CN116756634A - Intelligent building fault diagnosis method and device based on field self-adaption - Google Patents

Intelligent building fault diagnosis method and device based on field self-adaption Download PDF

Info

Publication number
CN116756634A
CN116756634A CN202310728138.2A CN202310728138A CN116756634A CN 116756634 A CN116756634 A CN 116756634A CN 202310728138 A CN202310728138 A CN 202310728138A CN 116756634 A CN116756634 A CN 116756634A
Authority
CN
China
Prior art keywords
data
source
trained
target domain
classifier
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310728138.2A
Other languages
Chinese (zh)
Inventor
李丹
涂刚
郑子彬
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Sun Yat Sen University
Original Assignee
Sun Yat Sen University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Sun Yat Sen University filed Critical Sun Yat Sen University
Priority to CN202310728138.2A priority Critical patent/CN116756634A/en
Publication of CN116756634A publication Critical patent/CN116756634A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/10—Pre-processing; Data cleansing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/096—Transfer learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/08—Construction

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Business, Economics & Management (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Marketing (AREA)
  • Primary Health Care (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Testing And Monitoring For Control Systems (AREA)

Abstract

The invention discloses a field-adaptive intelligent building fault diagnosis method and device, wherein the method comprises the following steps: carrying out data preprocessing on sensor data in the intelligent building by combining expert knowledge in the building field, and dividing the preprocessed building data into source domain data and target domain data; pseudo labels are given to the target domain data without labels according to the trained source classifier and a preset data mining model; performing challenge-based transfer learning to obtain a trained target encoder and an updated source encoder; parameter optimization is carried out on the trained target encoder and the updated source encoder by using a preset center loss optimizer so as to enhance the gaps among different categories in the public feature space; and performing intelligent building fault diagnosis by using the trained sharing classifier to obtain an intelligent building fault diagnosis result. The model disclosed by the invention has a simple structure, can solve the problem of potential type insensitivity in traditional migration learning, and can ensure that the migrated model has high accuracy.

