Electrical equipment fault diagnosis method based on category increment updating strategy and memory management optimization

Abstract

The invention relates to an electrical equipment fault diagnosis method based on a category increment updating strategy and memory management optimization. According to historical operation data of electrical equipment under various working conditions, feature distribution is optimized by utilizing a representation technology of converting one-dimensional time sequence data of multiple operation parameters into three-dimensional RGB images, and a fault image sample set is established; on the basis, a 34-layer residual neural network (ResNet-34) diagnosis model is constructed, and the fault type / grade of the electrical equipment is diagnosed; the method comprises the following steps of: establishing an image sample set containing newly-added fault types / grades in which the performance of the electrical equipment is continuously degraded by using the same method, nesting a sub-model which minimizes the overall classification loss of new and old faults in an original ResNet-34 diagnosis model, and based on a fault type increment updating strategy, diagnosing the types / grades of the novel faults possibly occurring in the electrical equipment; and by optimizing memory management and data storage, the high precision of the fault diagnosis model is ensured, the memory occupation is effectively reduced, and the model updating time is shortened.

Description

Technical Field

[0001] The present application relates to the technical field of electrical equipment fault diagnosis, and in particular to an electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization. Background Art

[0002] Substations are important power supply hubs, including a large number of primary equipment, secondary equipment and auxiliary equipment such as air conditioning equipment, which are prone to equipment failure. Fault diagnosis of electrical equipment in substations is of great significance for extending the service life of equipment, ensuring personnel safety, saving energy and reducing emissions, and ensuring stable, safe and efficient operation of power systems. In order to achieve accurate and efficient diagnosis of electrical equipment failures, traditional manual methods are unsustainable, and physical models based on expert knowledge also face challenges such as difficulty in model construction. Data-driven deep learning diagnostic models are increasingly showing advantages such as high diagnostic accuracy and strong versatility.

[0003] However, electrical equipment is a complex dynamic system that is constantly changing. Especially when the equipment is in service for an extended period of time and its performance is deteriorating, new faults may occur. Since the existing fault diagnosis model is trained with historical data, it cannot adapt to the distribution characteristics of new fault data, which leads to deterioration of model performance and misdiagnosis and missed diagnosis. Therefore, it is necessary to enable the existing fault diagnosis model to continuously learn real-time data and perform self-iterative updates, so as to quickly and accurately detect and diagnose new faults of electrical equipment and provide reliable support for the system's restorative control.

[0004] In addition, fault diagnosis model updates often require automatic reservation of sample memory. As the amount of data continues to increase, problems such as memory explosion and data redundancy will arise. If not resolved, this may result in reduced model update efficiency and insufficient accuracy. Currently, there is no fault diagnosis method that can effectively solve the above problems. Summary of the invention

[0005] The purpose of the present invention is to overcome the defect that the fault diagnosis model has performance degradation when facing new faults, and to provide an electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization.

[0006] The purpose of the present invention can be achieved by the following technical solutions: The present invention provides an electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization, the method comprising the following steps: Step 1: Collect historical operating data of electrical equipment and record the types of faults that occurred and the time period in which they occurred.

[0007] In this step, it is necessary to collect historical operating data of electrical equipment under normal operating conditions and specific fault conditions. At the same time, the fault type and its occurrence time should be recorded in detail to provide necessary information for subsequent data sample annotation.

[0008] Step 2: Preprocess the collected system operation data. Specifically, it includes the following steps: Step 2.1: Missing value filling. Search for missing values ​​in the collected data and fill them using the linear interpolation method; the linear interpolation method is shown in formula (1): (1) In the formula, , Respectively, they represent the time points before and after the missing values; , They are respectively represented as the values ​​corresponding to the time points before and after the missing value; , Represent the time point and fill value of missing value respectively.

[0009] Step 2.2: Eliminate data from the start-up and shutdown phases of the unit. Determine the time of the two non-steady-state operation phases of unit startup and shutdown, and directly eliminate the data from these two phases.

[0010] Step 2.3: Outlier removal: According to the Raida criterion (3𝜎 criterion), outliers in the data are searched and removed, and then filled using the linear interpolation method.

[0011] Raida Criteria (3 The criterion is as shown in formula (2): (2) In the formula, Represents the time series of operating data of a certain parameter of the system; Indicated in The value of this parameter at a point in time, ; Indicates the length of the time series; for expectations; for The standard deviation of for The outlier in At each time point, the standard deviation is usually calculated using the Bessel formula: ; express The average value of . Remove outliers Then, linear interpolation is performed using formula (1).

[0012] Step 3: Use the Fisher Score feature selection algorithm to screen the system operating parameters. Feature selection based on the Fisher Score algorithm specifically includes the following steps: Step 3.1: Use Fisher Score to measure the ability of each feature to distinguish samples and perform preliminary dimensionality reduction on the data. Specifically: Fisher Score is a filtering supervised feature selection algorithm. Its basic idea is to select features with small intra-class differences and large inter-class differences, and measure the classification ability of features based on the ratio of inter-class divergence to intra-class divergence. The larger the ratio, the stronger the feature classification ability. The definitions of inter-class divergence and intra-class divergence are as follows: Order a sample The feature set is represented as , any class sample is represented as ,Include Fault categories, then the characteristics The inter-class divergence and Class samples in The intra-class divergence under They are respectively expressed as formula (3) and formula (4): (3) (4) in, It is The amount of data for class fault samples, It is All samples of class fault have features The mean on are all samples in the feature The mean of Type of fault Samples in The value of . Then the characteristic The Fisher Score is expressed as formula (5): (5) Fisher Score penalizes intra-class variance and rewards inter-class variance. That is, the more a feature can distinguish samples, the higher its Fisher Score. Features with a score of 0 are removed to perform preliminary dimensionality reduction on the data.

