Mechanical heterogeneous fault diagnosis method and equipment based on multi-model collaborative neural architecture search

Abstract

The invention belongs to the related technical field of mechanical fault diagnosis, and discloses a mechanical heterogeneous fault diagnosis method and equipment based on multi-model collaborative neural architecture search, and the method comprises the steps: (1) processing a vibration acceleration signal to form a training set and a test set; (2) constructing a multi-model collaborative network, training the multi-model collaborative network by using the training sample, and optimizing the weight of each round of collaborative network; after each round of collaborative network training is completed, calculating an award obtained by the current collaborative network by adopting a test set, determining whether to update the structure of the current collaborative network according to a difference value between the award of the current collaborative network and the award of the previous round of collaborative network, and generating a candidate structure of the next round of collaborative network by adopting a neural architecture search framework based on an avoidance strategy; and (3) after the multi-model collaborative network identifies a task scene based on the sensor parameters of the multi-channel sensor, fault diagnosis is carried out. According to the invention, the diagnosis accuracy is improved while the heterogeneous fault diagnosis of the mechanical equipment is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field related to mechanical fault diagnosis, and more specifically, relates to a mechanical heterogeneous fault diagnosis method and device based on multi-model collaborative neural architecture search. Background Art

[0002] In modern industrial production, mechanical equipment plays an important role in all walks of life, and their intelligent operation and maintenance have attracted the attention of many scholars and engineers in recent years. With the development of artificial intelligence, mechanical intelligent fault diagnosis methods based on deep learning have made great progress, and rotating machinery data sets in many laboratories, such as gears and bearings, have been deeply analyzed. While these data-driven fault diagnosis methods continue to improve the diagnostic accuracy, they still have some limitations.

[0003] Most methods focus on diagnostic objects with complete data acquisition, such as bearing fault simulation experimental platforms, that is, the input of deep learning models is always fixed data shape and channel dimension. Few studies consider differentiated scenarios between multiple devices, but in industry, the original data may be incomplete due to communication anomalies, sensor failures, and differences in diagnostic capabilities between different downstream companies. It is difficult to handle such heterogeneous fault diagnosis tasks with different signal dimensions through deep learning models with only a single type of input. It is usually necessary to construct a new deep learning model architecture to adapt to changes in different data structures and distributions, which greatly increases the time cost and consumption of computing resources.

[0004] Although the fault diagnosis method based on deep learning avoids manual feature extraction, scholars generally need to continuously conduct structural design and experimental trial and error after mastering rich professional knowledge and field experience when selecting model structures and hyperparameters. Deep learning models are highly sensitive to changes in data sets, and even to vibration signals of different parts of a machine collected by sensors. In different network structures, how to choose the right model for specific diagnostic scenarios and signals, researchers still do not have a unified answer.

[0005] Neural architecture search is an automated method for designing neural network models. It optimizes candidate networks through search strategies and verification feedback to obtain high-performance models suitable for specific tasks. However, as the search space expands and the number of model layers increases, the computing resources consumed by this method grows exponentially. For some small and medium-sized enterprises, the cost of intelligent operation and maintenance will increase significantly. In addition, the current mainstream neural architecture search methods are usually task-specific and perform detailed searches and hyperparameter optimizations on a single structure of a neural network. This method often does not have an advantage when facing multiple tasks. When companies use network searches, they will incur high costs to solve a machine failure problem. In summary, these problems hinder the promotion of fault diagnosis based on deep learning methods in industrial applications. Summary of the invention

[0006] In response to the above defects or improvement needs of the prior art, the present invention provides a mechanical heterogeneous fault diagnosis method and device based on multi-model collaborative neural architecture search, which aims to realize heterogeneous fault diagnosis of mechanical equipment while improving the accuracy of diagnosis.