Description

Intelligent building fault diagnosis method and device based on field self-adaption
Technical Field
The invention belongs to the technical field of fault diagnosis, and particularly relates to a field self-adaption-based intelligent building fault diagnosis method and device.
Background
Along with the rapid development of information technology and sensor networks, intelligent buildings bring much convenience to the life of people, and the intelligent buildings are internally provided with intricate and complex network structures, so that the intelligent buildings are very high in energy consumption as the buildings, and particularly the heating, ventilation and air conditioning systems in the intelligent buildings. The energy consumption of building hvac systems is statistically 40% of the world, most of which is wasted energy due to equipment maintenance deficiencies, component aging, improper operation and control. The intelligent building fault detection and diagnosis can help to improve the energy saving capability and comfort and safety.
With the development of machine learning algorithms and the development of deep learning techniques, there are various methods or models for detecting and diagnosing faults of intelligent buildings in the prior art, but it is very difficult to design a perfect model for different buildings running under various conditions caused by different climates, seasons, outdoor temperatures and the like. The existing model or transfer learning algorithm is too complex in structure, so that the problems of high calculation cost, long training time and the like are caused, and the model accuracy is reduced or the requirement on the complexity of model design is increased due to seasonal changes, normal working condition changes of devices and the like.
Disclosure of Invention
The invention aims to provide a field-adaptive intelligent building fault diagnosis method and device, which are used for solving the technical problems of complex structure and lower detection result precision of the existing intelligent building fault detection model.
The aim of the invention can be achieved by the following technical scheme:
the intelligent building fault diagnosis method based on the field self-adaption comprises the following steps:
carrying out data preprocessing on sensor data in an intelligent building by combining expert knowledge in the building field to obtain preprocessed building data, and dividing the preprocessed building data into labeled source domain data and target domain data only with partial labels;
pseudo labels are given to the target domain data without labels according to a trained source classifier and a preset data mining model, the target domain data given with the pseudo labels are regarded as the target domain data with labels, and the trained source classifier is obtained by training the source domain data with labels;
performing challenge-based migration learning to obtain a trained target encoder and an updated source encoder, so as to migrate both the source domain data and the target domain data to a new public feature space;
Performing parameter optimization on the trained target encoder and the updated source encoder by using a preset center loss optimizer so as to enhance differences among different categories in the common feature space;
and performing intelligent building fault diagnosis by using the trained shared classifier to obtain an intelligent building fault diagnosis result, wherein the trained shared classifier is obtained by training the source domain data and the target domain data with the labels.
Optionally, performing data preprocessing on sensor data in the intelligent building in combination with expert knowledge in the building field to obtain preprocessed building data includes:
performing data inspection on the sensor data and deleting missing values;
extracting features of the sensor data, performing dimension reduction processing on the extracted features by adopting an extremely random tree, and reserving sensitive features meeting preset conditions;
and combining expert knowledge in the building field to obtain expert features, and taking the expert features and the sensitive features as preprocessed building data.
Optionally, labeling the target domain data without labels according to the trained source classifier and a preset data mining model includes:
Updating the trained source classifier by using target domain data with labels to obtain an updated source classifier, and classifying data samples in the target domain data by using the updated source classifier to obtain first-class labels;
classifying the data samples by using the preset data mining model with the tag diffusion algorithm to obtain second class tags;
and when the first type tag and the second type tag are the same, taking the first type tag or the second type tag as a pseudo tag of the data sample.
Optionally, updating the trained source classifier by using the tagged target domain data to obtain an updated source classifier includes:
initializing a target domain encoder by using parameters of the source domain encoder;
mapping the tagged target domain data using the target domain encoder;
and updating the trained source classifier by using the mapped target domain data with the labels to obtain an updated source classifier.
Optionally, the loss function of the central loss optimizer is:
wherein ,is center loss, x i I data point representing each category, +.>Representing the center point of the category, batch_size represents the amount of data that participates in the calculation at the time of the current training.
The invention also provides an intelligent building fault diagnosis device based on the field self-adaption, which comprises:
the data preprocessing module is used for carrying out data processing on sensor data in the intelligent building by combining expert knowledge in the building field to obtain preprocessed building data, and dividing the preprocessed building data into labeled source domain data and target domain data only with partial labels;
the pseudo tag giving module is used for giving pseudo tags to the target domain data without tags according to a trained source classifier and a preset data mining model, and regarding the target domain data given with the pseudo tags as the target domain data with the tags, wherein the trained source classifier is obtained by training the source domain data with the tags;
the resistance transfer learning module is used for performing resistance-based transfer learning to obtain a trained target encoder and an updated source encoder so as to transfer the source domain data and the target domain data to a new public feature space;
the inter-class gap enhancement module is used for carrying out parameter optimization on the trained target encoder and the updated source encoder by using a preset center loss optimizer so as to enhance the gaps among different classes in the public feature space;
The fault diagnosis module is used for performing intelligent building fault diagnosis by using the trained shared classifier to obtain an intelligent building fault diagnosis result, and the trained shared classifier is obtained by training the source domain data and the target domain data with the labels.
Optionally, the data preprocessing module performs data preprocessing on the sensor data in the intelligent building in combination with expert knowledge in the building field to obtain preprocessed building data, where the preprocessing includes:
the data preprocessing module performs data inspection on the sensor data and deletes missing values;
extracting features of the sensor data, performing dimension reduction processing on the extracted features by adopting an extremely random tree, and reserving sensitive features meeting preset conditions;
and combining expert knowledge in the building field to obtain expert features, and taking the expert features and the sensitive features as preprocessed building data.
Optionally, the pseudo tag assigning module assigns a pseudo tag to the target domain data without the tag according to the trained source classifier and a preset data mining model, including:
the pseudo tag giving module updates the trained source classifier by using target domain data with tags to obtain an updated source classifier, and classifies data samples in the target domain data by using the updated source classifier to obtain a first type of tag;
Classifying the data samples by using the preset data mining model with the tag diffusion algorithm to obtain second class tags;
and when the first type tag and the second type tag are the same, taking the first type tag or the second type tag as a pseudo tag of the data sample.
Optionally, the updating the trained source classifier by the pseudo tag giving module by using the tagged target domain data to obtain an updated source classifier includes:
the pseudo tag giving module initializes the target domain encoder by using parameters of the source domain encoder;
mapping the tagged target domain data using the target domain encoder;
and updating the trained source classifier by using the mapped target domain data with the labels to obtain an updated source classifier.
