CN117951613A

CN117951613A - Online learning anomaly detection method, device, equipment and medium based on self-adaptive weighting

Info

Publication number: CN117951613A
Application number: CN202410162430.7A
Authority: CN
Inventors: 陶艳君; 应文豪
Original assignee: Suzhou Beijie Environmental Protection Equipment Co ltd; Changshu Institute of Technology
Current assignee: Suzhou Beijie Environmental Protection Equipment Co ltd; Changshu Institute of Technology
Priority date: 2024-02-05
Filing date: 2024-02-05
Publication date: 2024-04-30

Abstract

The disclosure provides an online learning anomaly detection method, device, equipment and medium based on self-adaptive weighting, which comprises the steps of obtaining first data, wherein the first data comprises data streams generated by sensors, and the sensors comprise an air pressure sensor, a water pressure sensor, a temperature sensor and an air speed sensor; extracting time sequence features in the data stream; and updating the data flow through machine learning, and calculating the gradient of the loss function relative to the model parameters to realize the online adjustment of the model parameters. The invention realizes real-time, accuracy and self-adaptability of anomaly detection and provides powerful support for safety and environmental protection of industrial production.

Description

Online learning anomaly detection method, device, equipment and medium based on self-adaptive weighting

Technical Field

The invention relates to the technical field of anomaly detection, in particular to an online learning anomaly detection method, device, equipment and medium based on self-adaptive weighting.

Background

With the improvement of the industrial automation level, the dust collecting system plays a key environmental protection role in industrial production. These systems maintain the cleanliness of the production environment and the health of personnel by collecting, filtering and disposing of dust and particulates generated during the production process. However, under the influence of long-term operation and complex industrial environments, dust collection systems may face various abnormal conditions, such as equipment failure, pipeline blockage, fan abnormality, etc., which may cause system performance degradation, affect production efficiency, and even cause potential safety hazards.

There are some limitations to the conventional dust collection system monitoring methods. Traditional offline monitoring methods are generally used for monitoring by manually setting a threshold value or based on a static model, and are difficult to adapt to dynamic changes of industrial production environments. Because the factors such as dust concentration, environmental conditions and the like change at any time in the production process, the methods can cause false alarm or missing alarm, and the requirements of real-time monitoring and self-adaption cannot be met. Meanwhile, there are various and dynamic anomalies in industrial production, such as equipment failure, operational anomalies, etc., which may affect the proper operation of the dust collection system. Therefore, improving real-time monitoring and accurate identification of dust collection system anomalies is an urgent need in industrial processes.

In the field of machine learning, the traditional processing learning method generally needs offline training and cannot adapt to the dynamic change characteristics of an industrial dust collection system. In order to solve this problem, researchers have made remarkable progress in the field of online learning in recent years. Online learning allows the model to continuously adapt to new data, maintaining flexible adaptation to data distribution by dynamically updating model parameters. This makes online learning an ideal choice for anomaly detection in industrial dust collection systems. However, the application of the existing online learning method in the abnormality detection of the industrial dust collecting system also faces some challenges. Dust collection system data often have high dynamic, non-stationarity and isomerism, and an anomaly detection method is required to have good adaptability and robustness. Furthermore, the anomaly detection method also needs to be able to handle unbalanced data distribution, ensuring sensitivity to few classes of anomalies.

Disclosure of Invention

The invention aims to provide an online learning anomaly detection method, device, equipment and medium based on self-adaptive weighting, which can be better adapted to continuously-changing industrial environments, realize the instantaneity, the accuracy and the self-adaptation of anomaly detection and provide powerful support for the safety and the environmental protection of industrial production.

According to an aspect of the present disclosure, there is provided an online learning anomaly detection method based on adaptive weighting, including: acquiring first data, wherein the first data comprises data streams generated by sensors, and the sensors comprise an air pressure sensor, a water pressure sensor, a temperature sensor and a wind speed sensor; extracting time sequence features in the data stream; and updating the data flow through machine learning, and calculating the gradient of the loss function relative to the model parameters to realize the online adjustment of the model parameters.