[0013] Step 3.2: Sort the reduced-dimensional features in the dataset based on the Fisher Score. After calculating the Fisher Score for all features, sort them according to the scores. The top-ranked features are excellent features that are highly correlated with the fault label and have low redundancy between features.

[0014] Step 3.3: Calculate the average Fisher Score of all current features, and continue to screen features that are higher than this average. Generally, features with a Fisher Score greater than 20 are considered to have better discrimination ability, but for specific data sets and problems, the high and low standards of the Fisher Score may be different, and there is no obvious boundary for evaluating a feature based on the Fisher Score. Therefore, the present invention believes that the classification effect of features with a Fisher Score greater than the average is obvious and is an important feature for the model. Then determine the final feature subset based on the required image size so that the image size is an integer multiple of the feature dimension to facilitate subsequent work.

[0015] Step 4: The electrical equipment operation data after feature selection is classified according to the fault type, and Z-score is standardized to obtain its standard normal distribution. Then the operation data is converted into red green blue (RGB) three-dimensional images to construct an image sample set. Furthermore, the process of establishing the RGB image sample set is as follows: Step 4.1: Assume that the dataset has A fault that The signal sequence of the fault data is ,in yes Fault data The signal at time The value on It is the length of the signal, which can also be understood as the data recording duration. The RGB image representation of fault data consists of the following series of operations.

[0016] Step 4.2: After feature selection, the fault data is classified according to the fault type and Z-score normalized to obtain its standard normal distribution. , Middle Elements The calculation method of is as follows: (6) in, and They are some kind of fault data Signal Let N represent the total number of signals, Represents the size of the image. The fault signals of N channels with a length of S are standardized according to formula (6), and each group of N signals is arranged end to end in time sequence to form a one-dimensional fault signal sequence. At this time, the data length is S×N. Then, according to the image size The length is Time window Take a set of signal samples .

[0017] Step 4.3: Order data Represents a signal sample The strength value of The basic two-dimensional matrix value representing the image. Fault signal sample After calculation by formula (7), it is converted into Basic Matrix .

[0018] (7) RGB images are in three-dimensional matrix format. , represents the RGB three-dimensional matrix, which are the red (p=1), green (p=2) and blue (p=3) channels.

[0019] Step 4.4: Basic Matrix The maximum and minimum values ​​are normalized to , as shown in formula (8). Finally, It can be obtained by formula (9). The red, green, and blue pixel values ​​are scaled to 0~255.

[0020] (8) (9) After the above operation, a fault image can be generated. The next one is generated by the time window. The image is generated by sliding interception according to the specified time step H. In order to make full use of the data, the time step is determined according to the time granularity of data acquisition. Finally, an RGB image sample set is formed.

[0021] Step 5: Randomly shuffle all fault categories in the established image sample set to simulate the random occurrence of faults, and use "initial-m, increment-n" to represent the process of gradually increasing fault categories. m represents the number of fault categories that can be identified by the fault diagnosis model in the initial stage, and n represents the number of fault categories added each time. Specifically, the following steps are included: Step 5.1: Reserve sample set During the incremental category update process, the fault diagnosis model can retain a smaller set of historical fault category samples, called a reserved sample set, to review old knowledge. With the help of the reserved sample set, the model can avoid forgetting during the update, that is, it can only diagnose new faults but cannot distinguish old faults.

[0022] The present invention claims to save a fixed number of samples. For the known number of fault categories , the model is retained in each type of fault samples, of which Represents a floor function. This helps keep fixed-size samples in memory and reduces storage burden.

[0023] Step 5.2: Sample selection In order to establish a reserved sample set, it is necessary to select the most representative samples for each type of fault from the known fault samples. The Herding strategy is used for sample selection. Given a certain type of fault All samples of , Herding first calculates the class center of the current data , as shown in formula (9): (9) in, Represents the distance from each sample to the class center, and then selects the corresponding number of samples according to the distance.

[0024] Step 5.3: Divide the dataset into training and testing sets First, all fault categories in the image sample set are randomly shuffled to simulate the random occurrence of faults. Afterwards, "initial-m, incremental-n" is used to represent the process of gradually increasing fault categories. In the initial stage, the image sample set D containing m fault categories is divided into training sets with a ratio of 8:2 D 1 and test set D 2. After that, the training set D 1. Create a reserved sample set training set D* by selecting samples through the Herding strategy.

[0025] Step 6: Build a 34-layer ResNet fault image diagnosis model (ResNet-34), whose structure mainly includes convolutional layers, pooling layers, multiple residual blocks, fully connected layers, and classification layers. Change the number of residual blocks to achieve a network depth of 34 layers. The model learns weights and biases to achieve accurate fault diagnosis by minimizing the loss function that measures the difference between the diagnostic output category and the true fault category.

[0026] Step 7: Using the fault category samples of the electrical equipment in the initial stage, the ResNet-34 fault diagnosis model is trained and tested, which specifically includes the following steps: Step 7.1: Graphically represent the equipment operation data and generate several RGB image samples of different operation states. Compared with the original data, the generated image set has more obvious features and a sample size that is more suitable for deep network training.