[0007] To achieve the above object, according to one aspect of the present invention, a mechanical heterogeneous fault diagnosis method based on multi-model collaborative neural architecture search is provided, and the method comprises the following steps:

[0008] (1) The vibration acceleration signal of the mechanical equipment to be diagnosed collected by the multi-channel sensor is normalized, and the normalized vibration acceleration signal is resampled by using a sliding window to obtain a series of time window samples, and labels are added to the time window samples to form a training set and a test set;

[0009] (2) Use multiple models to construct a multi-model collaborative network, and use training samples to train the multi-model collaborative network and optimize the weight of each round of collaborative network; after each round of collaborative network training is completed, use the test set to calculate the reward obtained by the current collaborative network, and determine whether to update the structure of the current collaborative network based on the difference between the rewards of the current collaborative network and the previous round of collaborative network, and use a neural architecture search framework based on an avoidance strategy to generate candidate structures for the next round of collaborative networks;

[0010] (3) Multi-model collaborative network based on sensor parameters of multi-channel sensors After identifying the task scenario, the corresponding sub-model is activated to perform fault diagnosis.

[0011] Furthermore, the training set includes sensor data of all channels, and the test set includes sensor data of some or all channels.

[0012] Furthermore, the training samples of the entire channel are divided into multiple one-dimensional single-channel data, and each sub-model extracts features through the single-channel data; all sub-models are connected to the same decision network.

[0013] Furthermore, for each sub-model, the multi-label based binary classification loss is calculated:

[0014]

[0015] in To use the bth model for n c The predicted value of the i-th sample of the category; during the collaborative training process, the final inter-ring loss L in Affected by each sub-model, it is expressed as:

[0016]

[0017] Where N i and B are the number of batch samples and sub-models participating in each round of collaborative training.

[0018] Furthermore, the training of the multi-model collaborative network includes inner-loop learning and outer-loop learning based on neural architecture search with avoidance strategy; the outer-loop learning includes performance evaluation of the trained multi-model collaborative network in the current e-th period and generation of candidate multi-model collaborative networks in the (e+1)-th period.

[0019] Furthermore, each multi-model collaborative network architecture state comes from the search space of the neural network, which contains n potential combinations of states S = {S1, S2, ..., S n}, each state space Contains two pieces of information, namely the branch index to be updated The index of the candidate submodel in model space And the model structure

[0020] Furthermore, based on the task information p j Reconstruction state is S e In the multi-model collaborative network, the devices corresponding to the test set samples do not overlap with the devices corresponding to the training set.

[0021] Furthermore, the calculation formula for the average cross-individual accuracy is:

[0022]

[0023] in is the index parameter of the sensor, N p is the total number of heterogeneous tasks, and Evl(·) represents the classic neural network evaluation function.

[0024] The present invention also provides a mechanical heterogeneous fault diagnosis system based on multi-model collaborative neural architecture search, the fault diagnosis system includes a memory and a processor, the memory stores a computer program, and the processor executes the mechanical heterogeneous fault diagnosis method based on multi-model collaborative neural architecture search as described above when executing the computer program.

[0025] The present invention also provides a computer-readable storage medium, which stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the mechanical heterogeneous fault diagnosis method based on multi-model collaborative neural architecture search as described above.

[0026] In general, compared with the prior art, the above technical solution conceived by the present invention, the mechanical heterogeneous fault diagnosis method and device based on multi-model collaborative neural architecture search provided by the present invention mainly have the following beneficial effects:

[0027] 1. The present invention uses the collected multi-channel vibration acceleration signals as the basis for multi-model training and testing, and selects a separate model for the vibration acceleration signal of each channel for decision-making, thereby reducing the risk of overfitting of a single model for specific data, significantly improving the accuracy of mechanical equipment fault diagnosis, and being able to simultaneously realize the diagnosis of mechanical heterogeneous faults.

[0028] 2. A neural architecture search framework based on an avoidance strategy is used to generate candidate structures for the next round of collaborative networks, which can achieve end-to-end diagnosis without complex expert knowledge, greatly reducing search time and computational consumption. There is no need to refine the model from scratch, and it can prevent the model from falling into the local optimal situation, thereby improving search efficiency.

[0029] 3. The equipment corresponding to the test set samples does not overlap with the equipment corresponding to the training set. The model can perform fault diagnosis on any unknown machinery of the same model through known machinery, which greatly promotes the application of deep learning-based fault diagnosis in industrial scenarios.