According to some embodiments of the present disclosure, the extracting the time-series features in the data stream by combining the structured data generated by the sensor device specifically includes: the pressure change trend features extracted by the air pressure sensor are used for capturing the air flow state; the water pressure fluctuation characteristic extracted by the water pressure sensor is used for monitoring abnormal liquid flow behavior; the temperature change characteristics extracted by the temperature sensor are used for detecting abnormal fluctuation of the working temperature of the system; the wind speed change characteristics extracted by the wind speed sensor are used for analyzing the influence of wind power on the dust collection system.

According to some embodiments of the disclosure, the extracting the time series features in the data stream further includes normalizing and filtering the data.

According to some embodiments of the disclosure, the updating the data stream through machine learning, and calculating a gradient of a loss function with respect to a model parameter, to implement an online adjustment process of the model parameter, specifically further includes: and (3) detecting abnormality in the online adjustment process of the model parameters, generating a corresponding alarm signal when abnormality is detected, and providing abnormality detection information for operators in real time.

According to some embodiments of the disclosure, the weights of the positive and negative categories are weighted by adopting a weighting strategy which is self-adaptive according to the distribution design inside the time sequence batch samples, wherein the weighting strategy adopts the ratio of the number of normal samples and abnormal samples to the total number of samples in the small batch as the weight values of the normal samples and the abnormal samples respectively.

According to a second aspect of the present disclosure, there is provided an online learning anomaly detection device based on adaptive weighting, including: the data acquisition module is used for acquiring first data, wherein the first data comprises data streams generated by sensors, and the sensors comprise a gas pressure sensor, a water pressure sensor, a temperature sensor and a wind speed sensor; the feature extraction module is used for extracting time sequence features in the data stream; and the online learning module is used for updating the data stream through machine learning, calculating the gradient of the loss function on the model parameters and realizing online adjustment of the model parameters.

According to a third aspect of the present disclosure, there is provided an electronic device comprising a processor, and a memory and a network interface connected to the processor; the network interface is connected with a nonvolatile memory in the server; the processor retrieves the computer program from the nonvolatile memory through the network interface during operation, and runs the computer program through the memory to execute the online learning anomaly detection method based on adaptive weighting.

According to a fourth aspect of the present disclosure, there is provided a storage medium having stored thereon computer program instructions which, when executed by a processor, implement an adaptive weighting based online learning anomaly detection method as described above.

According to the technical scheme, the method, the device, the equipment and the medium for detecting the abnormality in the online learning based on the self-adaptive weighting are provided, and the innovative application of the multi-dimensional time sequence feature extraction, ELASTIC WEIGHT Consolidation algorithm, the weighted loss function and the weighted cross entropy loss function is combined.

Drawings

FIG. 1 shows a schematic diagram of an adaptively weighted online learning anomaly detection method according to the present disclosure;

FIG. 2 shows a schematic diagram of EWC operation according to the present disclosure;

FIG. 3 shows a schematic of an equipment anomaly flow according to the present disclosure;

FIG. 4 shows a schematic diagram of a computing device according to an embodiment of the present disclosure.

Detailed Description

The technical solutions of the embodiments of the present invention will now be clearly and completely described with reference to the drawings in the embodiments of the present invention, and it is apparent that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

It is noted that relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.

The disclosure provides an online learning anomaly detection method based on adaptive weighting, which comprises the following steps of (1): acquiring first data, wherein the first data comprises data streams generated by sensors, and the sensors comprise an air pressure sensor, a water pressure sensor, a temperature sensor and a wind speed sensor; s2: extracting time sequence features in the data stream; s3: and updating the data flow through machine learning, and calculating the gradient of the loss function relative to the model parameters to realize the online adjustment of the model parameters.

The method comprises the following key modules:

(1) A data stream receiving module: after the explosion-proof dry-wet dust collector is started, the system enters an automatic flow. The data flow receiving module receives data flow from a dust collecting system sensor in real time, and the data flow comprises an air pressure sensor, a water pressure sensor, a temperature sensor, a wind speed sensor and the like.

(2) And the feature extraction module is used for: the module extracts key information from raw data collected by each sensor. By analyzing the spectral characteristics, amplitude variations, etc., various time series features are extracted. The importance of this step in machine learning is not negligible, which helps reduce data dimensionality, remove noise, and improve model performance.

(3) And an online learning module: the ELASTIC WEIGHT Consolidation (EWC) machine learning algorithm is adopted as a core algorithm of online learning, and the module dynamically updates an anomaly detection model so as to adapt to new data flow and improve detection performance. In addition, the module comprises parameter updating rules, and online adjustment of model parameters is achieved by calculating a loss function gradient, so that the model can keep robustness.