[0027] Step 7.2: Train the ResNet network.

[0028] First, according to the image size m×m after the system running data conversion, the input image size of the ResNet model is adjusted to m×m.

[0029] Subsequently, the ResNet structure is initialized and the official pre-trained weight file is downloaded to the ResNet call. Then, a fully connected layer and a softmax classifier layer are added. The number of hidden neurons in the fully connected layer is 128, and the number of neurons in the softmax is determined according to the label category of the electrical equipment fault dataset.

[0030] Finally, the training sample set of various faults of electrical equipment in the initial stage is D 1 is used as input, its features are obtained through ResNet, and the newly added fully connected layer and softmax classifier layer are trained for fault classification. During the training process, the Adam optimizer is used for training optimization, the initial learning rate is set to 0.001, the batch size is 64, and 50 epochs are trained. The early stopping strategy is used to prevent overfitting.

[0031] Step 7.3: The training set is input into the diagnostic model in batches for training to improve the diagnostic accuracy of the model. After multiple trainings, the model converges and the test set is used. D 2. The performance of the model is evaluated, using precision, recall, specificity and accuracy as the main evaluation indicators. The stability and generalization ability of the model are evaluated through 10-fold cross validation to obtain the historical fault diagnosis model.

[0032] Step 8: For the dataset containing n newly added faults, use the same input features and image representation technology to establish an image sample set E, which is divided into training sample sets according to the ratio of 8:2 E 1 and test sample set E 2 In chronological order, the first 80% of the samples are used as the training set, and the last 20% of the samples are used as the test set. E 2 Ensure that the number of new and old fault samples is the same to ensure that the test results after the model update are credible; use the test sample set E 2. After testing the historical fault diagnosis model, it was found that the performance was degraded, so the fault diagnosis model was expanded by distillation network to enhance the model representation capability; Step 9: Based on the existing fault diagnosis model ResNet-34, the reserved sample set D* and the newly added fault training sample set E 1. Reconstruct and update the model incrementally, including the following steps: Step 9.1: When the fault diagnosis model cannot diagnose a new fault type, the model needs to perform incremental category updates based on old knowledge. By expanding the network structure based on the ResNet-34 model, the fault diagnosis model can not only diagnose new faults, but also remember the old fault types. The goal is to fit a model , so that the model's classification loss for all faults is minimized. The function is expressed as formula (10): (10) in is the hypothesis space, It is an indicator function that outputs 1 when the expression is true, otherwise it outputs 0. This makes the model discriminable for all categories, not only can it handle new categories well, but also can retain the knowledge of the original categories.

[0033] Step 9.2: Expand the ResNet-34 network under a limited memory budget. Specifically, it trains a new ResNet-34 classification model for the new fault diagnosis task, but it adds an additional model compression process through knowledge distillation to extract common features from the old backbone and integrate them with the new backbone to reduce redundancy. The difference between the new and old model backbones is represented by the function shown in equation (11): (10) By minimizing the difference between the two, the common points of the new and old ResNet classification models are extracted. This can reduce the forgetting of old knowledge, keep the number of backbone networks basically constant, and avoid catastrophic memory expansion. This model update strategy is called distillation network expansion.

[0034] Step 10: By embedding the hierarchical reinforcement learning framework, the reconstructed new and old training sample sets D*+ E 1. Optimize memory management and allocate memory when updating the fault diagnosis model, so as to achieve better memory management and data storage and improve the model update efficiency. This memory management framework is called RMM (Reinforced Memory Management). It includes the following steps: Step 10.1: Define two levels of policy functions, one for determining the memory allocation between samples and new data, called Level-1 memory; the other for allocating memory for each category, called Level-2 memory.

[0035] Step 10.2: Define the state, action, and reward during the model update process. The state refers to the variable that represents the degree of data imbalance when the model is incrementally updated. The action refers to the memory allocation operation generated by the policy function based on the state. The reward is to evaluate the quality of the policy function based on the effect of memory allocation.

[0036] Step 10.3: The concept of pseudo-category incremental updates is introduced for training the policy function. By generating pseudo-tasks, RMM can learn a policy function that adapts to different categories of incremental updates. In this way, the policy function of RMM can be directly applied to the target task without additional computation.

[0037] Step 10.4: Enable the RMM memory management technology to interact with the distilled network extension framework in real time to form an RMM-distilled network extension fault diagnosis model, dynamically update the model for newly added fault samples, and ensure the effectiveness of real-time fault diagnosis of electrical equipment.

[0038] Step 11: In the incremental model update phase, the training set includes the reserved sample set of the old fault category and the new fault category samples. After each incremental phase, the fault diagnosis model will be tested to verify the model update effect. The test set contains all the fault types that the current model is exposed to. The specific steps include: Step 11.1: Reconstruct the old and new training sample sets D*+ E 1. An extended network for training the historical diagnosis model, extracting common features from the old backbone through knowledge distillation and integrating them with the new backbone; Step 11.2: According to the principle of gradient descent, use the Adam optimizer to optimize the variable parameters in the model so that the model can learn the current sample knowledge and complete an update; Step 11.3: Set the test sample set E 2 Input into the diagnosis model, the output layer of the ResNet-34 network model outputs the probability of the sample belonging to each category through the Softmax function, and the category with the largest probability is the result diagnosed by the model; Step 11.4: Based on the test results, the accuracy of the diagnostic model is evaluated and the results are fed back to the RMM-distilled network extension framework. The model is then updated again by adjusting the state, action, reward, and policy functions until the model meets the accuracy requirements.