Optionally, the loss function of the central loss optimizer in the inter-class gap enhancement module is:
wherein ,is center loss, x i I data point representing each category, +.>Representing the center point of the category, batch_size represents the number of data points in each category.
The invention provides a field-adaptive intelligent building fault diagnosis method and device, wherein the method comprises the following steps: carrying out data preprocessing on sensor data in an intelligent building by combining expert knowledge in the building field to obtain preprocessed building data, and dividing the preprocessed building data into labeled source domain data and target domain data only with partial labels; pseudo labels are given to the target domain data without labels according to a trained source classifier and a preset data mining model, the target domain data given with the pseudo labels are regarded as the target domain data with labels, and the trained source classifier is obtained by training the source domain data with labels; performing challenge-based migration learning to obtain a trained target encoder and an updated source encoder, so as to migrate both the source domain data and the target domain data to a new public feature space; performing parameter optimization on the trained target encoder and the updated source encoder by using a preset center loss optimizer so as to enhance differences among different categories in the common feature space; and performing intelligent building fault diagnosis by using the trained shared classifier to obtain an intelligent building fault diagnosis result, wherein the trained shared classifier is obtained by training the source domain data and the target domain data with the labels.
In view of this, the beneficial effects brought by the invention are:
the data sources of the data set in the invention not only directly measure the data obtained by the fault device by the sensor, but also have some characteristics summarized by expert knowledge in the building field, the characteristics reflect the difference of certain indexes under normal and fault conditions, and a more accurate diagnosis result is provided for the model; the trained source classifier and an agreement mechanism of a preset data mining model are utilized to label the target domain data without labels, so that the reasonable application of enough label-free data can be ensured, and meanwhile, the high accuracy of the label-free data is realized; the field self-adaptive method in the transfer learning is used, the problem that the data distribution difference of different categories disappears after the two-field data is transferred is concerned, the data distribution difference between the different categories can be maintained, the potential category insensitivity problem in the traditional transfer learning can be solved, and therefore the transferred model is guaranteed to have high accuracy.
Drawings
FIG. 1 is a schematic flow chart of the method of the present invention;
FIG. 2 is a schematic diagram of steps in an embodiment of the method of the present invention;
fig. 3 is a schematic view of the structure of an embodiment of the device of the present invention.
Detailed Description
Term interpretation:
transfer learning (Transfer Learning, TL): transfer learning is a machine learning technique that enables models trained on a task to be updated or otherwise manipulated for use on a new task that is related to but different from the original task. When available data are limited, a new model cannot be trained from scratch, or a new task is similar to the original task enough, the pre-trained model can adapt to a new problem only by slightly modifying, and transfer learning can well help us to complete.
Domain adaptation (Domain Adaptation, DA): domain adaptation is a class of methods for transfer learning, which has more specific usage scenarios and requirements. In a field-adaptive scenario, we have two fields, one is a source field with a complete data tag and one is a target field with usually no tag or few tags. What we need to do is that the model trained on the source domain can be used on the tasks of the target domain through correlation operations.
Semi-supervised learning: semi-supervised learning is a scenario under machine learning. In semi-supervised learning, only a part of data is tagged, and the part of tagged data and non-tagged data need to be used to complete tasks.
Pseudo tag: in a semi-supervised learning scenario, some data may not have tags. By means of the algorithm, we can initially infer their labels by means of the links between them and the tagged data. The label thus obtained is a pseudo label.
Label diffusion algorithm (labeldiffusion): this is a semi-supervised algorithm. The labeled data points propagate label information to unlabeled data points.
Extreme random trees (ExtraTrees), an ensemble learning technique, essentially consists of a number of decision trees. The input data for each decision tree in the extremely random tree is a subset of the original training samples. If used as a feature selection, consider that all data is entered, including training sets and test sets. Each tree is trained by randomly selecting a certain number of features, and the importance of one feature is calculated as a segmentation decision mathematical index brought by the feature.
Full tie layer: in deep neural networks, if each input element is connected to each output element in a layer, such a layer is called a fully connected layer
Center Loss function (Center Loss), which is calculated as the square of the distance of each class of data points from the feature Center point, the optimizer is responsible for optimizing the two-domain mapper parameters so that each class of data points can be infinitely close to its Center point.
The embodiment of the invention provides a field self-adaption-based intelligent building fault diagnosis method and device, which are used for solving the technical problems that an existing intelligent building fault detection model is complex in structure and low in detection result accuracy.
In order that the invention may be readily understood, a more complete description of the invention will be rendered by reference to the appended drawings. Preferred embodiments of the present invention are shown in the drawings. This invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and/or" as used herein includes any and all combinations of one or more of the associated listed items.
It was reported that in 2017, the cost of building failure maintenance exceeded $70 billion. Another report of Reports and Data in 2021 predicts that the global building analysis market for intelligent building fault detection and diagnosis methods and real-time performance monitoring will reach $152.1 billion by 2026. The statistics show that the intelligent building fault detection and diagnosis has great research prospect and potential market.
Over the past few years, as machine learning algorithms and deep learning techniques have evolved, initially people have utilized tagged data to obtain a number of methods or models that can be used for intelligent building fault detection and diagnosis. For example, training a neural network, given an input dataset and an output label dataset, one can easily give the parameters of the neural network to direct training. In practice, however, the data obtained by the intelligent building management system from the sensors is unlabeled. If these data are manually marked, not only a lot of time costs and effort are incurred, but also no real-time performance is provided at all.
The prior art mainly has the following defects:
(1) The designed model or the migration learning algorithm has too complex structure, so that the problems of high calculation cost, long training time and the like are caused. In the invention, the diagnosis model is a simple deep neural network, the migration algorithm step only comprises five steps of pre-training, pseudo-labeling, anti-migration, center loss optimization and shared classifier training, and each step has very simple implementation details and clear flow, thereby being very suitable for the requirements of rapid use in an industrial scene;
(2) And after the knowledge of the source domain model is not concerned to migrate to the target domain model, the distribution condition of the two-domain data after the two-domain data are processed by the respective models is not concerned. This is likely to cause the data distribution differences between each class of the two domains to disappear, and overfitting is likely to occur when training the classifier available for the target domain data. After the migration is realized, the invention can further adjust according to the data distribution condition, so as to avoid the condition of fitting when the classifier is trained later;
(3) The universality is not good. The prior art taught above is more suitable for handling the situation that the similarity of two-domain data is already high, and if the similarity of two-domain data is very low, the performance is greatly reduced, because the migration result is affected, and errors are easily caused when the classifier is trained later. According to the invention, as the data distribution situation after migration is checked, the same kind of data is gathered together, so that different kinds of data are dispersed as far as possible, and the migration result is influenced even though the original similarity of the two-domain data is low, and the negative influence caused by the two-domain data can be removed after the intermediate step adjustment of the invention;