(4) An abnormality detection module: and performing real-time anomaly detection by using the model dynamically updated by the online learning module. And introducing a weighted loss function to balance the contributions of the normal sample and the abnormal sample, so as to ensure that the abnormal detection can be effectively carried out under different data distribution. In addition, the weighted cross entropy loss function is adopted, so that the abnormality detection performance in an unbalanced data scene is improved.

(5) An alarm generation module: when the abnormality detection module detects an abnormality, the alarm generation module generates a corresponding alarm signal. This ensures that when an abnormality occurs in the dust collection system, an operator can obtain feedback in time, and take necessary measures to improve the safety and reliability of the system.

In the invention, the characteristic extraction module analyzes the multidimensional time series characteristics of the sensor data. The method specifically comprises the following steps: the pressure change trend characteristics extracted from the air pressure sensor are used for capturing the change condition of the air flow state; the water pressure fluctuation characteristics extracted from the water pressure sensor are used for monitoring abnormal behaviors of liquid flow; detecting abnormal fluctuation of the working temperature of the system by the temperature change characteristics extracted from the temperature sensor; analyzing the influence of wind power on a dust collection system according to the wind speed change characteristics extracted from the wind speed sensor; the feature extraction module further comprises preprocessing steps such as data normalization and filtering so as to ensure the accuracy and stability of the sensor data.

Further, for the online learning module, the present invention employs a machine learning algorithm ELASTIC WEIGHT Consolidation (EWC), which is a consideration for several reasons. First, EWC can preserve knowledge of previous tasks, in which the model needs to continuously adapt to new data streams, but at the same time, also needs to preserve knowledge of previous tasks, i.e., historical data. Traditional online learning algorithms may suffer from a "forget" problem in that previously learned knowledge is easily forgotten when learning a new task. The EWC algorithm better preserves knowledge of previous tasks by applying a penalty to the model parameters that makes it less likely to modify weights important to previous tasks. Moreover, the EWC has strong adaptability, and the data stream in the industrial environment generally shows high dynamic property, non-stationarity and isomerism. The EWC algorithm has strong adaptability, and can keep the performance of the model under the continuously changing data distribution. By dynamically updating the model parameters, the EWC algorithm can adapt to new data streams in time, and robustness of the model can be ensured when the model faces to dynamic changes of an industrial dust collection system. Then, the EWC can prevent catastrophic forgetfulness, which is a significant adverse effect on the performance of the previous task when a new task is learned in online learning. The EWC algorithm effectively prevents catastrophic forgetting through punishment of key parameters, and ensures that the model does not lose important knowledge learned before when adapting to new data. The following is that the industrial dust collection system's operating conditions may experience conceptual drift over time, i.e., changes in data distribution, in response to the conceptual drift. The EWC algorithm can better adapt to conceptual drift through dynamic adjustment of parameters, and improves generalization of the model. Finally, the EWC algorithm avoids the complex process of retraining the whole model through online parameter adjustment. This is particularly important for industrial online learning scenarios, as it allows models to be learned and adjusted online without interrupting operation, ensuring continuous monitoring and anomaly detection of the dust collection system. Based on its significant improvement in model parameter stability in online learning scenarios and its ability to retain knowledge of previous tasks. The robustness of the model is improved by adapting to new data streams with the knowledge of previous tasks preserved. The module also comprises a parameter updating rule, and the on-line adjustment of the model parameters is realized by calculating the gradient of the loss function on the model parameters so as to ensure that the model can adapt to the changed industrial environment in time.

Further, the invention designs a new loss function, namely a weighted loss function, in the anomaly detection so as to balance the contributions of the normal sample and the anomaly sample. A weighted cross entropy loss function is introduced to improve anomaly detection performance in an unbalanced data scenario. The weight item of the loss function can be adaptively adjusted according to the distribution in the time sequence batch samples, so that few categories are more focused in the training process, and the abnormality detection performance is improved.