[0030] 4. The present invention directly uses multiple high-level model structures as the basic elements of the search, which greatly reduces the search space and computational cost, while making full use of the knowledge and achievements in vertical fields. At the same time, the inner loop learning only trains and updates the modified sub-model structure and weights, and the remaining branches are frozen to further reduce the computational cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of a mechanical heterogeneous fault diagnosis method based on multi-model collaborative neural architecture search provided by the present invention;

[0032] Figure 2 It is the T-SNE diagram of the original data;

[0033] Figure 3 It is a schematic diagram of data collection and segmentation;

[0034] Figure 4 Schematic diagram of the iterative optimization process of outer-loop neural architecture search. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0036] First, the problem is defined. For multiple motors on an industrial production line, each motor contains multi-channel acceleration vibration sensor signals. The motor data of the training set is represented as M represents the number of motors, n i represents the number of samples, represents the i-th short-time data sample obtained from the m-th motor, L represents the length of the sample, and C represents the width of the sample, which is determined by the number of sensors on a motor. Indicates the corresponding label.

[0037] Heterogeneous faults can be viewed as a multi-task scenario characterized by differences in data dimensionality between the test set and the training set. This heterogeneity arises from changes in device states or sudden data loss, such as communication problems, system upgrades or modifications, etc. Therefore, the target device signal of the test set is defined as c is an integer between 1 and C, and the additional parameter Used to encode the sensor position.

[0038] According to general research in the field of fault diagnosis, the distribution differences of signal data collected from different machines are significant, namely:

[0039]

[0040] The main purpose of the present invention is to establish a universal classification model suitable for multi-task processing to handle heterogeneous signal inputs and improve diagnostic accuracy. In addition, it is also required to search for the optimal collaborative neural network structure for heterogeneous tasks as efficiently as possible in the model level space.

[0041] See also Figure 1 , Figure 2 , Figure 3 and Figure 4 The present invention provides a mechanical heterogeneous fault diagnosis method based on multi-model collaborative neural architecture search, and the fault diagnosis method mainly includes the following steps:

[0042] Step 1: Normalize the vibration acceleration signal of the mechanical equipment to be diagnosed collected by the multi-channel sensor, and resample the normalized vibration acceleration signal using a sliding window to obtain a series of time window samples, add labels to the time window samples, and form a training set and a test set.

[0043] Normalize the vibration acceleration signal to initially reduce the inherent differences and instantaneous impact effects at different locations. Use a traditional sliding window to sample and construct a batch of time window samples to create composite fault labels for subsequent training and testing. The training set contains sensor data from all channels, and the test set contains sensor data from some or all channels.

[0044] Step 2: Use multiple models to construct a multi-model collaborative network, and use training samples to train the multi-model collaborative network and optimize the weights of each round of collaborative network; after each round of collaborative network training is completed, use the test set to calculate the reward obtained by the current collaborative network, and determine whether to update the structure of the current collaborative network based on the difference between the rewards of the current collaborative network and the previous round of collaborative network, and use a neural architecture search framework based on an avoidance strategy to generate candidate structures for the next round of collaborative network, and repeat this cycle until the predetermined number of training times is reached.

[0045] This step mainly includes the inner-loop learning of the multi-model collaborative network and the outer-loop learning of the neural architecture search based on the avoidance strategy. Different from the traditional single model, multiple models are used for collaborative training and fusion decision-making to form a multi-model collaborative network (MMCN). The neural architecture search framework based on the avoidance strategy is used to optimize the structure of the MMCN. The training samples are used to train and optimize the weights of each round of the collaborative network. The test samples are used to calculate the rewards obtained by the current collaborative network and guide whether to update the collaborative network structure.

[0046] The multiple models used in this implementation are mainly some classic deep learning models and advanced domain generalization models:

[0047] 1) WDCNN: A classic and widely compared convolutional neural network based fault diagnosis model, used as the initial model structure of each branch.

[0048] 2) MSMCNN: A multi-scale hybrid convolutional neural network for multi-channel signal fault diagnosis of industrial robots.

[0049] 3) MSRFCNN: A multi-stage residual fusion convolutional neural network for in-situ fault detection of motors in CNC machine tool spindles.

[0050] 4) MSRCN: A cross-machine fault diagnosis model for complex faults in industrial motors.

[0051] 5) CFSPT: A lightweight Transformer model for cross-machine fault diagnosis of electric motors.

[0052] 6) MSPT: Multi-Scale Pruned Transformer Model for Cross-Machine Fault Diagnosis.

