An Industrial Internet of Things Equipment Fault Diagnosis System with Adaptive Triggering Incremental Learning

Abstract

The present invention discloses an industrial Internet of Things equipment fault diagnosis system with adaptively triggered incremental learning, wherein a fault diagnosis module monitors equipment operation data through sensors, determines whether the equipment is in a normal state based on the data, and if a fault occurs, determines what type of fault the equipment has. A concept drift detection module divides the equipment operation data by time window, uses historical data and current data to detect whether concept drift occurs in the data, and if concept drift is detected, activates an incremental update module to incrementally update the fault diagnosis module. The incremental update module uses representative sample data of old categories and new sample data to incrementally update the fault diagnosis module. In response to the situation where data changes constantly during the operation of industrial equipment, the present invention ensures that the fault diagnosis model can be updated in a timely manner, and does not require retraining of the fault diagnosis model, thereby maintaining a relatively fast update speed and a high fault diagnosis accuracy.

Description

Technical Field

[0001] The present invention relates to an industrial Internet of Things equipment fault diagnosis system with adaptively triggered incremental learning, belonging to the field of equipment fault diagnosis. Background Art

[0002] In recent years, to meet production demands, mechanical equipment has often operated in complex and harsh environments with high loads, high temperatures, and high operating rates. This can lead to a certain degree of damage, resulting in a series of adverse consequences. The stable and safe operation of mechanical equipment has become increasingly important in the industrial production industry. Therefore, fault diagnosis of mechanical equipment can avoid unnecessary maintenance costs and reduce the possibility of catastrophic accidents. With the development and widespread application of Industrial Internet of Things technology, the industrial field has accumulated a massive amount of equipment operation data through dense sensor networks, making data-driven fault diagnosis systems the mainstream.

[0003] Traditional data-driven fault diagnosis systems can only demonstrate excellent diagnostic accuracy and robustness in static data scenarios. Therefore, such systems typically treat equipment operating data as static data. However, in complex industrial production environments, equipment generates dynamic streaming data under time-varying operating conditions. Compared to static data, streaming data is large in scale and dynamic. Traditional data-driven fault diagnosis systems lack incremental learning and real-time update capabilities, making them difficult to adapt to industrial streaming data scenarios. Summary of the Invention

[0004] In order to solve the existing technical problems, the present invention provides an industrial Internet of Things equipment fault diagnosis system based on adaptive triggered incremental learning, which can solve the problem that the fault diagnosis model in the current fault diagnosis system cannot be updated incrementally or in real time. It aims to effectively diagnose the faults of mechanical equipment in industrial flow data, avoid damage and shutdown of mechanical equipment, and malicious operation of the system, improve the stability of automated production, ensure the long-term operation of the entire industrial system, and promote the further development of industrial manufacturing.

[0005] The technical solutions mainly adopted in the present invention are:

[0006] An industrial Internet of Things equipment fault diagnosis system based on adaptive triggered incremental learning includes a fault diagnosis module, a concept drift detection module and an incremental update module; wherein,

[0007] The fault diagnosis module monitors the operating data of industrial equipment through industrial Internet of Things sensors and uses the fault diagnosis model to determine whether the equipment is in a normal state. If it is in a faulty state, it determines what kind of fault the equipment has;

[0008] The concept drift detection module divides the equipment operation data by time window, uses historical data and current data to detect whether concept drift occurs in the data, and if concept drift is detected, activates the incremental update module to incrementally update the fault diagnosis module;

[0009] The incremental update module uses new data monitored by sensors to incrementally update the fault diagnosis module.

[0010] Preferably, the fault diagnosis module uses a one-dimensional convolutional neural network to construct a fault diagnosis model. To construct a fault diagnosis model based on a one-dimensional convolutional neural network, it is necessary to prepare a certain amount of operation monitoring data of the health status and fault status of the equipment in advance, and label the data with corresponding labels. Then, the prepared data Adam optimizer is used to train the fault diagnosis model. After the model training is completed, it can be used to diagnose equipment faults.

[0011] The concept drift detection module uses an OS-ELM-based concept drift detection algorithm to detect concept drift of data. The specific steps are as follows:

[0012] S3-1: Construct a time window and use it to divide the data. Input the data from the current time window and the historical time window into the algorithm.

[0013] S3-2: Initialize the online extreme learning machine model using historical data in the historical time window and update the OS-ELM using current data in the current time window;

[0014] S3-3: Detect whether concept drift occurs in the data by comparing the differences between the model hidden layer parameters obtained by historical and current data training;

[0015] S3-4: If concept drift is detected, the incremental update process of the fault diagnosis model is started.