[0039] Compared with the prior art, the beneficial effects that can be achieved by at least one of the above technical solutions adopted in the embodiments of this specification include at least: (1) The present invention uses the time series signal image characterization technology to convert the electrical equipment time series fault signal into RGB format and establish a fault image sample set. The feature selection algorithm based on Fisher Score is used to process the data to reduce feature redundancy and effectively improve the convergence speed and diagnostic accuracy of the model.

[0040] (2) The present invention constructs an end-to-end model based on ResNet to classify electrical equipment fault images and realize the diagnosis of the fault type / level of the electrical equipment. ResNet effectively deepens the model depth, enhances the image feature extraction capability and fault classification mapping effect, and makes the model have the characteristics of fast convergence and high accuracy.

[0041] (3) The present invention uses a category incremental learning algorithm to continuously update the model so that the model conforms to the real-time data distribution. The distillation network expansion technology is used to update the model using new and old data. Through knowledge distillation and parameter training, the phenomenon of catastrophic forgetting due to the model update process is avoided, making the model more adaptable to the newly added fault data stream.

[0042] (4) The present invention applies the RMM memory optimization algorithm to reconstruct new and old fault samples during the model update process, and iterates and feeds back the results of model training to reward effective memory allocation actions. While ensuring high diagnostic accuracy, it effectively reduces memory usage, reduces model update time, and saves computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0044] Figure 1 A schematic diagram of a specific implementation flow of an electrical equipment fault diagnosis method based on a category incremental update strategy and memory management optimization provided by an embodiment of the present invention; Figure 2 A schematic diagram of the process of converting fault data into RGB images based on Fisher Score in step S1 and step S2 provided in an embodiment of the present invention; Figure 3 The difference comparison of the fault image before and after the feature selection in step S2 provided in the embodiment of the present invention; Figure 4 A schematic diagram of the network structure and parameter settings in the ResNet-34-based fault diagnosis model provided in an embodiment of the present invention; Figure 5 A model update strategy framework based on category increment provided by an embodiment of the present invention; Figure 6 A fault diagnosis model update strategy for distilled network expansion provided by an embodiment of the present invention (a comparison with a direct network expansion update strategy is given); Figure 7A flowchart of using the RMM algorithm to optimize memory management during the model update process provided by an embodiment of the present invention; Figure 8 Comparison of the diagnostic accuracy of small-scale new faults between the electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization and the benchmark diagnosis method provided by the embodiment of the present invention; Fig. 9 Comparison of the diagnostic accuracy of large-scale new faults between the electrical equipment fault diagnosis method based on the category incremental update strategy and memory management optimization and the benchmark diagnosis method provided by the embodiment of the present invention; Fig.10 Comparison of average accuracy and model update duration of the updated model before and after memory management optimization provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0046] The following describes the implementation methods of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The present application can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that, in the absence of conflict, the following embodiments and the features in the embodiments can be combined with each other. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work belong to the scope of protection of the present application.

[0047] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present application, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspect described herein can be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this device and / or practice this method.

[0048] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present application. The drawings only show components related to the present application rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0049] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the examples can be practiced without these specific details.

[0050] The technical solutions provided by various embodiments of the present application are described below in conjunction with the accompanying drawings.

[0051] The present invention relates to an electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization, aiming to solve the problem that new fault categories are generated during the operation of electrical equipment and the performance of existing fault diagnosis models deteriorates. An efficient model update technology is provided, which enables the original model to perform self-iterative updates according to the constantly changing real-time data stream, while effectively reducing the risks of model memory explosion and data redundancy, and can quickly and accurately detect and diagnose new fault types / levels of electrical equipment, thereby ensuring the safe and stable operation of electrical equipment.

[0052] Equipment fault diagnosis strongly relies on accurate and efficient diagnostic models. For complex electrical equipment, diagnostic models based on expert knowledge are facing challenges such as difficulty in modeling and difficulty in identifying model parameters. However, with the increasing automation of equipment system data collection and the increasing computing power of large computers, data-driven diagnostic models only need to rely on equipment history or online operation data to complete training and perform online fault diagnosis, which has increasingly shown advantages such as high diagnostic accuracy and strong versatility. However, due to the high coupling between the operating parameters of electrical equipment and the complex characteristics, shallow fault diagnosis models have insufficient diagnostic accuracy due to their weak ability to extract data features. Based on this, deep learning fault diagnosis models with deeper network structures have emerged.

[0053] As a classic deep learning algorithm, convolutional neural network (CNN) has the advantages of local perception, weight sharing, sparse connection, and suitability for processing image data. It can convert the one-dimensional time series data of various complex operating parameters of electrical equipment into three-dimensional RGB (Red, Green, Blue) images. While retaining the time series characteristics of the data, it can effectively refine the coupling characteristics between data. Converting data samples into image samples and then using CNN to identify the image can achieve more accurate equipment fault diagnosis.

[0054] Electrical equipment is a complex dynamic system that is constantly changing. Over time, the operating conditions, system structure, and operating performance of the equipment will change. If the historical fault diagnosis model trained with old data is still used to diagnose new data, misdiagnosis and missed diagnosis may occur. Therefore, the fault diagnosis model should not only have the ability to self-update the training data set and iteratively update the model, but also effectively solve problems such as insufficient memory and data redundancy during the model update process, which poses new challenges to the real-time fault diagnosis of electrical equipment.