For a better understanding of the present invention, please refer to the specific implementation of the system module in conjunction with fig. 1:

(1) In the embodiment, in dust treatment equipment, such as a chuck type dust collector, a cloth belt type dust collector or a cyclone type dust collector, a plurality of sensor devices are arranged inside the dust collector, and data in the operation process of the dust collector is monitored and collected in real time. The data flow receiving module is connected with the dust collecting system sensor to acquire the data flow generated by the sensor in real time. The sensor comprises an air pressure sensor, a water pressure sensor, a temperature sensor, a wind speed sensor and the like. The data stream is transmitted in a time sequence form and comprises data of a plurality of observation time points;

(2) The feature extraction module performs in an embodiment the extraction of multi-dimensional time series features. By analyzing the spectral characteristics, amplitude changes and the like of sensor data such as air pressure, water pressure, temperature, wind speed and the like, the characteristics which have important significance for an abnormality detection task are extracted;

(3) In an embodiment, the online learning module employs ELASTIC WEIGHT Consolidation (EWC) algorithm to adapt to the new data stream by dynamically adjusting the model parameters. In the initialization stage, the module sets initial model parameters, and receives new data at each time point for online learning. The EWC algorithm reserves the knowledge of the previous task by calculating the gradient of the loss function, so that the robustness of the model is improved;

(4) The anomaly detection module uses the model dynamically updated by the online learning module to detect in real time. A weighted loss function and a weighted cross entropy loss function are introduced to balance the contributions of normal and abnormal samples and to improve performance in unbalanced data scenarios. The module judges whether abnormality occurs or not through threshold setting of model output and generates corresponding alarm signals;

(5) The alarm generation module generates a corresponding alarm signal when an abnormality is detected, and provides real-time feedback to an operator. Through modes such as an interface and alarm information, operators can timely take necessary measures, and the safety and reliability of the system are improved.

The invention introduces ELASTIC WEIGHT Consolidation (EWC) algorithm in the dust collecting system abnormality detection method of online self-adaptive learning, which aims to solve the forgetting problem in continuous learning task. According to the method, the EWC regularization term is introduced into the loss function, so that key model parameters of a previous task are protected, and the model is prevented from forgetting knowledge of the previous task when learning a new task. The regular term of EWC can be substituted into the loss function in the present invention, and the specific expression is as follows:

Where λ is the regularized weight used to control the degree of protection of the previous task, θ _i is a parameter of the model, Is the optimal value of the parameter on the previous task. Fisher _i is a Fisher information matrix, reflects the importance of parameters, measures the sensitivity of model output relative to the parameters, and is calculated as follows:

Wherein E [ ] represents the expectation of the expression in brackets; specifically, this equation pertains to an element in the Fisher information matrix that represents the desired square of the log-likelihood function with respect to the second partial derivative of the model parameter θ _i; this is typically used to evaluate the uncertainty and information content of the estimate in statistical inference.

Where p (D; θ) is a likelihood function of the model with respect to the parameter θ, D is data; e [ ] represents the expectation of the expression in brackets, specifically, this equation is about an element in the Fisher information matrix that represents the square of the expectation of the log-likelihood function about the second partial derivative of the model parameter θ _i, which is commonly used to evaluate the uncertainty and information quantity of the estimate in statistical inference.

The main reason for using the EWC algorithm is that it can prevent catastrophic forgetfulness when learning a new task. For dust collection system anomaly detection, the knowledge of the previous task includes features of the system's normal operation, which are of significant value for anomaly detection. The EWC protects key parameters of the previous task by introducing regularization terms into the loss function, so that the model can be better adapted to the old task when learning the new task. In the field of dust collection system anomaly detection, there are other continuous learning algorithms, including Replay algorithms, which prevent forgetting by periodically reviewing the data of previous tasks and re-injecting them into the training data. However, replay may face storage and computing resource challenges. There is also LwF (Learning without Forgetting) algorithm that attempts to alleviate the forgetting problem by preserving the output distribution of the previous task. However, in high-dimensional data and complex tasks, the effect of LwF may be limited.

EWC is introduced in the dust collecting system anomaly detection method for online self-adaptive learning, and the problem that prior task knowledge is forgotten in continuous learning is solved through a unique regularization mechanism of the EWC. The robustness and the practicability of the system are improved. Compared with other algorithms, the EWC is excellent in preventing forgetting, and an effective solution is provided for abnormality detection in an online learning scene.

