-tuning the dynamic update model, only the output layer parameters of the dynamic update model are updated, while the backbone network parameters of the dynamic update model remain unchanged.

[0107] In this embodiment, a benchmark anchoring distribution module trains a probability distribution model based on a verified dataset of normal operation of the electromechanical system. This generates a benchmark anchoring distribution of normal energy consumption under the target operating condition of the electromechanical system and freezes all parameters of the probability distribution model. The fixed benchmark anchoring distribution is used as the confidence score benchmark for energy consumption operation samples. Then, a confidence-weighted incremental update module obtains newly added non-alarm energy consumption operation samples, calculates the probability density value corresponding to each energy consumption operation sample, and maps it to the corresponding training weight to obtain a weighted sample set. A dynamic update model is constructed with the probability distribution model corresponding to the benchmark anchoring distribution as the initial state. The dynamic update model is incrementally fine-tuned based on the weighted sample set. The upper quantile value of the dynamic energy consumption distribution is taken as the energy consumption alarm threshold. This allows for dynamic adaptive adjustment of the alarm threshold according to the operating condition while maintaining the stability of the benchmark statistical characteristics. This effectively alleviates the threshold offset problem caused by the accumulation of normal samples at the boundary and improves the accuracy and adaptability of energy consumption alarms.

[0108] Furthermore, the self-learning method for energy consumption alarm threshold of electromechanical system operating status can run the aforementioned self-learning system for energy consumption alarm threshold of electromechanical system operating status. For specific implementation, please refer to the method embodiment, which will not be repeated here.

[0109] Based on the above embodiments, such as Figure 4 As shown, the present invention also provides an electronic device, the electronic device comprising:

[0110] The processor 22 includes at least one processor 22, at least one memory 21, a communication interface 23, and a communication bus 24, wherein the processor 22 is communicatively connected to the memory 21.

[0111] In this embodiment, the memory 21 can be implemented in any suitable manner, for example, the memory 21 can be a read-only memory, a hard disk drive, a solid-state drive, or a USB flash drive, etc.; the memory 21 is used to store at least one executable instruction executed by the processor;

[0112] In this embodiment, the processor 22 can be implemented in any suitable manner. For example, the processor 22 can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers, etc.; the processor is used to execute the executable instructions to implement a self-learning method for energy consumption alarm thresholds of an electromechanical system operating state as described above.

[0113] Based on the above embodiments, the present invention also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the above-described self-learning method for energy consumption alarm threshold of electromechanical system operating status.

[0114] Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0115] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, equipment, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0116] In the several embodiments provided in this application, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or units may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or equipment, and may be electrical, mechanical, or other forms.

[0117] The modules described as separate components may or may not be physically separate. The components shown as modules 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 units can be selected to achieve the purpose of this embodiment according to actual needs.

[0118] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0119] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program instructions, such as USB flash drives, portable hard drives, read-only storage servers, random access storage servers, magnetic disks, or optical disks.

[0120] Furthermore, it should be noted that the combination of the various technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.

[0121] It should be noted that the above examples are merely specific embodiments of the present invention, and the present invention is obviously not limited to the above embodiments, with many similar variations. All modifications that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should fall within the protection scope of this invention.

[0122] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A self-learning system for energy consumption alarm thresholds in the operating state of an electromechanical system, characterized in that, This includes a baseline anchoring distribution module and a confidence-weighted incremental update module; The benchmark anchoring distribution module is used to train a probability distribution model based on a verified dataset of normal operation of the electromechanical system, and generate a benchmark anchoring distribution of normal energy consumption under the target operating conditions of the electromechanical system. After training, all parameters of the probability distribution model are frozen, so that the benchmark anchor distribution remains fixed during system operation and serves as the confidence score benchmark for energy consumption operation samples; wherein, the probability density value of the energy consumption operation sample under the benchmark anchor distribution is the confidence score of the energy consumption operation sample. The confidence-weighted incremental update module is used to obtain newly added energy consumption operation samples without alarms, input each energy consumption operation sample into the benchmark anchoring distribution, and calculate the probability density value corresponding to each energy consumption operation sample. The probability density value is mapped to the training weights corresponding to the energy consumption operation sample to obtain a weighted sample set; a dynamic update model is constructed with the probability distribution model corresponding to the benchmark anchor distribution as the initial state, and the dynamic update model is incrementally fine-tuned based on the weighted sample set to obtain the dynamic energy consumption distribution at the current time. The preset upper quantile value of the dynamic energy consumption distribution is taken as the energy consumption alarm threshold at the corresponding time.