(4) The sensor measurement data is directly utilized, and other indexes influencing the fault diagnosis result are not added according to expert knowledge summary. The invention combines the advantages of the rule-based and data-driven building fault diagnosis method, and utilizes expert summarized knowledge to enable the data-driven method to have higher diagnosis accuracy.
In a real situation, models learned in other positions or under working conditions can be applied to tasks needing to be processed at present after certain processing, so that the design, calculation and training costs of a plurality of models can be saved. Such ideas come from transfer learning, where the model that has been learned is referred to as the source domain model and the data used for this model training is referred to as the source domain data. Accordingly, the data and models used by the currently processed task are referred to as target domain data and target domain models.
The traditional machine learning method often requires that the source domain and the target domain are identical in data distribution, and the limitation greatly restricts the expansion of application scenes and breaks through bottlenecks. In recent years, as a hot branch in machine learning, transfer learning attracts attention of a large number of researchers, and is considered as a driving force for promoting progress in the machine learning era. By using the deep neural network which is popular at present, the deep migration learning technology can be realized, the knowledge of the source domain model is migrated into the target domain model, and an excellent target domain model is trained under the condition that the target domain has no label or few labels.
Deep migration learning can be divided into the following three categories: (1) example-based. The method considers that if the two-domain model has certain mobility, the two-domain data also have some common characteristics, and some data in the source domain can be directly used when the target domain model is trained. The specific method is that after a proper weight value is applied to the used source domain data, the source domain data and the target domain data are mixed together and used for training a target domain model; (2) model-based. Such methods consider that the model structure of the two domains and a part of the parameters can be shared. The most classical approach of this kind is pre-training-updating adjustment, i.e. training out the source domain model and then using the target domain data to adjust the parameters. (3) feature-based. Such methods attempt to find a data characterization expression that is applicable to both fields. The three methods have a certain progressive relationship, namely, transition from concrete data to a model for processing the data and then to abstract high-level characteristic expression.
With the development of machine learning algorithms, data driven methods make use of rich data, which is unique in fault detection and diagnosis, and various implementations exist. The method is an information explosion age, the data acquisition amount is greatly increased, and the demand of the machine learning model is also increased. While in the real world many tasks are similar. If appropriate adjustments can be made to the model previously trained in a field, it can be migrated to the target task. This is transfer learning, which can reduce model adaptation problems due to differences between data domains. After the source domain and target domain data are mapped to the same feature space, the difference between each class may disappear, which may cause a problem of over-fitting when the classification task is finally done, and thus needs to be considered.
Therefore, an efficient fault diagnosis scheme is particularly important, which can reduce or avoid property loss, energy loss and even casualties caused by faults. There are roughly three types of methods for fault detection and diagnosis: model-based, signal-based, and data-driven. The model-based approach compares the relationship between sensor inputs and model outputs from the underlying hardware level by modeling the physical system. The signal-based method utilizes the correlation of specific faults and corresponding system output signals to judge. The data driven method uses the existing fault types to mine the hidden information behind a large amount of sensor data.
In the invention, a domain adaptation model (transfer learning model) sensitive to the class is provided, after a source encoder of a source domain, a target encoder of a target domain, a discriminator and a source classifier are trained, the difference between classes is increased in a mapped feature space, namely the difference between the classes is enhanced, and then a shared classifier is trained. The method is applied to judging the same fault type in different seasons, and the efficiency can be found. In the training process, the invention combines expert summarized knowledge, and adds some characteristic indexes summarized by the expert on the basis of the sensor measurement data, and the indexes can influence the fault diagnosis result after being evaluated.
In a first aspect, referring to fig. 1, the present invention provides an embodiment of a smart building fault diagnosis method based on domain adaptation, including:
s100: carrying out data preprocessing on sensor data in an intelligent building by combining expert knowledge in the building field to obtain preprocessed building data, and dividing the preprocessed building data into labeled source domain data and target domain data only with partial labels;
s200: pseudo labels are given to the target domain data without labels according to a trained source classifier and a preset data mining model, the target domain data given with the pseudo labels are regarded as the target domain data with labels, and the trained source classifier is obtained by training the source domain data with labels;
s300: performing challenge-based migration learning to obtain a trained target encoder and an updated source encoder, so as to migrate both the source domain data and the target domain data to a new public feature space;
s400: performing parameter optimization on the trained target encoder and the updated source encoder by using a preset center loss optimizer so as to enhance differences among different categories in the common feature space;
S500: and performing intelligent building fault diagnosis by using the trained shared classifier to obtain an intelligent building fault diagnosis result, wherein the trained shared classifier is obtained by training the source domain data and the target domain data with the labels.
In step S100, performing data preprocessing on sensor data in a smart building in combination with expert knowledge in the building field to obtain preprocessed building data includes:
performing data inspection on the sensor measurement data and deleting missing values;
extracting features of sensor measurement data, performing dimension reduction processing on the extracted features by adopting an extremely random tree, and reserving sensitive features meeting preset conditions;
and combining expert knowledge in the building field to obtain expert features, and taking the expert features and the sensitive features as preprocessed building data.
In the invention, the original data is sensor measurement data from a building, the original data is preprocessed by the used model, and other characteristics which have influence on fault diagnosis results and are summarized by intelligent building fault experts are added on the basis of the original sensor data. Mainly comprises the following steps: (1) Checking data, namely deleting the measurement data of the time point with the missing value; (2) feature extraction: selecting a plurality of features with the largest correlation with the fault type; (3) adding expert knowledge: on the basis of the original sensor data, other characteristics which are summarized by intelligent building fault experts and have influence on fault diagnosis results are added.
In some embodiments of the present invention, the data preprocessing process is specifically:
(1) Data checking, deleting the measurement data at the time point with the missing value.
Specifically, the data used in the embodiment of the invention is measurement data returned by each sensor in the building at a corresponding time point, and the set of fault labels to which the source domain data and the target domain data belong is the same. There may be some missing values of the data transmitted back from the sensor for various reasons. When the missing values are less, the quantity of the time node data containing the missing values does not greatly affect the model migration effect, for example, the missing values do not exceed 10% of the quantity of the original data, and then the missing values can be directly selected to be deleted.