[0055] Based on the above content, the design idea of ​​the present invention is to use residual neural network (ResNet) to effectively deepen the depth of CNN network, build a fault diagnosis model, and ensure the stability and robustness of the model. The operation data of electrical equipment is converted into RGB image samples to train the model, and a fault diagnosis model with high accuracy and fast convergence is obtained.

[0056] On this basis, the category incremental learning algorithm is used to update the model accordingly, so that the model conforms to the current data distribution. For the historical model trained based on historical data, the distillation network expansion technology is used to update the model using new and old data. Through knowledge distillation and parameter training, the fault diagnosis model can achieve the purpose of continuous learning. At the same time, during the model update process, the RMM memory optimization algorithm is used to reconstruct new and old fault samples, and interactive feedback iteration with the results of model training is performed to reward effective memory allocation actions. While ensuring high diagnostic accuracy, it reduces memory usage, reduces model update time, and saves computing resources.

[0057] The electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization described in this embodiment is implemented by Python 3.6 and pytorch1.1, using Pycharm as the programming environment. The experiment is carried out on a computer equipped with Intel(R) Core(TM) i7-10870H CPU @ 2.20GHz, 16.0 GB memory and NVIDIA GeForce GTX 1650 TiGPU.

[0058] Furthermore, in this embodiment, the electrical equipment fault diagnosis method based on the category incremental update strategy and memory management optimization is as follows: Figure 1 As shown, the following steps are included: S1: Collect historical operating data of electrical equipment under various working conditions D , and pre-process the historical operation data; S2: Fisher Score algorithm is used for feature selection to select several system parameters as input features; S3: Convert the one-dimensional time series data after feature selection into a three-dimensional RGB image, and attach corresponding fault labels to the RGB image according to the system operating conditions of the time series data to establish an image sample set D ; S4: Divide the image sample set in step S3 into a training sample set D 1 and test sample set D 2; S5: Construct a 34-layer residual neural network ResNet as a fault diagnosis model, using the fault training sample set divided in step S4 D 1. Model training, using the test sample set D 2. Test the model until it converges to obtain a fault diagnosis model based on historical data; S6: For the dataset containing the newly added faults, use the input features of step S2 and the image representation technology of step S3 to establish an image sample set E , according to step S4, the image sample set is divided into a training sample set E 1 and test sample set E 2; S7: Using the test sample set E 2. The historical fault diagnosis model in step S5 is tested and it is found that the performance is degraded, so the fault diagnosis model is expanded by distillation network to enhance the model representation capability; S8: For the training sample set in step S4 D 1 Automatically generate reserved sample sets D* and the training sample set in step S6 E 1. Reconstruct image samples; S9: performing real-time interactive feedback between the extended network generated in step S7 and the reconstructed training sample set based on memory management optimization in step S8; S10: embed the interactive process of step S9 in the fault diagnosis model based on historical data of step S5, so as to minimize the overall classification loss of the model for existing fault types and complete the update of the model; S11: Using the test sample set in step S6 E 2. Test the model in step S10 until the model converges to obtain a fault diagnosis model based on category incremental update.

[0059] The data samples used in the following embodiments come from ① experimental data from a commercial building in Iowa at Drexel University, ② data from the automatic fault detection and diagnosis project of Lawrence Berkeley National Laboratory, and ③ ASHRAE project 1312-RP in the United States. There are five open source fault data sets in total, and the fault detection equipment is air handling equipment (AHU). Among them, the first data set is recorded as sample one, which generates faults of different severity on three components such as fresh air valves, hot water and cooling coil valves, and introduces a total of 4 operating faults. It contains 10,800 time series data samples; the second data set is recorded as sample two, which contains 8 faults and 10,800 time series data samples; the ASHRAE project has three data sets, which are fault samples in spring, summer and winter, recorded as sample three, sample four and sample five, and five major faults are considered. Each fault has about three degrees of severity, and there are 20, 19 and 13 fault types respectively. Samples one and two are called small-scale fault categories, and samples three, four and five are called large-scale fault categories. The results of the embodiment will be displayed in this group.

[0060] The electrical equipment fault diagnosis method based on the category incremental update strategy and memory management optimization of the present invention is verified based on the experimental data in the embodiment. The adopted data set is subjected to the image representation processing based on the Fisher Score feature selection. The specific implementation steps are as follows: Figure 2 shown.

[0061] The first step is data preprocessing.

[0062] Firstly, the missing values ​​in the data are filled by linear interpolation. Secondly, the non-steady-state operation data of the chiller during the start-up and shutdown stages are eliminated. Furthermore, the abnormal values ​​in the parameters of the chiller operation data are detected according to the Raida criterion (3𝜎 criterion), and then the abnormal values ​​are eliminated and filled by linear interpolation. Finally, in order to avoid the adverse effects caused by the order of magnitude differences between the parameter data, the data are normalized according to the maximum and minimum values ​​of each parameter, and the data are unified to the range of 0 to 1.

[0063] The second step is feature selection.

[0064] Take sample 3 as an example. The data collected the operation data of AHU at one-minute intervals. Each fault day can generate 1440 data samples, and each sample in the data set has 160 features. The average FisherScore of the 160 features of the data set used is 0.3933, and a total of 46 features are higher than this average. In order to ensure that the number of features is an integer multiple of the image size, 48 feature variables are selected for analysis in each season, including air temperature, flow rate, damper, and valve position.