[0053] Inner-loop learning of multi-model collaborative networks: In inner-loop learning, unlike traditional single models, multiple models participate in training and decision output. In order to adapt to various channel dimensions, the training samples of the full channel are divided into multiple one-dimensional single-channel data, and each sub-model extracts features through single-channel data. All sub-models are connected to the same decision network, which means that the weight update and output of each sub-model are synchronized. For each sub-model, the binary classification loss based on multiple labels is calculated:

[0054]

[0055] in To use the bth model for n c The predicted value of the i-th sample of the category. During the collaborative training process, the final inter-ring loss L in Affected by each sub-model, it can be expressed as:

[0056]

[0057] Where N i and B are the number of batch samples and sub-models participating in each round of collaborative training. After the outer loop structure is updated, only the changed sub-model participates in training and weight update, which further reduces the computational consumption. Multiple deep learning models participate in the training process to form a multi-model collaborative network MMCN, and perform multiple iterations based on the classic back-propagation algorithm.

[0058] Outer loop neural architecture search learning based on avoidance strategy: This part mainly includes the performance evaluation of the trained MMCN in the current e-th period and the generation of candidate MMCN in the (e+1)th period. Each MMCN architecture state comes from the search space of the neural network, which contains a potential combination of n states S = {S1, S2, ..., S n}, each state space Contains two pieces of information, namely the branch index to be updated The index of the candidate submodel in model space And the model structure Different from the traditional neural architecture search with meticulous basic element objects (such as the number of nodes and layers), this diagnostic method directly takes multiple high-level model structures as the basic elements of the search, which greatly reduces the search space and computational cost while making full use of the knowledge and achievements in vertical fields.

[0059] At the beginning of each outer loop, two indexes generated according to the avoidance strategy and like Figure 4 As shown. The algorithm receives the e-th round and As the starting point of the search, it receives a fixed-length stack I S ∈R 2×K , this stack provides short-term storage for tracking the model search history within the last K attempts. An avoidance-based diversity search strategy (ADS) is used to randomly generate new candidate values ​​and perform a first check to see if they are the same as the input value:

[0060]

[0061]

[0062] Where I represents the index space, and the two generated values ​​also need to be different from the most recent K combinations to further enhance the diversity of the search.

[0063] Then execute the second judgment:

[0064]

[0065] once Through these two judgments, it is considered to be a valid and diverse candidate indicator group. At the same time, the oldest group in the stack is removed to maintain the size of the stack as K.

[0066] In order to complete the multi-task performance evaluation, based on the task information p j Reconstruction state is S e MMCN, where the devices corresponding to the test set samples do not overlap with the devices corresponding to the training set to ensure generalization performance, and calculate the average cross-individual accuracy (CIA):

[0067]

[0068] in is the index parameter of the sensor, which can guide the selection of the corresponding sub-model, N pis the total number of heterogeneous tasks, and Evl(●) represents the classic neural network evaluation function. Generated according to the avoidance-based diversity search strategy Get the status as S e+1 The candidate MMCN is trained in the inner loop to obtain a CIA similar to the above formula e+1 Note that the inner loop learning only trains and updates the modified sub-model structure and weights, and the remaining branches are frozen to further reduce the computational cost. The reward R of the (e+1)th search e+1 It can be calculated as:

[0069] R e+1 =CIA e+1 -CIA e

[0070] Whether MMCN updates the structure in the (e+1)th search of the outer loop is based on the reward R e+1 Among them, E M Represents the model-level search space. After N rounds of iterations, the outer loop automatically searches and outputs the optimal MMCN structure. and the corresponding weights.

[0071]

[0072] Step 3: Multi-model collaborative network based on sensor parameters of multi-channel sensors After identifying the task scenario, the corresponding sub-model is activated to perform fault diagnosis.

[0073] Based on sensor parameters Able to identify the corresponding sensor and the channel and location of the sensor.

[0074] The purpose of this step is to diagnose signals of various input dimensions. The multi-model collaborative network receives sensor parameters to recognize the task scenario; the corresponding neural networks are activated and reassembled into a matching collaborative network for the final diagnosis.