[0016] Preferably, the OS-ELM-based concept drift detection algorithm is responsible for dividing the device operation data and detecting whether concept drift occurs in the data. The specific calculation process of the algorithm is as follows:

[0017] S4-1: Set two time windows TW HD With TW CD , among which TW HD Represents the historical time window, TW CD Represents the current time window, the length of the time window is L T . Assume that the initial time is TW HD The historical data in is defined as Indicates the interval L starting from time t=0 T The length of time, that is, [0,LT ] The data collected during the initial time TW CD The current data in is defined as Indicates that from time t = L T Start interval L T The length of time, that is, [L T ,2L T ] data collected within the time period.

[0018] S4-2: If no concept drift is detected, the right boundary of the historical time window is moved forward by the length of the time window, that is, the time interval is [0,2L T ], current time window TW CD The overall length of the forward time window is [2L T ,3L T If concept drift is detected, the historical time window TW is updated. HD , TW HD Move to Taiwan CD The moment of the previous time window length.

[0019] S4-3: Set the historical time window TW HD The data S1 in is input into the OS-ELM model. For a single hidden layer OS-ELM with L hidden units, the network model can be expressed as:

[0020]

[0021] Among them, g(·) is the activation function, W=[w1,…,w L ] T and B=[b1,…,b L ] T are the weight and bias values ​​between the input layer and the hidden layer, β i is the weight value from the hidden layer to the output layer. This formula can be transformed into the existence of β i , w i and b i , making

[0022]

[0023] The above formula can be expressed in matrix form, namely:

[0024] Hβ=T (4);

[0025] Where H is the hidden layer output value matrix, and β is the weight value matrix from the hidden layer to the output layer.

[0026] Using S1 data to train the OS-ELM model can obtain the model M(t0) at time t0, and the model parameters are This model is called the baseline model and represents the data distribution of the old data.

[0027] S4-4: Set the current data time window TW CD The data S2 in the input is input into the OS-ELM model for model update. Since OS-ELM only modifies the model output weight β, we can obtain t1 = t0 + L T Model parameters at time The model trained at this time is called M(t1), which represents the data distribution of the new data.

[0028] S4-5: By calculating the model parameters and The Euclidean distance between them can measure the difference between model parameters trained using data at different times, thereby reflecting whether concept drift occurs in streaming data. The formula for calculating the difference in model parameters is:

[0029]

[0030] S4-6: Set a detection threshold Th so that when the difference between the model before and after updating exceeds the threshold, that is, D(M(t0), M(t1))>Th, it can be determined that concept drift has occurred in the streaming data.

[0031] Preferably, the incremental update module uses an incremental learning algorithm based on forgetting constraints to incrementally update the fault diagnosis module. The specific update process is as follows:

[0032] S5-1: Construct a representative sample dataset D of old categories RE ,This dataset saves a constant number of representative samples for each category;

[0033] S5-2: Save data S of N time windows N For incremental updates, the new sample needs to carry a label;

[0034] S5-3: The old category representative sample dataset D RE The labeled samples of each old category in the data are input into the model together with the newly arrived samples in the streaming data for incremental update.

[0035] Preferably, the incremental learning algorithm based on forgetting constraint is responsible for learning knowledge from the newly input equipment monitoring data and incrementally updating the existing fault diagnosis module to ensure that the fault diagnosis module can adapt to the new data. The algorithm specifically modifies the loss function of the fault model; wherein the original loss function of the fault diagnosis model of the one-dimensional convolutional neural network is the classification loss function L c (x), the specific formula is as follows:

[0036]

[0037] Where |C| is the number of fault categories, y i is the label value of the sample, p i is the probability of the corresponding category obtained by the softmax function;

[0038] On the basis of the original loss function of the model, an incremental learning loss function L based on forgetting constraint is added g (x), the formula of the loss function is:

[0039]

[0040] in, and are the normalized features extracted by the original model and the current model, L g (x) The features extracted by the current model are encouraged to be as similar as possible to the features extracted by the original model. The loss function of the final fault diagnosis model is the cross entropy loss function L c (x) and the incremental learning loss function L based on the forgetting constraint g The weighted sum of (x) is as follows:

[0041] L(x)=λL g (x)+(1-λ)L c (x) (8);

[0042] Among them, L g The weight λ of (x) is set as a dynamic weight, and the formula is:

[0043]

[0044] Among them, |C n | and |C o | is the number of new and old category samples in each stage, λ base is a fixed constant for each dataset. In general, when the ratio of the number of new categories to the number of old categories increases, λ increases, and vice versa. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a system component module diagram of the present invention;

[0046] Figure 2 It is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0047] In order to help those skilled in the art better understand the technical solutions in this application, the technical solutions in the embodiments of this application are clearly and completely described below. Obviously, the described embodiments are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0048] Experimental Example 1:

[0049] This example uses a bearing dataset publicly available from Case Western Reserve University and bearing vibration data collected using the PT700 bearing test platform. The vibration signal sampling frequency in the dataset is 12 kHz, the load is set to 0 to 3 HP, and the bearing health status is divided into four states: normal, inner race fault, outer race fault, and rolling fault.