[0065] The third step is RGB imaging representation.

[0066] According to the image characterization technology described in the present invention, an RGB image sample set is established. After experimental verification, the image size of the present invention is determined to be 96×96. Then, the difference between the fault image before and after feature selection is compared. The results are as follows: Figure 3 As shown, it can be seen that the fault image after feature selection is more layered and organized, the overall color is deepened, and the feature expression is enhanced.

[0067] Build a ResNet-34 fault diagnosis model, optimize hyperparameters, and obtain the best diagnosis model on historical fault data. The model structure and basic parameters are as follows: Figure 4 As shown, the specific steps include the following steps.

[0068] The first step is to generate image sample sets and divide the data. After the fault data is represented by images, 700 to 1,200 images are generated according to the amount of data, and the data set is divided according to the experimental settings. Compared with the original data, the generated image set has more obvious features and a sample size that is more suitable for deep network training.

[0069] In the second step, four residual blocks are designed, their parameters and quantity are adjusted, input layer, convolution layer, pooling layer, fully connected layer and output layer are added at the beginning and end, and the depth of ResNet network is set to 34 layers. The parameter settings of each layer are shown in Figure 4 illustrate.

[0070] The third step is to train the ResNet-34 history model.

[0071] (1) The input image size of the ResNet model is adjusted to 96×96. For all electrical equipment samples, the fault categories in the initial stage are set to 3. For samples 1 and 2 of the small-scale fault category, 1 and 2 faults are added each time respectively; for samples of large-scale faults, 3 faults are added each time.

[0072] (2) Initialize the ResNet structure and download the official initial weight file to ResNet. Then, add a fully connected layer and a softmax classifier layer. The number of hidden neurons in the fully connected layer is 128, and the number of softmax neurons is determined according to the label category of the air handling unit fault sample set.

[0073] (3) Taking the RGB images of various faults of the air handling unit in the initial stage as input, the features are obtained through ResNet, and the newly added fully connected layer and softmax classifier layer are trained for fault classification. During the training process, the Adam optimizer is applied for training optimization.

[0074] The fourth step is to test the model and verify the effect of feature selection on model performance. In order to explore the impact of feature selection on model accuracy, the image datasets before and after feature selection were selected to train the ResNet-34 model separately, keeping the same data partitioning and model settings. The experiment shows that the images after feature selection can improve the diagnostic accuracy by 3.5%.

[0075] Figure 5 The flowchart of incremental model update is shown in Figure 1. It is based on the existing fault diagnosis model ResNet-34, adds new fault category samples, and when the model performance degrades, expands the network structure of the original model to perform incremental category update. Specifically, it includes the following steps: Based on the existing ResNet-34 fault diagnosis model, the commonly used network expansion method is to directly copy the backbone network to enhance the representation capability, which is called direct network expansion, such as Figure 6 (b) It trains a new ResNet classification model for each new task and aggregates features through a larger fully connected layer. The model output aggregation is expressed as follows:

[0076] in, is the backbone layer of the old fault classification model, is the newly initialized backbone layer. represents feature aggregation, is a newly initialized fully connected layer. During the model update process, the old model classification backbone layer is frozen to keep the old knowledge and avoid forgetting. However, the disadvantage of direct network expansion is that a large amount of memory must be provided to train a backbone for each incremental task.

[0077] The embodiment of the present invention adopts a distillation network expansion algorithm. Figure 6 As shown in (a), it also trains a new ResNet classification model for each new task, but it adds an additional model compression process through knowledge distillation to extract common features from the old backbone and integrate them with the new backbone to reduce redundancy. The difference between the new and old model backbones is characterized by the following formula:

[0078] By minimizing the difference between the two, the common points of the new and old ResNet classification models are extracted. It can be seen that distillation network expansion can not only reduce the forgetting of old knowledge, but also keep the number of backbone networks basically constant and avoid catastrophic memory expansion.

[0079] In addition to reducing the memory usage of the network structure itself, by embedding a RMM hierarchical reinforcement learning framework, memory allocation is performed on the reconstructed new and old fault samples when the fault diagnosis model is updated, which can achieve better memory management and data storage and improve the model update efficiency. This is called the RMM-distillation network expansion model update strategy. Figure 7 As shown in the figure. First, two levels of policy functions are defined. One is used to determine the memory allocation between samples and new data, called Level-1 memory; the other is used to allocate memory for each category, called Level-2 memory. Then, the state, action, and reward in the model update process are defined. The state refers to the variable that indicates the degree of data imbalance when the model is incrementally updated. The action refers to the memory allocation operation generated by the policy function according to the state. The reward is to evaluate the quality of the policy function based on the effect of memory allocation. Finally, the concept of pseudo-category incremental update is introduced for training the policy function. By generating pseudo tasks, RMM can learn policy functions that adapt to incremental updates of different categories. In this way, the policy function of RMM can be directly applied to the target task without additional calculations.

[0080] Figure 8 It represents the diagnostic performance of the fault diagnosis model before and after the incremental update of the category on a small-scale fault sample. Each data point in the figure represents the diagnostic accuracy of the model for the currently known fault. Figure 8 Take (a) as an example. When no new faults occur, that is, the first data point on the left side of the figure, the model performance reaches about 99%. As the equipment runs longer, a new fault appears, and the original diagnosis model's accuracy for the new fault drops to 25%, and it is unable to identify the old fault. After the incremental category update, the model accuracy remains at about 99%. Similarly, when observing Figure 8 (b) also has similar results.

[0081] Fig. 9 It shows the incremental diagnostic accuracy of each type of working condition before and after the model is updated by the present invention in the incremental update stage of large-scale fault sample categories. Figure 8 The results are similar to those of , and the model cannot accurately diagnose the increasing fault types only by relying on the existing fault sample knowledge, and the diagnostic accuracy rate plummets to as low as about 10%. However, using the updated model of the present invention, the diagnostic accuracy rate of the model for each sample set at each stage is basically maintained at about 95%.

[0082] Fig.10It is demonstrated that before and after the application of RMM memory management optimization technology, the diagnostic model based on distillation network expansion can achieve higher diagnostic accuracy more quickly with a smaller sample size during the category incremental update process. In the five sample sets verified by the present invention, the application of RMM technology saved an average of about 17% to 30% of the training time for model update, and the fault diagnosis accuracy was maintained at a high level.

[0083] The above results show that the electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization provided by the present invention has superior effect in adapting to new fault data.

[0084] The present invention converts the time series data of electrical equipment under different conditions into a three-dimensional RGB image, which not only retains the time series characteristics of the original data, but also can extract the coupling characteristics between the various characteristics. On this basis, in order to address the problem of performance degradation of the fault diagnosis model in the scenario of newly added fault types, the present invention proposes an RMM-distillation network expansion model update strategy to achieve self-update of the model according to new fault category samples, and has low requirements on the amount of data, fast model convergence speed and high accuracy, thereby ensuring the reliability and safety of electrical equipment fault detection and diagnosis. The above embodiments prove that the electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization can accurately diagnose a variety of faults of air handling units, and the model is continuously updated as the fault type increases, maintaining a high diagnostic accuracy, which provides a feasible solution for intelligent and real-time fault diagnosis of electrical equipment.

[0085] The present invention relates to an electrical equipment fault diagnosis method based on a category incremental update strategy and memory management optimization. The method optimizes feature distribution based on historical operating data of electrical equipment under various working conditions by using a characterization technology that converts one-dimensional time series data of multiple operating parameters into three-dimensional RGB images, and establishes a fault image sample set; based on this, a 34-layer residual neural network (ResNet-34) diagnosis model is constructed to diagnose the types / levels of electrical equipment faults; the same method is used to establish a new fault type / level image sample set containing the continuously deteriorating performance of electrical equipment, and a sub-model that minimizes the overall classification loss of new and old faults is embedded in the original ResNet-34 diagnosis model. Based on this fault category incremental update strategy, new fault types / levels that may occur in electrical equipment are diagnosed; and by applying RMM, a hierarchical reinforcement learning framework and pseudo-category incremental learning training, memory management and data storage are optimized, and while ensuring the high accuracy of the fault diagnosis model, memory usage is effectively reduced, model update time is reduced, and computing resources are saved. Compared with the prior art, the model of the present invention has fast convergence speed and high accuracy, can self-update according to new fault category samples, has low data volume requirements, and can improve the reliability and safety of electrical equipment operation.

[0086] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the product embodiment described later, since it corresponds to the method, the description is relatively simple, and the relevant parts can be referred to the partial description of the system embodiment.

[0087] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A method for fault diagnosis of electrical equipment based on category incremental update strategy and memory management optimization, characterized in that: The following steps are involved: 1) Collect historical operating data of electrical equipment under various working conditions D , and pre-process the historical operation data; 2) Fisher Score algorithm is used for feature selection to select several system parameters as input features; 3) Convert the one-dimensional time series data after feature selection into a three-dimensional RGB image, and attach corresponding fault labels to the RGB image according to the system operating conditions of the time series data to establish an image sample set D ; 4) Divide the image sample set in step 3) into a training sample set D 1 and test sample set D 2; 5) Construct a 34-layer residual neural network ResNet as a fault diagnosis model, using the fault training sample set divided in step 4) D 1. Model training, using the test sample set D 2. Test the model until it converges to obtain a fault diagnosis model based on historical data; 6) For the dataset containing the newly added faults, use the input features of step 2) and the image representation technology of step 3) to establish an image sample set E , according to step 4), the image sample set is divided into a training sample set E 1 and test sample set E 2; 7) Using the test sample set E 2. The historical fault diagnosis model in step 5) was tested and its performance was found to be degraded. Therefore, the fault diagnosis model was expanded by distillation network to enhance the model representation capability. 8) For the training sample set in step 4) D 1 Automatically generate reserved sample sets D* , and the training sample set in step 6) E 1. Reconstruct image samples; 9) performing real-time interactive optimization on the extended network generated in step 7) and the sample reconstructed based on memory management in step 8); 10) The interactive process of step 9) is embedded in the fault diagnosis model based on historical data in step 5) to minimize the overall classification loss of the model for existing fault types and complete the model update; 11) Using the test sample set in step 6) E 2. Test the model in step 10) until the model converges to obtain a fault diagnosis model based on category incremental update.

2. The electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization according to claim 1 is characterized in that: In step 1), the specific steps of preprocessing the historical operation data include: 11) Retrieve the location of missing values ​​in historical operation data and fill in the missing values ​​using linear interpolation method; 12) Eliminate the data during the start-up and shutdown phases of the unit; 13) Eliminate outliers in system operation data according to the Raida criterion, and use linear interpolation method to fill in data.