The loss function is a key component of model optimization that quantifies the difference between the predicted outcome of the model and the actual target. In the dust collecting system abnormality detection method of the present invention, the purpose of the loss function is to measure the accuracy of prediction of the model for the system state, especially the performance in detecting abnormalities. In the present invention, the penalty function may include new task terms and ELASTIC WEIGHT Consolidation (EWC) regularized terms. The specific expression is as follows:

TotalLoss＝NewTaskLoss+EWCRegularizer (3)

The Total Loss is a concept of comprehensively considering new task Loss and regularization items, and balancing the adaptability to the new task and the reservation of old task knowledge when learning the new task; new Task Loss represents a penalty term for the New Task, and EWC Regularizer represents an EWC regularization term. The new task loss term is the error that the model produces when learning new data. It typically includes the differences between model predictions and actual observations. In the present invention, this term reflects the predictive performance of the model on the dust collection system state, especially when faced with new data. The regularization term in EWC is used to prevent model overshoot as new tasks are learned to preserve knowledge of previous tasks. The expression comprises a weight term, a Fisher information matrix and the square of parameter variation. This regularization enables the model to adapt to new data as it learns new tasks, while preserving the adaptability to previous tasks. The total loss function becomes the optimization target of the model. During the training process, the parameters of the model are adjusted to minimize the loss function, thereby improving the accuracy and robustness of the model to anomaly detection.

In the present invention, the abnormality detection task of the dust collection system has a certain specificity in which the data may exhibit an unbalanced distribution. Unbalanced data refers to the situation where the number of samples of one class is much smaller than that of another class. In dust collection systems, samples of normal operating conditions may occupy a substantial majority, while anomalies are relatively rare. Such imbalance may result in models that tend to predict a greater number of categories while learning and predicting, and perform poorly for fewer categories. In the process of anomaly detection, the invention adopts a Weighted Cross entropy (Weighted Cross-Entropy) loss function in order to solve the problem of unbalanced data classification. The loss function gives different weights to different classes of samples when calculating the loss in order to better handle unbalanced data. The basic expression of the weighted cross entropy is as follows:

L_{weight_cross_entropy}(y_pred,y_ture)-W_posyturelog(y_pred)-W_neg(1-y_ture)log(1-y_pred) (4)

Where y _pred is the predictive probability of the model, y _true is the true label, and W _pos and W _neg are the weights of the positive and negative categories. The weight terms W _pos and W _neg function to balance the contributions of normal and abnormal samples in the loss calculation. By adjusting these two weights, the degree of interest of the model in normal and abnormal samples can be controlled more flexibly.

In view of this, the present invention transitions to the problem of unbalanced data for anomaly detection tasks and emphasizes the importance of considering unbalanced data classification. In practical application, it is particularly important to timely and accurately detect the abnormality of the dust collecting system, so that it is important to fully consider the problem of unbalanced data in the model design.

In order to further optimize the adaptability of the model to unbalanced data, the invention improves the weighted cross entropy. This improvement takes into account the difference in the number of normal and abnormal samples within a batch, dynamically builds a weight term as follows:

here, N _normal and N _anomaly represent the number of data stream batches at this time of normal and abnormal samples, respectively, and N _total is the total number of the data stream batches. By hooking the construction of the weight term with the number of batches, we make the model more flexible to adapt to the sample distribution in different batches, especially when it comes to a dynamically changing dataset.

The improved weighted cross entropy loss function is beneficial to improving the performance of the model on an abnormality detection task in practical application, and particularly provides a more reliable solution for abnormality detection of a dust collection system in terms of coping with unbalanced data distribution.

In an embodiment of the present invention, the specification of the patent details an alarm generation module, as shown in fig. 3, which plays a key role in dust collection system abnormality detection. The alarm generation module is an important component of man-machine interaction, and the practicability and the user friendliness of the system are improved through feeding back abnormal detection results in real time.

The alert generation module receives the output of the anomaly detection model and determines whether to trigger an alert based on predefined thresholds or rules. When an abnormal event occurs, the module can quickly generate a corresponding alarm signal to give a timely warning to an operator. This real-time feedback is critical to taking the necessary measures in time in the dust collection system to prevent potential equipment failure or system anomalies.

In implementations, the alert generation module may take a variety of forms including, but not limited to, a graphical interface, a warning light, an audible cue, and the like. Through the graphical interface, operators can clearly know the result of the abnormality detection and take corresponding actions when needed. Meanwhile, through modes such as an alarm lamp and an audio prompt, the module can effectively convey abnormal information when the working environment is noisy or an operator cannot directly see the display screen.