[0050] The process of the fault diagnosis module is as follows:

[0051] (1) First, a portion of the data for each bearing state in the dataset is selected and added to the representative sample set of the old category, and the fault diagnosis model is initialized using the old category samples;

[0052] (2) Input the operating data of the industrial Internet of Things equipment monitored by the sensor into the fault diagnosis model, perform fault diagnosis, and provide the diagnosis results;

[0053] (3) Over time, due to reasons such as equipment aging, workload changes, or noise interference, the concept drift of equipment monitoring data occurs, and the fault diagnosis model needs to be incrementally updated to ensure that the fault diagnosis model can adapt to data changes;

[0054] (4) The updated fault diagnosis model continues to be used for fault diagnosis until the next update.

[0055] The concept drift detection module detects whether concept drift occurs in the continuously input industrial IoT device operation data to trigger the incremental update process. The concept drift detection module process is as follows:

[0056] (1) Construct a time window, divide the data using the current time window and the historical time window, and input the data into the algorithm;

[0057] (2) Initialize the online extreme learning machine model using historical data in the historical time window and update the OS-ELM using current data in the current time window;

[0058] (3) Detect whether concept drift occurs in the data by calculating the Euclidean distance between the model hidden layer parameters obtained by historical and current data training and comparing the detection threshold;

[0059] (4) If concept drift is detected, the incremental update process of the fault diagnosis model is initiated.

[0060] After receiving the signal of concept drift, the incremental update module starts to incrementally update the fault diagnosis module. The specific steps are as follows:

[0061] (1) Save new input data S for N time windows N Used for incremental updates to determine whether new samples meet the update conditions; the update conditions are whether the number of samples reaches the updateable number NUM and whether the number of data in each category is equal or similar. N and NUM are both preset parameters;

[0062] (2) If the above update conditions are met, the old category representative sample dataset D RE The labeled samples of each old category in the data are input into the incremental learning algorithm based on forgetting constraints together with the newly arrived samples in the streaming data to incrementally update the fault diagnosis model;

[0063] (3) If the update condition is not met, it enters the waiting stage and waits for N extend time window to wait for the update condition to be met. If you wait until N extend If the update conditions are not met after the time window ends, the update will be started directly. extend are preset parameters;

[0064] (4) After the incremental update of the fault diagnosis module and the old category representative sample dataset D RE Afterwards, the incremental update module enters the waiting phase, waiting for the next round of incremental updates.