(2) And extracting the characteristics, and selecting a plurality of characteristics with the largest correlation with the fault type.
In order to reduce the calculation cost during training the neural network, improve the sensitivity of the network to input features and improve the model performance, certain dimension reduction processing is required for the original data with a large number of features. In one embodiment of the invention, an extremely random tree scheme is selected. The expert summarizes the characteristics, and selects several types of characteristics with the greatest influence on fault diagnosis results from various types of data directly measured by the sensor. The selection of the number of extracted features can calculate the importance index of each feature by using extremely random tree for the source domain of the original data, the features with higher selection index are remained, and the threshold can be set at 10 -5 ～10 -3 。
(3) New features are added using expert summarized knowledge.
The data measured by the sensor directly on the faulty device reflects the measurement results of some characteristics under the fault condition, and the expert summarizes the data by utilizing the difference between the data under the fault condition and the data under the normal condition to obtain some new data characteristics. The characteristics summarized by the experts are added into the data set, so that more objective and accurate can be achieved.
Training a source classifier by using source domain data, training a discriminator by using source domain and target domain data, adding the difference between different classes, optimizing the mappers of the two domains, and finally training a sharing classifier. The source data and the target data refer to the construction data after pretreatment.
The overall flow is shown in fig. 2, where the dashed line indicates that the portion, except the data and tag, does not require training and the solid line indicates that it is required. The structures of the encoder, the classifier and the discriminator are all neural networks, the model comprises two parts of the encoder and the classifier, namely the source model comprises two parts of the source encoder and the source classifier, and the target model comprises two parts of the target encoder and the target classifier.
As a challenge-based migration learning method, embodiments of the present invention desirably allow the data distribution of the source domain and the target domain to be similar, thereby enabling migration, including encoders and classifiers. As a first step, pre-training is provided by the source encoder M for migration s Sum source classifier C s 。
Compared with the prior method which utilizes a complex decision tree structure or a complex neural network, the embodiment of the invention has the advantages that the model structure has less sensitivity to data because the data is subjected to feature extraction and expert knowledge is added, that is, the model structure (source encoder M s Sum source classifier C s ) The design can be simpler, so that a neural network connected by a full-connection layer is adopted, the set number of neurons in each layer is recommended to be less than 1000, and the specific number does not greatly influence the final model performance.
Designed source encoder M s Sum source classifier C s After the neural network structure, the embodiment of the invention carries out regular initialization on the linear parameters, then trains by using the source data with the labels, and hopes that Cs can have higher classification accuracy on the data mapped by Ms. During training, the loss lambda of the source classifier is minimized Cs (X s ,Y s ) Namely, the optimization target shown in the formula (1) is realized:
wherein min represents that the loss function optimization objective is minimization, M s and Cs Representing the source encoder and source classifier obtained by this optimization, x s The source domain is represented by a representation of the source domain,x t representing the target domain, E representing the desire, E (xs,y)～(Xs,Ys) The set of labels obtained by the source domain model representing the source domain data expected to be input by the embodiment can be distributed as the original set of source domain data and labels (the same shall apply to the following formulas), y s Refers to the label of model prediction, y s K represents the predicted label as the original label, and K represents the maximum label number (starting from 0).
It should be noted that, after pre-training, the trained source encoder parameters and source classifier parameters can be obtained.
In one embodiment of the present invention, pseudo-tagging the untagged target domain data according to a trained source classifier and a preset data mining model comprises:
updating the trained source classifier by using the target domain data with the labels to obtain an updated source classifier, and classifying the data samples in the target domain data by using the updated source classifier to obtain first-class labels; classifying the data samples by using a preset data mining model with a tag diffusion algorithm to obtain a second type tag; when the first type of tag and the second type of tag are the same, the first type of tag or the second type of tag is used as a pseudo tag of a data sample in target domain data without the tag.
Because the data labels are needed later when the gap between classes is enhanced, only a part of the target domain data is labeled in the semi-supervised learning scene, wherein the data with the same label number in each class is needed, and the pseudo labels are needed to be given to the rest of the unlabeled target domain data.
In the embodiment of the invention, a preset agreement mechanism of the data mining model and the neural network model is adopted to assign the pseudo tag. Specifically, for one data sample in the unlabeled target domain data, when the data mining model and the neural network model (source classifier) agree on the label judged by the data mining model and the neural network model, the agreed label is taken as a pseudo label of the data sample.
In a preferred embodiment, the data mining model in this embodiment adopts a label diffusion algorithm, and the idea is to diffuse information about labels from labeled data to unlabeled data, so that the data mining model is a very excellent semi-supervised learning algorithm.
In this embodiment, the neural network model used when the label is pseudo-tagged is an adjusted source domain model updated with the labeled target domain data. The specific process of updating and adjusting is as follows: firstly, initializing target domain encoder parameters by using trained source domain encoder parameters; then, inputting the target domain data with the labels into a target domain encoder, and performing mapping processing such as linear transformation on the target domain data with the labels by using the target domain encoder; and inputting the mapped target domain data with the labels into a source classifier, and performing fault classification by using the source classifier to obtain corresponding class labels.
When training is performed by using the labeled target domain data, the classifier adopts the source domain classifier, and at this time, the parameters of the source domain classifier are updated along with the training to obtain the updated source classifier. In each iteration, the embodiment of the invention can adopt an agreement mechanism, specifically: a data mining model based on similarity is selected, the model can directly give a pseudo tag according to the similarity between label-free data and labeled data, and when the data mining model and a target domain model agree on the tag judgment of a certain data sample, the tag is used as the pseudo tag of the data sample and is equivalent to the target domain data with the tag, so that the updating and adjustment of the next round of source domain model are influenced.
In step S300, a target encoder trained based on countermeasure-based migration learning and an updated source encoder are performed to migrate both the source domain data and the target domain data to a new common feature space.
In challenge-based migration learning, there are generally two choices for migration direction: migration of one domain to another domain is typically due to the choice of the active encoder and source classifier to migrate the target domain to the source domain, or both domains to a new space. The method selects the latter, so that the problem of overfitting of a source encoder and a source classifier to source domain data during pre-training caused by overlarge distribution difference of two domain data can be avoided.
Based on the idea of generating-fighting, we first train a discriminant D that can judge that a given data is from the source domain (i.e., belonging to x s ) Or the target domain (i.e. belonging to x t ) Then using a source encoder M s Model parameters to initialize the target encoder M t Training the target encoder, updating and adjusting the source encoder, so that the discriminator cannot recognize that the input data comes from the domain after the two-domain encoder maps the data, and specifically realizing that the data is jointly demapped to a new domain label on the domain label instead.