The demonstration experiments of the method of the patent use the actual dust collection system working data to verify the effectiveness of the dust collection system in practical application. We have selected dust collection system operating data over a range of times, including normal operating conditions and various anomalies. Experiments aim to evaluate the practicability of the method in the real working environment through the abnormal detection performance of the model. The experimental data comprises real-time monitoring data from a plurality of sensors such as an air pressure sensor, a water pressure sensor, a temperature sensor, a wind speed sensor and the like. The injection of abnormal conditions simulates various problems that dust collection systems may face, such as pipe blockage, fan failure, etc. Such a comprehensive data set can comprehensively cover various abnormal scenes that may be encountered by the dust collection system. In the experiment, the patent classifies the working state of the dust collecting system into two types of normal and abnormal, and uses the method of the patent for abnormality detection. This patent carries out supervised learning according to the label of experimental data, trains the model in order to adapt to the difference of normal and abnormal state. In order to evaluate the performance of the model, the patent uses indexes such as accuracy, recall rate, F1 value and the like to perform comprehensive performance analysis.

Comparison of results from different methods

Method of	Precison	Recall	F1
				Replay	80.2	78.48	79.33
LwF	88.45	89.03	88.71
				EWC	93.84	96.75	95.27
The method of the invention	97.67	98.01	97.84

In the above table, we compare the performance of different methods (Replay, lwF, EWC and the method of the present invention) in the dust collection system anomaly detection task, and comprehensively evaluate their performance by means of three indexes Precision, recall and F1. The Replay method performs well in terms of accuracy, recall and F1 values, but relatively low overall performance. The LwF method has a significant improvement in all indexes compared with the Replay method, and is more excellent in performance, but has a gap compared with the two subsequent methods. The EWC method reaches a high level in all indexes, showing superiority in the dust collecting system abnormality detection task. The method of the invention is excellent in all indexes and reaches the highest level. The high accuracy and recall, and the high F1 value indicate that the method achieves significant performance improvement in dust collection system anomaly detection. Compared with other methods, the method provided by the invention has higher performance level on all indexes, and has better accuracy and recall rate. The EWC method has good effect in the experiment, but the method has more excellent performance, and is suitable for the actual application scene of the abnormality detection of the dust collecting system.

Embodiments of the present invention also provide a computer readable storage medium having computer program instructions stored thereon; when the computer program instructions are executed by the processor, the training method of the machine learning algorithm EWC provided by the embodiment of the invention is realized.

It should be understood that the invention is not limited to the particular arrangements and instrumentality described above and shown in the drawings. For the sake of brevity, a detailed description of known methods is omitted here. In the above embodiments, several specific steps are described and shown as examples. The method processes of the present invention are not limited to the specific steps described and shown, but various changes, modifications and additions, or the order between steps may be made by those skilled in the art after appreciating the spirit of the present invention.

The functional blocks shown in the above-described structural block diagrams may be implemented in hardware, software, firmware, or a combination thereof. When implemented in hardware, it may be, for example, an electronic Circuit, application SPECIFIC INTEGRATED Circuit (ASIC), appropriate firmware, plug-in, function card, or the like. When implemented in software, the elements of the invention are the programs or code segments used to perform the required tasks. The program or code segments may be stored in a machine readable medium or transmitted over transmission media or communication links by a data signal carried in a carrier wave. A "machine-readable medium" may include any medium that can store or transfer information. Examples of machine-readable media include electronic circuitry, semiconductor Memory devices, read-Only Memory (ROM), flash Memory, removable Read-Only Memory (Erasable Read Only Memory, EROM), floppy disks, compact discs (Compact Disc Read-Only Memory, CD-ROM), optical discs, hard disks, fiber optic media, radio Frequency (RF) links, and so forth. The code segments may be downloaded via computer networks such as the internet, intranets, etc.

It should also be noted that the exemplary embodiments mentioned in this disclosure describe some methods or systems based on a series of steps or devices. The present invention is not limited to the order of the above-described steps, that is, the steps may be performed in the order mentioned in the embodiments, or may be performed in a different order from the order in the embodiments, or several steps may be performed simultaneously.

Aspects of the present disclosure are described above with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions/acts specified in the flowchart and/or block diagram block or blocks. Such a processor may be, but is not limited to being, a general purpose processor, a special purpose processor, an application specific processor, or a field programmable logic circuit. It will also be understood that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware which performs the specified functions or acts, or combinations of special purpose hardware and computer instructions.

In the foregoing, only the specific embodiments of the present invention are described, and it will be clearly understood by those skilled in the art that, for convenience and brevity of description, the specific working processes of the systems, modules and units described above may refer to the corresponding processes in the foregoing method embodiments, which are not repeated herein. It should be understood that the scope of the present invention is not limited thereto, and any equivalent modifications or substitutions can be easily made by those skilled in the art within the technical scope of the present invention, and they should be included in the scope of the present invention.

Claims

1. An online learning anomaly detection method based on adaptive weighting is characterized by comprising the following steps:

Acquiring first data, wherein the first data comprises data streams generated by sensors, and the sensors comprise an air pressure sensor, a water pressure sensor, a temperature sensor and a wind speed sensor;

Extracting time sequence features in the data stream;

and updating the data flow through machine learning, and calculating the gradient of the loss function relative to the model parameters to realize the online adjustment of the model parameters.

2. The adaptive weighting-based online learning anomaly detection method of claim 1, wherein the extracting time-series features in the data stream extracts a plurality of time-series features by analyzing a first characteristic, the first characteristic comprising a spectral characteristic, a magnitude variation.

3. The method for online learning anomaly detection based on adaptive weighting of claim 2, wherein the extracting of time series features in the data stream is performed by combining structured data generated by a sensor device, specifically comprising:

the pressure change trend features extracted by the air pressure sensor are used for capturing the air flow state;

the water pressure fluctuation characteristic extracted by the water pressure sensor is used for monitoring abnormal liquid flow behavior;

The temperature change characteristics extracted by the temperature sensor are used for detecting abnormal fluctuation of the working temperature of the system;

The wind speed change characteristics extracted by the wind speed sensor are used for analyzing the influence of wind power on the dust collection system.

4. An adaptive weighting based online learning anomaly detection method according to claim 3 wherein the extracting of time series features in the data stream further comprises normalization and filtering of the data.

5. The online learning anomaly detection method based on adaptive weighting according to claim 3, wherein the updating of the data stream by machine learning and the calculation of the gradient of the loss function with respect to the model parameters realizes the online adjustment process of the model parameters, and specifically further comprises:

And (3) detecting abnormality in the online adjustment process of the model parameters, generating a corresponding alarm signal when abnormality is detected, and providing abnormality detection information for operators in real time.

6. The adaptive weighting-based online learning anomaly detection method of claim 5, wherein the anomaly detection employs a weighted cross entropy loss function that is:

L_{weight_cross_entrypy}(y_pred,y_true)＝-W_posy_truelog(y_pred)-W_neg(1-y_true)log(1-y_pred)

Where y _pred is the predictive probability of the model and y _true is the true label; w _pos and W _neg are the weights of the positive and negative categories, respectively, L _{weight_cross_entropy}(y_pred,y_ture) are weighted cross entropy loss functions.

7. The method for online learning anomaly detection based on adaptive weighting according to claim 6, wherein the weights of the positive and negative categories are weighting strategies designed adaptively according to the distribution inside the time series batch samples, and the weighting strategies adopt the ratio of the number of normal samples, abnormal samples and the total number of samples in the small batch as the weight values of the normal samples and the abnormal samples, respectively.

8. An online learning anomaly detection device based on adaptive weighting, comprising:

The data acquisition module is used for acquiring first data, wherein the first data comprises data streams generated by sensors, and the sensors comprise a gas pressure sensor, a water pressure sensor, a temperature sensor and a wind speed sensor;

the feature extraction module is used for extracting time sequence features in the data stream;

and the online learning module is used for updating the data stream through machine learning, calculating the gradient of the loss function on the model parameters and realizing online adjustment of the model parameters.

9. An electronic device, comprising a processor, and a memory and a network interface connected with the processor; the network interface is connected with a nonvolatile memory in the server; the processor, when running, retrieving a computer program from the non-volatile memory via the network interface and running the computer program via the memory to perform the method of any of the preceding claims 1-7.

10. A storage medium, characterized by: computer program instructions stored on a computer storage medium, which when executed by a processor, implement the method of any one of claims 1-7.