[0065] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. An industrial IoT equipment fault diagnosis system with adaptively triggered incremental learning, characterized by The system includes: a fault diagnosis module, a concept drift detection module and an incremental update module; The fault diagnosis module reads the operating data of industrial equipment through the industrial Internet of Things sensor and uses the fault diagnosis model to determine whether the equipment is in a normal state. If it is in a faulty state, it determines what kind of fault the equipment has; The concept drift detection module divides the equipment operation data by time window, uses historical data and current data to detect whether concept drift occurs in the data, and if concept drift is detected, activates the incremental update module to incrementally update the fault diagnosis module; The incremental update module uses the old category representative sample data and the new sample data to incrementally update the fault diagnosis module; The concept drift detection module uses an OS-ELM-based concept drift detection algorithm to detect concept drift of data. The specific steps are as follows: S3-1: Construct a time window and use it to divide the data. Input the data from the current time window and the historical time window into the algorithm. S3-2: Initialize the online extreme learning machine model using historical data in the historical time window and update the OS-ELM using current data in the current time window; S3-3: Detect whether concept drift occurs in the data by comparing the differences between the model hidden layer parameters obtained by historical and current data training; S3-4: If concept drift is detected, the incremental update process of the fault diagnosis model is started; The OS-ELM-based concept drift detection algorithm is responsible for partitioning device operation data and detecting whether concept drift occurs in the data. The specific calculation process is as follows: S4-1: Set two time windows TW HD With TW CD , among which TW HD Represents the historical time window, TW CD Represents the current time window, the length of the time window is L T ; Assume that TW is the initial HD The historical data in is defined as Indicates the interval L starting from time t=0 T The length of time, that is, [0,L T ] Data collected during the initial time TW CD The current data in is defined as Indicates that from time t = L T Start interval L T The length of time, that is, [L T ,2L T ]Data collected within a certain time period; S4-2: If no concept drift is detected, the right boundary of the historical time window is moved forward by the length of the time window, that is, the time interval is [0,2L T ], current time window TW CD The overall length of the forward time window is [2L T ,3L T ], and so on; if concept drift is detected, the historical time window TW is updated HD , TW HD Move to Taiwan CD The moment of the previous time window length; S4-3: Set the historical time window TW HD The data S1 in is input into the OS-ELM model. For a single hidden layer OS-ELM with L hidden units, the network model is expressed as: Among them, g(·) is the activation function, W=[w1,…,w L ] T and B=[b1,…,b L ] T are the weight and bias values ​​between the input layer and the hidden layer, β i is the weight value from the hidden layer to the output layer; Formula (1) is transformed into the existence of β i , w i and b i , making Formula (3) is expressed in matrix form, namely: Hβ=T (4); Where H is the hidden layer output value matrix, β is the weight value matrix from the hidden layer to the output layer; Using S1 data to train the OS-ELM model can obtain the model M(t0) at time t0, and the model parameters are Model M(t0) is called the baseline model, which represents the data distribution of the old data; S4-4: Set the current data time window TW CD The data S2 in the input is input into the OS-ELM model for model update. Since OS-ELM only modifies the model output weight β, it can obtain t1 = t0 + L T Model parameters at time The model trained at this time is called M(t1), which represents the data distribution of the new data; S4-5: By calculating the model parameters and The Euclidean distance between them can measure the difference between the model parameters trained using data at different times, so as to reflect whether concept drift occurs in the streaming data. The calculation formula for the model parameter difference is: S4-6: Set a detection threshold Th so that when the difference between the model before and after updating exceeds the threshold, that is, D(M(t0), M(t1))>Th, it can be determined that concept drift has occurred in the streaming data.

2. The industrial Internet of Things equipment fault diagnosis system with adaptively triggered incremental learning according to claim 1, characterized in that: The fault diagnosis module constructs a fault diagnosis model based on a one-dimensional convolutional neural network. Constructing a fault diagnosis model based on a one-dimensional convolutional neural network requires preparing a number of data of different equipment operation status categories in advance and labeling the data with corresponding labels. Then, the prepared data is used to train the fault diagnosis model using the Adam optimizer. After the model training is completed, the equipment fault is diagnosed based on the monitored equipment operation data.

3. The industrial Internet of Things equipment fault diagnosis system with adaptively triggered incremental learning according to claim 1, characterized in that: The incremental update module uses an incremental learning algorithm based on forgetting constraints to incrementally update the fault diagnosis module. The specific update process is as follows: S5-1: Construct a representative sample dataset D of old categories RE ,This dataset saves a constant number of representative samples for each category; S5-2: Save data S of N time windows N For incremental updates, the new sample data needs to carry labels; S5-3: The old category representative sample dataset D RE The labeled samples of each old category in the data are input into the model together with the newly arrived samples in the streaming data for incremental update.

4. The industrial Internet of Things equipment fault diagnosis system with adaptively triggered incremental learning according to claim 3 is characterized by: The incremental learning algorithm based on forgetting constraint is responsible for learning knowledge from the newly input equipment monitoring data and incrementally updating the existing fault diagnosis module to ensure that the fault diagnosis module can adapt to the new data. Specifically, the loss function of the fault model is modified; the original loss function of the fault diagnosis model of the one-dimensional convolutional neural network is the classification loss function L c (x), the specific formula is as follows: Where |C| is the number of fault categories, y i is the label value of the sample, p i is the probability of the corresponding category obtained by the softmax function; On the basis of the original loss function of the model, an incremental learning loss function L based on forgetting constraint is added g (x), loss function L g The formula for (x) is: in, and are the normalized features extracted by the original model and the current model, L g (x) The features extracted by the current model are encouraged to be as similar as possible to the features extracted by the original model. The loss function of the final fault diagnosis model is the cross entropy loss function L c (x) and the incremental learning loss function L based on the forgetting constraint g The weighted sum of (x) is as follows: L(x)=λL g (x)+(1-λ)L c (x) (8); Among them, L g The weight λ of (x) is set as a dynamic weight, and the formula is: Among them, |C n | and |C o | is the number of new and old category samples in each stage, λ base is a fixed constant for each dataset; when the ratio of the number of new categories to the number of old categories increases, λ increases, and vice versa.

Images