In summary, assuming that the source domain and the target domain are labeled 0 and 1, respectively, and the new domain is labeled 0.5, we can generalize the resistance migration to the optimization of the following calculation formula:
wherein ,ΛMs Representing a loss function, Λ, when the arbiter D discriminates the source domain data Mt Then represents the loss function of the arbiter D when discriminating the target domain data, Λ D Finally, after mapping the two-domain data, the two-domain encoder is enabled to judge that the domain label of the mapped data is a loss function of 0.5 by the arbiter D.
Wherein, equation (2) and equation (3) represent training of the arbiter, and equation (4) represents training of the target encoder and update adjustment of the source encoder. Combining the three formulas together can obtain the optimization target of the resistance migration: The method has the specific significance that parameters of the two-domain encoder are adjusted by means of the discriminator, so that two-domain data mapped by the two-domain encoder can reach a new public feature space.
In step S400, parameter optimization is performed on the trained target encoder and the updated source encoder by using a preset center loss optimizer to enhance the gap between different categories in the common feature space.
After the resistance migration is completed, although the source domain data and the target domain data are mapped to a new public feature space, the situation that the inter-class difference disappears may occur, which may cause a fitting problem when the classifier is trained later. Therefore, in the embodiment of the invention, in order to realize the difference between the salient classes after mapping, the same class is concentrated as much as possible and different classes are dispersed as much as possible, thereby improving the accuracy of the classifier, and introducing a center loss optimizer to optimize the two-domain encoder.
The loss function in the center loss optimizer is the center loss shown in equation (5), which is calculated as the square of the distance of each class of data point from the feature center point, and the optimizer is responsible for optimizing the two-domain encoder parameters so that each class of data point can be compared to its center point c yi Infinite access.
wherein ,is center loss, x i I data point representing each category, +.>Representing the center point of each category, batch_size represents how many pieces of data are currently being trained, i.e., the amount of data that participates in the calculation when currently being trained.
With respect toGradient sum->The updated formula is shown in formula (6) and formula (7):
after the completion of the resistance migration and the center loss optimization, the classifier needs to be retrained once due to the change of the mapping structure, and the classifier is trained by using the labeled source domain data and the labeled target domain data, so as to obtain a trained shared classifier C sh The method is suitable for data after two-domain mapping, and the optimization target is shown in a formula (8):
the testing process is to test the performance of the final model, namely the target encoder and the shared classifier, which is shown on the target domain test set data, and the main evaluation index is the classification accuracy.
In step S500, the intelligent building fault diagnosis is performed by using the trained shared classifier, which is obtained by training the source domain data and the labeled target domain data, to obtain the intelligent building fault diagnosis result.
Specifically, target data without labels are input into a trained shared classifier, and a fault classification result of the intelligent building is obtained.
According to the invention, expert summarized knowledge is utilized to compare measurement differences under fault and normal working conditions, and new characteristics are added; an agreement mechanism is used for solving the problem that the accuracy and the credibility of the pseudo tag in semi-supervised learning are not high enough; the field self-adaptive framework used by the invention combines semi-supervised learning and data mining algorithms, focuses on the problem that the data distribution difference of different categories disappears after two-domain data migration, and can solve the problem of potential type insensitivity in traditional migration learning.
In the embodiment of the invention, the data sources of the data set not only directly measure the data obtained by the fault device by the sensor, but also have some characteristics summarized by expert knowledge in the building field, and the characteristics reflect the difference of certain indexes under normal and fault conditions, so that a more accurate diagnosis result is provided for the model; the trained source classifier and an agreement mechanism of a preset data mining model are utilized to label the target domain data without labels, so that the reasonable application of enough label-free data can be ensured, and meanwhile, the high accuracy of the label-free data is realized; the field self-adaptive method in the transfer learning is used, the problem that the data distribution difference of different categories disappears after the two-field data is transferred is concerned, the data distribution difference between the different categories can be maintained, the potential category insensitivity problem in the traditional transfer learning can be solved, and therefore the transferred model is guaranteed to have high accuracy.
In a second aspect, referring to fig. 3, the present invention further provides an embodiment of a smart building fault diagnosis apparatus based on field adaptation, including:
the data preprocessing module 11 is used for carrying out data processing on sensor data in the intelligent building by combining expert knowledge in the building field to obtain preprocessed building data, and dividing the preprocessed building data into labeled source domain data and target domain data only with partial labels;
a pseudo tag assigning module 22, configured to assign a pseudo tag to the target domain data without a tag according to a trained source classifier and a preset data mining model, and consider the target domain data with the pseudo tag as the target domain data with the tag, where the trained source classifier is trained by using the source domain data with the tag;
a resistance transfer learning module 33, configured to perform a resistance-based transfer learning to obtain a trained target encoder and an updated source encoder, so as to transfer both the source domain data and the target domain data to a new common feature space;
an inter-class gap enhancement module 44, configured to perform parameter optimization on the trained target encoder and the updated source encoder by using a preset center loss optimizer, so as to enhance a gap between different classes in the common feature space;
The intelligent building fault diagnosis module 55 is configured to perform intelligent building fault diagnosis by using a trained shared classifier to obtain a diagnosis result of intelligent building fault, where the trained shared classifier is trained by using the source domain data and the labeled target domain data.
The data sources of the data set in the invention not only directly measure the data obtained by the fault device by the sensor, but also have some characteristics summarized by the expert, and the characteristics reflect the difference of certain indexes under normal and fault conditions, so that a more accurate diagnosis result is provided for the model.
The invention provides a pseudo tag giving algorithm under a data mining model and deep learning model agreement mechanism, which can ensure that enough non-tag data can be reasonably used and has higher pseudo tag accuracy.
The invention uses the field self-adaptive method in the migration learning, pays attention to the data distribution situation after migration, and can maintain the data distribution difference between different categories, thereby ensuring that the model after migration has high accuracy.
Compared with the prior art, the method has the advantages of simple and clear flow, simple model design for diagnosis, quick test and use, and convenient adjustment according to the requirement in the later period.
It will be clear to those skilled in the art that, for convenience and brevity of description, specific working procedures of the above-described systems, apparatuses and units may refer to corresponding procedures in the foregoing method embodiments, which are not repeated herein.
In the embodiments provided in the present application, it should be understood that the disclosed system, apparatus and method may be implemented in other manners. For example, the apparatus embodiments described above are merely illustrative, e.g., the division of the units is merely a logical function division, and there may be additional divisions when actually implemented, e.g., multiple units or components may be combined or integrated into another system, or some features may be omitted or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be an indirect coupling or communication connection via some interfaces, devices or units, which may be in electrical, mechanical or other form.