2. The energy consumption alarm threshold self-learning system for electromechanical system operation status according to claim 1, characterized in that, The benchmark anchoring distribution module includes a sample receiving unit, a quantile regression training unit, and a parameter freezing unit. The sample receiving unit is used to receive the verified steady-state normal operation dataset of the electromechanical system and output the training sample set; The quantile regression training unit is used to train a quantile regression model with the training sample set as input and a quantile loss function, obtain the probability distribution of normal energy consumption through multi-quantile fitting, and output the benchmark anchoring distribution. The parameter freezing unit is used to lock all model parameters of the quantile regression model after the training of the quantile regression model is completed.

3. The energy consumption alarm threshold self-learning system for the operating status of an electromechanical system according to claim 1, characterized in that, The confidence-weighted incremental update module includes a sample acquisition unit, a confidence calculation unit, a weight mapping unit, and an incremental update unit. The sample acquisition unit is used to construct a sliding window according to a preset window length and step size, acquire all energy consumption operation samples within the sliding window that simultaneously meet the conditions of not triggering alarms and not being marked as abnormal, and output a new sample set; During the initial operation phase of the system, the preset upper quantile value of the benchmark anchored distribution is used as the initial alarm threshold for screening the alarm status of the energy consumption operation sample. The confidence calculation unit is used to input each energy consumption operation sample in the newly added sample set into the benchmark anchoring distribution and calculate the probability density value corresponding to each energy consumption operation sample. The weight mapping unit is used to map the probability density value of each energy consumption operation sample to the corresponding training weight, and generate and output a weighted sample set. The incremental update unit is used to incrementally fine-tune the dynamic update model with the weighted sample set as input, and output the dynamic energy consumption distribution and corresponding energy consumption alarm threshold at the current time.

4. The energy consumption alarm threshold self-learning system for the operating status of an electromechanical system according to claim 3, characterized in that, The weight mapping unit uses a linear mapping rule to convert the probability density values ​​into training weights. The mapping formula is as follows: ； in, Let i be the training weights for the i-th energy consumption running sample. Let be the probability density value corresponding to the i-th energy consumption operation sample. The peak probability density value of the benchmark anchored distribution.

5. The energy consumption alarm threshold self-learning system for the operating status of an electromechanical system according to claim 3, characterized in that, The confidence-weighted incremental update module also includes a working condition determination unit; The operating condition determination unit is used to calculate the average confidence level of all energy consumption operation samples within the sliding window. When the average confidence of two consecutive sliding windows is lower than the preset confidence threshold, it is determined that the electromechanical system has a continuous operating condition deviation. The value of the global weight baseline is increased by a preset step size, and the increased global weight baseline does not exceed the preset upper limit value. When the average confidence level of the sliding window rises above the preset recovery threshold, the global weight baseline is restored to its initial value; the global weight baseline is a multiplicative coefficient used to correct the training weights. The weight mapping unit is further configured to multiply the global weight baseline with the mapped training weights to obtain the final training weights for each energy consumption operation sample.

6. The energy consumption alarm threshold self-learning system for the operating status of an electromechanical system according to claim 1, characterized in that, Both the probability distribution model corresponding to the baseline anchored distribution and the dynamic update model are quantile regression models with a neural network architecture, including a backbone network and an output layer. When performing incremental fine-tuning on the dynamic update model, only the output layer parameters of the dynamic update model are updated, while the backbone network parameters of the dynamic update model remain unchanged.

7. A self-learning method for energy consumption alarm thresholds in the operating state of an electromechanical system, characterized in that, Includes the following steps: Step 100: Train a probability distribution model based on the verified normal operation dataset of the electromechanical system to generate a benchmark anchored distribution of normal energy consumption under the target operating conditions of the electromechanical system. After training, all parameters of the probability distribution model are frozen, so that the benchmark anchor distribution remains fixed during system operation and serves as the confidence score benchmark for energy consumption operation samples; wherein, the probability density value of the energy consumption operation sample under the benchmark anchor distribution is the confidence score of the energy consumption operation sample. Step 200: Obtain newly added energy consumption operation samples without alarms, input each energy consumption operation sample into the benchmark anchoring distribution, and calculate the probability density value corresponding to each energy consumption operation sample. Step 300: Map the probability density value to the training weights corresponding to the energy consumption operation sample to obtain a weighted sample set; construct a dynamic update model with the probability distribution model corresponding to the benchmark anchor distribution as the initial state; perform incremental fine-tuning on the dynamic update model based on the weighted sample set to obtain the dynamic energy consumption distribution at the current time; and take the preset upper quantile value of the dynamic energy consumption distribution as the energy consumption alarm threshold at the corresponding time.

8. An electronic device, characterized in that, The electronic device includes: The processor and the memory are communicatively connected. The memory is used to store at least one executable instruction executed by the processor, and the processor is used to execute the executable instruction to implement the energy consumption alarm threshold self-learning method for the operating state of an electromechanical system as described in claim 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements a self-learning method for energy consumption alarm thresholds of an electromechanical system operating state as described in claim 7.

Images