[0075] The present invention also provides a mechanical heterogeneous fault diagnosis system based on multi-model collaborative neural architecture search, the fault diagnosis system includes a memory and a processor, the memory stores a computer program, and the processor executes the mechanical heterogeneous fault diagnosis method based on multi-model collaborative neural architecture search as described above when executing the computer program.

[0076] The present invention also provides a computer-readable storage medium, which stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the mechanical heterogeneous fault diagnosis method based on multi-model collaborative neural architecture search as described above.

[0077] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A mechanical heterogeneous fault diagnosis method based on multi-model collaborative neural architecture search, characterized in that: The method comprises the following steps: (1) The vibration acceleration signal of the mechanical equipment to be diagnosed collected by the multi-channel sensor is normalized, and the normalized vibration acceleration signal is resampled by using a sliding window to obtain a series of time window samples, and labels are added to the time window samples to form a training set and a test set; (2) Use multiple models to construct a multi-model collaborative network, and use training samples to train the multi-model collaborative network and optimize the weight of each round of collaborative network; after each round of collaborative network training is completed, use the test set to calculate the reward obtained by the current collaborative network, and determine whether to update the structure of the current collaborative network based on the difference between the rewards of the current collaborative network and the previous round of collaborative network, and use a neural architecture search framework based on an avoidance strategy to generate candidate structures for the next round of collaborative networks; (3) Multi-model collaborative network based on sensor parameters of multi-channel sensors After identifying the task scenario, the corresponding sub-model is activated to perform fault diagnosis.

2. The method for mechanical heterogeneous fault diagnosis based on multi-model collaborative neural architecture search according to claim 1, characterized in that: The training set contains sensor data of all channels, and the test set contains sensor data of some or all channels.

3. The method for mechanical heterogeneous fault diagnosis based on multi-model collaborative neural architecture search according to claim 1, characterized in that: The full-channel training samples are divided into multiple one-dimensional single-channel data, and each sub-model extracts features through the single-channel data; all sub-models are connected to the same decision network.

4. The method for mechanical heterogeneous fault diagnosis based on multi-model collaborative neural architecture search according to claim 3, characterized in that: For each sub-model, calculate the multi-label based binary classification loss: in To use the bth model for n c The predicted value of the i-th sample of the category; during the collaborative training process, the final inter-ring loss L in Affected by each sub-model, it is expressed as: Where N i and B are the number of batch samples and sub-models participating in each round of collaborative training.

5. The method for mechanical heterogeneous fault diagnosis based on multi-model collaborative neural architecture search according to claim 1, characterized in that: The training of the multi-model collaborative network includes inner-loop learning and outer-loop learning based on neural architecture search with avoidance strategy; the outer-loop learning includes the performance evaluation of the trained multi-model collaborative network in the current e-th period and the generation of candidate multi-model collaborative networks in the (e+1)-th period.

6. The method for mechanical heterogeneous fault diagnosis based on multi-model collaborative neural architecture search according to claim 5, characterized in that: Each multi-model collaborative network architecture state comes from the search space of the neural network, which contains n potential combinations of states S = {S1, S2, ..., S n }, each state space Contains two pieces of information, namely the branch index to be updated The index of the candidate submodel in model space And the model structure 7. The method for mechanical heterogeneous fault diagnosis based on multi-model collaborative neural architecture search according to claim 1, characterized in that: Based on the task information p j Reconstruction state is S e In the multi-model collaborative network, the devices corresponding to the test set samples do not overlap with the devices corresponding to the training set.

8. The method for mechanical heterogeneous fault diagnosis based on multi-model collaborative neural architecture search according to claim 1, characterized in that: The calculation formula for the average cross-individual accuracy is: in is the index parameter of the sensor, N p is the total number of heterogeneous tasks, and Evl(·) represents the classic neural network evaluation function.

9. A mechanical heterogeneous fault diagnosis system based on multi-model collaborative neural architecture search, characterized by: The fault diagnosis system includes a memory and a processor, the memory stores a computer program, and the processor executes the mechanical heterogeneous fault diagnosis method based on multi-model collaborative neural architecture search as described in any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are called and executed by the processor, the machine-executable instructions prompt the processor to implement the mechanical heterogeneous fault diagnosis method based on multi-model collaborative neural architecture search as described in any one of claims 1-8.

Images