The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
In addition, each functional unit in the embodiments of the present invention may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit. The integrated units may be implemented in hardware or in software functional units.
The integrated units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention may be embodied essentially or in part or all of the technical solution or in part in the form of a software product stored in a storage medium, including instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
The above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. The intelligent building fault diagnosis method based on the field self-adaption is characterized by comprising the following steps of:
carrying out data preprocessing on sensor data in an intelligent building by combining expert knowledge in the building field to obtain preprocessed building data, and dividing the preprocessed building data into labeled source domain data and target domain data only with partial labels;
pseudo labels are given to the target domain data without labels according to a trained source classifier and a preset data mining model, the target domain data given with the pseudo labels are regarded as the target domain data with labels, and the trained source classifier is obtained by training the source domain data with labels;
Performing challenge-based migration learning to obtain a trained target encoder and an updated source encoder, so as to migrate both the source domain data and the target domain data to a new public feature space;
performing parameter optimization on the trained target encoder and the updated source encoder by using a preset center loss optimizer so as to enhance differences among different categories in the common feature space;
and performing intelligent building fault diagnosis by using the trained shared classifier to obtain an intelligent building fault diagnosis result, wherein the trained shared classifier is obtained by training the source domain data and the target domain data with the labels.
2. The intelligent building fault diagnosis method based on field adaptation according to claim 1, wherein the data preprocessing of the sensor data in the intelligent building in combination with expert knowledge of the building field to obtain preprocessed building data comprises:
performing data inspection on the sensor data and deleting missing values;
extracting features of the sensor data, performing dimension reduction processing on the extracted features by adopting an extremely random tree, and reserving sensitive features meeting preset conditions;
And combining expert knowledge in the building field to obtain expert features, and taking the expert features and the sensitive features as preprocessed building data.
3. The domain-adaptive-based intelligent building fault diagnosis method according to claim 1, wherein labeling the target domain data without labels according to a trained source classifier and a preset data mining model comprises:
updating the trained source classifier by using target domain data with labels to obtain an updated source classifier, and classifying data samples in the target domain data by using the updated source classifier to obtain first-class labels;
classifying the data samples by using the preset data mining model with the tag diffusion algorithm to obtain second class tags;
and when the first type tag and the second type tag are the same, taking the first type tag or the second type tag as a pseudo tag of the data sample.
4. A field-adaptive-based intelligent building fault diagnosis method according to claim 3, wherein updating the trained source classifier with tagged target domain data to obtain an updated source classifier comprises:
Initializing a target domain encoder by using parameters of the source domain encoder;
mapping the tagged target domain data using the target domain encoder;
and updating the trained source classifier by using the mapped target domain data with the labels to obtain an updated source classifier.
5. The field-adaptive-based intelligent building fault diagnosis method according to claim 1, wherein the loss function of the center loss optimizer is:
wherein ,is center loss, x i I data point representing each category, +.>Representing the center point of the category, batch_size represents the amount of data that participates in the calculation at the time of the current training.
6. Intelligent building fault diagnosis device based on field self-adaptation, its characterized in that includes:
the data preprocessing module is used for carrying out data processing on sensor data in the intelligent building by combining expert knowledge in the building field to obtain preprocessed building data, and dividing the preprocessed building data into labeled source domain data and target domain data only with partial labels;
the pseudo tag giving module is used for giving pseudo tags to the target domain data without tags according to a trained source classifier and a preset data mining model, and regarding the target domain data given with the pseudo tags as the target domain data with the tags, wherein the trained source classifier is obtained by training the source domain data with the tags;
The resistance transfer learning module is used for performing resistance-based transfer learning to obtain a trained target encoder and an updated source encoder so as to transfer the source domain data and the target domain data to a new public feature space;
the inter-class gap enhancement module is used for carrying out parameter optimization on the trained target encoder and the updated source encoder by using a preset center loss optimizer so as to enhance the gaps among different classes in the public feature space;
the intelligent building fault diagnosis module is used for carrying out intelligent building fault diagnosis by utilizing the trained shared classifier to obtain an intelligent building fault diagnosis result, and the trained shared classifier is obtained by training the source domain data and the target domain data with the labels.
7. The field-adaptive intelligent building fault diagnosis apparatus according to claim 6, wherein the data preprocessing module performs data preprocessing on sensor data in the intelligent building in combination with expert knowledge of the building field to obtain preprocessed building data, comprising:
the data preprocessing module performs data inspection on the sensor data and deletes missing values;
Extracting features of the sensor data, performing dimension reduction processing on the extracted features by adopting an extremely random tree, and reserving sensitive features meeting preset conditions;
and combining expert knowledge in the building field to obtain expert features, and taking the expert features and the sensitive features as preprocessed building data.
8. The domain-adaptive-based intelligent building fault diagnosis apparatus according to claim 6, wherein the pseudo tag assignment module pseudo tags the target domain data without tags according to a trained source classifier and a preset data mining model, comprising:
the pseudo tag giving module updates the trained source classifier by using target domain data with tags to obtain an updated source classifier, and classifies data samples in the target domain data by using the updated source classifier to obtain a first type of tag;
classifying the data samples by using the preset data mining model with the tag diffusion algorithm to obtain second class tags;
and when the first type tag and the second type tag are the same, taking the first type tag or the second type tag as a pseudo tag of the data sample.
9. The domain-based adaptive intelligent building fault diagnosis apparatus according to claim 8, wherein the pseudo tag assignment module updates the trained source classifier with tagged target domain data to obtain an updated source classifier comprises:
the pseudo tag giving module initializes the target domain encoder by using parameters of the source domain encoder;
mapping the tagged target domain data using the target domain encoder;
and updating the trained source classifier by using the mapped target domain data with the labels to obtain an updated source classifier.
10. The field-adaptive-based intelligent building fault diagnosis apparatus according to claim 6, wherein the loss function of the center loss optimizer in the inter-class gap enhancement module is:
wherein ,is center loss, x i I data point representing each category, +.>Representing the center point of the category, batch_size represents the number of data points in each category.
CN202310728138.2A 2023-06-19 2023-06-19 Intelligent building fault diagnosis method and device based on field self-adaption Pending CN116756634A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310728138.2A CN116756634A (en) 2023-06-19 2023-06-19 Intelligent building fault diagnosis method and device based on field self-adaption