3. The electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization according to claim 1 is characterized in that: The specific steps of step 2) include: 21) Fisher Score is used to measure the ability of each feature to distinguish samples and perform preliminary dimensionality reduction on the data; 22) Rank the reduced-dimensional features in the dataset based on the Fisher Score; 23) Calculate the average Fisher Score of all features, and require that the Fisher Score of the selected feature is higher than the average; 24) Determine the final feature subset based on the required image size, so that the image size and the feature dimension are integer multiples, so that different fault image samples show specific regularities. The present invention recommends that the number of features be a multiple of 96, ranging from 12 to 48 features.

4. The electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization according to claim 1 is characterized in that: In step 3), the specific steps of establishing the RGB image sample set include: 31) The fault data after feature selection are classified according to the fault types, and Z-score standardization is performed to obtain their standard normal distribution; 32) Arrange each group of N characteristic data end to end in time sequence to form a one-dimensional fault signal sequence; 33) According to the image size m, a time window W is obtained to intercept a set of time series data samples, and after calculation, it is converted into an m×m basic matrix; 34) The basic matrix will be normalized to the maximum and minimum values ​​to generate an RGB fault image with three channels of red, green and blue pixels. The next image is generated by sliding the time window W at the specified time step H.

5. The electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization according to claim 1 is characterized in that: In step 4), the ratio of the number of training sample sets to the number of test sample sets is 8:

2.

6. The electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization according to claim 1 is characterized in that: In step 5), the ResNet is a deep convolutional neural network, which solves the degradation problem in the deep network training process by introducing the concept of "residual learning" to effectively deepen the network depth and improve the accuracy of fault diagnosis; its structure mainly includes a convolution layer, a pooling layer, multiple residual blocks, a fully connected layer and a classification layer; the number of residual blocks is changed to achieve a network depth of 34 layers; the 34-layer ResNet fault diagnosis model is trained using an image sample set based on historical data.

7. The electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization according to claim 1 is characterized in that: In step 6), the image sample set is divided into a training sample set E 1 and test sample set E 2, the first 80% of the samples in chronological order are used as the training set, and the last 20% of the samples are used as the test set; Among them, the test set E 2 Ensure that the number of new and old fault samples of various types is the same to ensure that the test results after the model is updated are credible.

8. The electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization according to claim 1 is characterized in that: In step 7), when the fault diagnosis model cannot diagnose the newly occurring fault type, the distillation network expansion strategy is used to incrementally update the model based on the old knowledge.

9. The electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization according to claim 8, characterized in that: It uses appropriate algorithms to expand the network under a limited memory budget; specifically, it trains a new ResNet classification model for each new task, but it adds an additional model compression process through knowledge distillation to extract common features from the old backbone and integrate them with the new backbone to reduce redundancy.

10. The electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization according to claim 1, characterized in that: In step 8), for the historical training sample set D 1 Use the Herding strategy to select samples and establish a reserved sample set D* , then E 1 and D* Will be refactored for model update.

11. The electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization according to claim 10, characterized in that: By embedding the RMM hierarchical reinforcement learning framework, memory management is optimized for the reconstructed new and old training sample sets, and memory allocation and data storage are optimized when the fault diagnosis model is updated, thereby improving the model update efficiency.

12. The electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization according to claim 1, characterized in that: In step 9), the network structure is expanded based on the ResNet-34 model so that the fault diagnosis model can not only diagnose new faults, but also remember the old fault types. The goal is to fit a model , so that the model's classification loss for all faults is minimized; during this period, RMM memory management interacts with the model training process in real time to achieve the optimal allocation of new and old data samples, thereby helping the model to be updated quickly.

13. The electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization according to claim 12, characterized in that: The RMM method propagates two levels of actions in a hierarchical manner, called Level-1 and Level-2, which represent the memory allocation between historical fault samples and new fault samples and the memory allocation for each individual fault category, respectively; then define the state, action, reward and hierarchical policy functions, create a pseudo-category incremental task learning policy function, and nest it with the distillation network expansion to form the RMM-distillation network expansion framework, and finally realize the category incremental update of the historical model, where the loss function of the category incremental update is: in, is the hypothesis space, It is an indicator function that outputs 1 when the expression is true, otherwise it outputs 0; so that the model is discriminable for all categories, not only can it handle new categories well, but also can retain the knowledge of the original categories.

14. The electrical equipment fault diagnosis method based on category incremental update strategy and memory management optimization according to claim 1, characterized in that: In step 10), the model updating and verification process includes the following steps: 101) The reconstructed new and old training sample sets D* + E 1. An extended network for training the historical diagnosis model, extracting common features from the old backbone through knowledge distillation and integrating them with the new backbone; 102) According to the principle of gradient descent, the Adam optimizer is used to optimize the variable parameters in the model so that the model can learn the current sample knowledge and complete an update; 103) Set the test sample set E 2 Input into the diagnosis model, the output layer of the ResNet-34 network model outputs the probability of the sample belonging to each category through the Softmax function, and the category with the largest probability is the result diagnosed by the model; 104) Based on the test results, the accuracy of the diagnostic model is evaluated and the results are fed back to the RMM-distilled network extension framework. Then, the model is updated again by adjusting the state, action, reward and strategy functions until the model meets the accuracy requirements.

Images