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310728138.2A CN116756634A (en) 2023-06-19 2023-06-19 Intelligent building fault diagnosis method and device based on field self-adaption

Publications (1)

Publication Number Publication Date
CN116756634A true CN116756634A (en) 2023-09-15

Family

ID=87958490

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310728138.2A Pending CN116756634A (en) 2023-06-19 2023-06-19 Intelligent building fault diagnosis method and device based on field self-adaption

Country Status (1)

Country Link
CN (1) CN116756634A (en)

Similar Documents

Publication Publication Date Title
CN110046743B (en) Public building energy consumption prediction method and system based on GA-ANN
Thai-Nghe et al. Deep learning approach for forecasting water quality in IoT systems
CN112906982A (en) GNN-LSTM combination-based network flow prediction method
CN110751186B (en) Cross-project software defect prediction method based on supervised expression learning
CN111126576B (en) Deep learning training method
CN108961468B (en) Ship power system fault diagnosis method based on integrated learning
WO2023116111A1 (en) Disk fault prediction method and apparatus
CN110162478A (en) A kind of defect code path orientation method based on defect report
CN113591971B (en) User individual behavior prediction method based on DPI time sequence word embedded vector
CN113743538A (en) Intelligent building energy consumption prediction method, equipment and medium based on IPSO-BP neural network
CN113762591B (en) Short-term electric quantity prediction method and system based on GRU and multi-core SVM countermeasure learning
Gökçe et al. Performance comparison of simple regression, random forest and XGBoost algorithms for forecasting electricity demand
Haroon et al. Application of machine learning in forensic science
CN109034182A (en) A kind of zero sample image identification new method based on attribute constraint
CN113033898A (en) Electrical load prediction method and system based on K-means clustering and BI-LSTM neural network
CN116756634A (en) Intelligent building fault diagnosis method and device based on field self-adaption
WO2023082788A1 (en) Method and apparatus for predicting oxygen content in flue gas and load, method and apparatus for selecting prediction model, and method and apparatus for predicting flue gas emission
Kuang [Retracted] Prediction of Urban Scale Expansion Based on Genetic Algorithm Optimized Neural Network Model
CN114566230A (en) Method for predicting compound properties based on multitask deep learning model
CN113723660A (en) Specific behavior type prediction method and system based on DNN-LSTM fusion model
CN114625831A (en) Classification evaluation feedback method for load identification of smart power grid
CN112348275A (en) Regional ecological environment change prediction method based on online incremental learning
CN109961143A (en) The point-by-point estimated amount of damage optimization method of RNN model
Wu et al. Association rule mining with a special rule coding and dynamic genetic algorithm for air quality impact factors in Beijing, China
KR102636461B1 (en) Automated labeling method, device, and system for learning artificial intelligence models

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination