Device status recognition method based on ESOINN

Abstract

The present invention relates to the field of equipment status monitoring technology, and discloses an ESOINN-based equipment status identification method, which aims to solve the problem of poor versatility and accuracy in existing methods for identifying equipment operating status. The solution mainly includes: collecting vibration data of the equipment according to a first preset period, and obtaining vibration data within multiple first preset periods according to a second preset period; incrementally updating the ESOINN network according to the obtained vibration data, and determining the label type corresponding to each vibration data based on the ESOINN network; determining the equipment operating status corresponding to each vibration data within each second preset period according to the label type and the neuron node data corresponding to each label type. The present invention improves the versatility and accuracy of equipment operating status judgment and is particularly suitable for industrial equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment status monitoring, and in particular to an equipment status identification method based on ESOINN. Background Art

[0002] The healthy and stable operation of industrial equipment has always been a key concern for manufacturers. Generally speaking, methods for obtaining equipment operating status fall into two categories: directly integrating with the equipment system to collect operating data, and non-invasively collecting data through external sensors and other methods to identify equipment status. Direct integration with the equipment system offers the advantage of accurate and reliable data, but drawbacks include high cost, limited versatility, and the difficulty of integrating with some older equipment. Non-invasive methods, on the other hand, offer greater versatility, are relatively simple to operate, and are relatively cost-effective.

[0003] Accurately and reasonably determining the status of equipment without intrusion is crucial for monitoring the operational status of IIoT devices. This helps manufacturers better understand the efficiency and health of their equipment, and promptly detect anomalies. There are two common methods for identifying status: one is to manually set thresholds, which requires prior knowledge and, because different devices have different operational states, is not universal or has poor accuracy. The other method uses artificial intelligence and machine learning algorithms, which offers greater applicability.

[0004] When it comes to industrial equipment status identification, the actual operating status of the equipment is usually unknown, making it difficult to use supervised learning methods such as neural networks and deep learning for modeling. On the other hand, due to the special nature of industrial equipment's long-term operation, new data is generated every moment. If the algorithm model is not updated in a timely manner, inaccurate status identification results may occur. Summary of the Invention

[0005] The present invention aims to solve the problems of poor versatility and accuracy in existing methods for identifying the operating status of equipment, and proposes an equipment status identification method based on ESOINN.

[0006] The technical solution adopted by the present invention to solve the above technical problems is:

[0007] On the one hand, a device status identification method based on ESOINN is provided, comprising the following steps:

[0008] Step 1: collecting vibration data of the device according to a first preset period, and obtaining vibration data within a plurality of first preset periods according to a second preset period;

[0009] Step 2: Incrementally update the ESOINN network based on the acquired vibration data, and determine the label type corresponding to each vibration data based on the ESOINN network;

[0010] Step 3: Determine the device operating status corresponding to each vibration data within each second preset period according to the tag type and the neuron node data corresponding to each tag type.

[0011] As a further optimization, the step 2 specifically includes:

[0012] Obtain historical vibration data of the device and pre-train the ESOINN network based on the historical vibration data;

[0013] The pre-trained ESOINN network is updated according to the acquired vibration data, and the label type corresponding to each vibration data is determined based on the pre-trained ESOINN network.

[0014] As a further optimization, the step 3 specifically includes:

[0015] If the label type corresponding to the vibration data is a first-class label, the corresponding vibration data is determined to be outlier data, and the device operation state corresponding to the vibration data is determined to be an abnormal state.

[0016] As a further optimization, the method further includes:

[0017] If there is vibration data corresponding to the abnormal operation status of the equipment, the abnormal information will be sent to the designated contact through the communication module.

[0018] As a further optimization, the step 3 specifically further includes:

[0019] Among all vibration data within the second preset period, if, except for vibration data with a tag type of the first type, the tag type corresponding to the remaining vibration data is one, it is determined that the device operating state corresponding to the remaining vibration data is the power-on state.

[0020] As a further optimization, step 3 specifically further includes:

[0021] Among all the vibration data within the second preset period, if, except for the vibration data with the label type of the first type of label, the remaining vibration data correspond to two types of labels, the neuron means of all the vibration data corresponding to the two types of labels are calculated respectively, and the device operation status of all the vibration data corresponding to the label type with a larger neuron node mean is determined to be the power-on state, and the device operation status of all the vibration data corresponding to the label type with a smaller neuron node mean is determined to be the power-off state.

[0022] As a further optimization, step 3 specifically further includes:

[0023] Among all the vibration data within the second preset period, if, except for the vibration data with the tag type of the first category tag, the tag types corresponding to the remaining vibration data are greater than two, then the weight values ​​corresponding to each type of tag except the first category tag are calculated respectively, and the device operation status of all the vibration data corresponding to the tag type with a positive and largest weight value is determined to be the power-on state, the device operation status of all the vibration data corresponding to the tag type with a positive and smallest weight value is determined to be the power-off state, and the device operation status of all the vibration data corresponding to the remaining tag types is determined to be the power-off state.

[0024] As a further optimization, the weight values ​​corresponding to each type of label except the first type of label are calculated separately, specifically including:

[0025] Determine the number of neuron nodes corresponding to each type of label except the first type of label, and determine the median corresponding to the number of neuron nodes;

[0026] Eliminate the label types whose number of neuron nodes is less than the median and set their weights to 0, obtaining m types of remaining labels;

[0027] The number of neuron nodes and the mean amplitude corresponding to each type of remaining label are determined respectively, and the weight value corresponding to each type of remaining label is calculated according to the mean amplitude and the number of neuron nodes.

[0028] As a further optimization, the vibration data is the amplitude data of the device in the x, y and z directions, and the calculation formula of the amplitude mean is as follows:

[0029]

[0030] Where μ i represents the mean amplitude of the i-th label in the x, y and z directions, x ij Indicates the characteristic amplitude value of the j-th neuron node of the i-th category label in the x direction, y ij Indicates the characteristic amplitude value of the jth neuron node of the i-th category label in the y direction, z ij Indicates the characteristic amplitude value of the j-th neuron node of the i-th category label in the z direction, l i The number of neuron nodes representing the i-th category label.

[0031] As a further optimization, the calculation formula of the weight value is as follows:

[0032]

[0033] in,

[0034] Where, δ iRepresents the weight value of the i-th category label, μ avg represents the average of the amplitude means of the remaining labels of each type, l avg Represents the average number of neuron nodes of each type of remaining labels; m represents the number of types of remaining labels.

[0035] The beneficial effects of the present invention are:

[0036] 1. Based on a non-invasive method, it supports any device that can be connected to a vibration sensor, with good applicability and low cost.

[0037] 2. Vibration data is transmitted to the ESOINN network in real time for identification and training, supporting online incremental updates to ensure the real-time performance of the algorithm model and good adaptability. Even if the overall operating characteristics of the equipment change, there is no need for manual intervention in the algorithm. The algorithm can gradually adjust the feature distribution without affecting the previous results.

[0038] 3. Through weight analysis, the operating status can be adaptively judged and the classification output by the algorithm can be converted into the equipment status without any human prior knowledge, and it has high versatility.

[0039] 4. Connect to the industrial Internet of Things platform to check the vibration and operating status of the equipment at any time and send abnormal notifications in time. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a schematic diagram of a first flow chart of the device status identification method based on ESOINN according to an embodiment of the present invention;

[0041] Figure 2 Schematic diagram showing the distribution of vibration amplitudes of the device A in three-dimensional space according to an embodiment of the present invention;

[0042] Figure 3 Schematic diagram of the topological structure of the pre-trained ESOINN network according to an embodiment of the present invention;

[0043] Figure 4 This is a second flow chart of the device status identification method based on ESOINN according to an embodiment of the present invention;

[0044] Figure 5 This is a third flow chart of the device status identification method based on ESOINN according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] The device status identification method based on ESOINN provided by the present invention includes the following steps: Step 1, collecting vibration data of the device according to a first preset period, and obtaining vibration data within multiple first preset periods according to a second preset period; Step 2, incrementally updating the ESOINN network according to the obtained vibration data, and determining the label type corresponding to each vibration data based on the ESOINN network; Step 3, determining the device operation status corresponding to each vibration data within each second preset period according to the label type and the neuron node data corresponding to each label type.

[0047] Enhanced Self-organizing Incremental Neural Network (ESOINN) is an algorithm based on neuron competitive learning. It generates a group of neurons distributed in the feature space to characterize the density distribution of the input data. The neurons are connected to each other to form multiple subgraphs, which represent the clustering results obtained by algorithm training and the topological structure of the data in the feature space. The biggest feature of ESOINN is that its training process is completely incremental, that is, it can dynamically generate new neurons adaptively to changes in the features of new data without affecting the original neuron structure. At the same time, ESOINN is different from commonly used clustering algorithms such as k-means. It does not require the number of clusters to be specified in advance, is not sensitive to the initial cluster center, and has better applicability when prior knowledge is insufficient. Based on this, the present invention uses the ESOINN network to classify the vibration data of the equipment, and judges the equipment operating status corresponding to each vibration data based on the type of label and the neuron node data corresponding to each label type.

[0048] Example

[0049] The device status identification method based on ESOINN according to the embodiment of the present invention is as follows: Figure 1 As shown, the following steps are included:

[0050] Step 1: collecting vibration data of the device according to a first preset period, and obtaining vibration data within a plurality of first preset periods according to a second preset period;

[0051] In this embodiment, a vibration sensor can be used to obtain the vibration amplitude data of the device in the x, y and z directions, and the collection and reporting frequency is once per minute, that is, the first preset period is 1 minute. Figure 2 The figure shows the distribution of the vibration amplitude of a real device A in the x, y, and z directions in three dimensions in three dimensions in μm, with a total of 33,520 items.

[0052] This embodiment obtains 10 pieces of vibration amplitude data, that is, the second preset period is 10 minutes, and the device operating status corresponding to these 10 pieces of vibration data is determined based on the ESOINN network.

[0053] Step 2: Incrementally update the ESOINN network based on the acquired vibration data, and determine the label type corresponding to each vibration data based on the ESOINN network;

[0054] In this embodiment, the collected vibration data is stored in a database for initial training of the ESOINN network. Furthermore, the data is directly brought into the pre-trained ESOINN network via MQTT for classification and incremental updates. It should be noted that the ESOINN algorithm also supports skipping pre-training with historical data and directly training with incremental updates. However, given the accuracy issues of the ESOINN network in its early stages, where insufficient data may result in the inability to form effective clusters or the resulting clusters being single and unrecognizable, pre-training with historical vibration data ensures that the ESOINN network achieves good classification results early on.

[0055] Figure 3 The topological structure obtained by pre-training the ESOINN algorithm on the historical data set of device A is shown. It can be clearly seen that the vibration data in the three-dimensional space is divided into two parts. Combined with the actual situation, it can be seen that the neuron subgraph on the left represents the shutdown state, while the neuron subgraph on the right represents the power-on state.

[0056] Step 3: Determine the device operating status corresponding to each vibration data within each second preset period according to the tag type and the neuron node data corresponding to each tag type.

[0057] like Figure 4 As shown in the figure, after obtaining the pre-trained ESOINN network, the latest vibration data of device A is obtained through MQTT. The vibration amplitude in the three directions of x, y and z is set to (8.5, 9.7, 0.8), and its unit is μm. The pre-trained ESOINN network is loaded and the new data is input to obtain a new ESOINN network. According to the output results of the algorithm code of the ESOINN network, the neuron topology structure generates a total of 124 neuron nodes, and there are 61 with the label "0", corresponding to Figure 3 In the left part, there are 63 labels with the value "1", corresponding to Figure 3 On the right side of the , the label of the new data classification is "1". If the new data does not belong to any category of the existing structure, the output label is "-1". Figure 3 Therefore, the topology of the ESOINN network has not changed.

[0058] Since the ESOINN algorithm can only classify data and cannot directly give the corresponding device operating status, this embodiment adopts a method based on the tag type and the neuron node data corresponding to each tag type to determine the device operating status corresponding to each vibration data, such as Figure 5 As shown, specifically including:

[0059] A. If the label type corresponding to the vibration data is a first-class label, the corresponding vibration data is determined to be outlier data, and the device operation state corresponding to the vibration data is determined to be an abnormal state.

[0060] Specifically, for vibration data that does not belong to any existing type of structure, that is, outlier data, the ESOINN network will output a label of "-1". Therefore, combined with reality, it is not difficult to understand that such points should be considered "abnormal", that is, the label "-1" is marked as "abnormal".

[0061] B. For vibration data with a label other than "-1", continue to determine the corresponding device operating status using the following method:

[0062] Since the number of subgraphs obtained by the ESOINN network (used to represent the number of label types) is uncontrollable, there are three cases:

[0063] a. Among all vibration data within the second preset period, if, except for vibration data with a tag type of the first type, the tag type corresponding to the remaining vibration data is one, it is determined that the device operating state corresponding to the remaining vibration data is the power-on state.

[0064] It can be understood that among the 10 vibration data within the second preset period, if the label type corresponding to the remaining vibration data whose label is not "-1" is one, that is, there is only one sub-graph, then the device operating status corresponding to the remaining vibration data will be set to the power-on status.

[0065] b. Among all the vibration data within the second preset period, if, except for the vibration data with the label type of the first type of label, the remaining vibration data correspond to two types of labels, the neuron means of all the vibration data corresponding to the two types of labels are calculated respectively, and the device operation status of all the vibration data corresponding to the label type with a larger neuron node mean is determined to be the power-on state, and the device operation status of all the vibration data corresponding to the label type with a smaller neuron node mean is determined to be the power-off state.

[0066] It can be understood that among the 10 vibration data within the second preset period, if the remaining vibration data whose labels are not "-1" correspond to two types of labels, that is, the number of sub-graphs is equal to 2, then it is not difficult to understand in combination with reality that the group with a larger mean value of neuron nodes is in the running state, and the group with a smaller mean value of neuron nodes is in the shutdown state, that is, the operating state of the equipment is judged by the mean value of its neuron nodes.

[0067] c. Among all the vibration data within the second preset period, if, except for the vibration data with the tag type of the first category tag, the tag types corresponding to the remaining vibration data are greater than two, then the weight values ​​corresponding to each type of tag except the first category tag are calculated respectively, and the device operation status of all the vibration data corresponding to the tag type with a positive and largest weight value is determined to be the power-on state, the device operation status of all the vibration data corresponding to the tag type with a positive and smallest weight value is determined to be the power-off state, and the device operation status of all the vibration data corresponding to the remaining tag types is determined to be the power-off state.

[0068] It can be understood that among the 10 vibration data within the second preset period, if the remaining vibration data with a label other than "-1" corresponds to more than two types of labels, that is, the number of sub-graphs is greater than 2, it is necessary to calculate the weight value corresponding to each label except the label "-1" respectively, and determine the device operating status corresponding to the vibration data of each type of label based on the weight value, specifically including:

[0069] Step c1: Determine the number of neuron nodes l corresponding to each type of label except the first type of label i , and determine the median l corresponding to the number of neuron nodes me , where l me =Median(l i ), l i The number of neuron nodes representing the i-th category label;

[0070] Step c2: remove the label types whose number of neuron nodes is less than the median and set their weights to 0, obtaining m types of remaining labels. The purpose of removing labels is to exclude label types with a smaller number to reduce the system's computational pressure.

[0071] Step c3: determine the number of neuron nodes and the mean amplitude corresponding to each type of remaining label, and calculate the weight value corresponding to each type of remaining label based on the mean amplitude and the number of neuron nodes.

[0072] Specifically, first calculate the amplitude mean μ corresponding to each type of remaining label based on the amplitude data and the number of neuron nodes corresponding to each type of remaining label i , the calculation formula is as follows:

[0073]

[0074] Where μ i represents the mean amplitude of the i-th label in the x, y and z directions, x ij Indicates the characteristic amplitude value of the j-th neuron node of the i-th category label in the x direction, y ij Indicates the characteristic amplitude value of the jth neuron node of the i-th category label in the y direction, z ij Indicates the characteristic amplitude value of the j-th neuron node of the i-th category label in the z direction, l i The number of neuron nodes representing the i-th category label.

[0075] Then calculate the average value μ of the amplitude mean of each type of remaining labels avg , the calculation formula is as follows:

[0076]

[0077] Then calculate the average value l of the number of neuron nodes of each type of remaining labels avg , the calculation formula is as follows:

[0078]

[0079] Finally, the weight values ​​δ of the remaining labels are calculated based on the mean amplitude and the number of neuron nodes. i , the calculation formula is as follows:

[0080]

[0081] From the above formula, we can see that the parameters that determine the size of the weight value are mainly the amplitude mean and the number of neuron nodes. Combined with actual analysis, when a certain amount of data has been accumulated and the industrial equipment is working stably, the neuron nodes falling into the power-on and power-off states should account for the majority of the proportion, and the abnormal and intermediate states should account for a relatively small proportion. From the above formula, we can see that after taking the median, the weight value corresponding to the smaller number of label types will be 0, and the weight value corresponding to the larger number of label types will be greater than 0. At the same time, according to the above formula, the larger the amplitude mean corresponding to the label, the greater the weight value. Among the labels with weight values ​​greater than 0, the largest one is selected as the "power-on" state, the smallest one as the "power-off" state, "-1" represents the "abnormal" state, and the other labels are regarded as "intermediate states."

[0082] The advantage of this method is that it is not easily affected by individual special nodes, such as individual minimum or maximum values. The median operation effectively filters out the interference of individual minimum values, and does not cause other categories to be classified as "small values" due to excessive maximum values. In the weight calculation formula, the influence of quantity on the weight value is weakened, and the influence of amplitude mean on the weight value is strengthened. This is to avoid the phenomenon of weight inversion caused by too many small values. On the other hand, if the operating status of the device changes, the algorithm will automatically adjust the status corresponding to each label through data accumulation.

[0083] In order to facilitate the staff to know the operating status of the equipment, this embodiment further includes: if there is vibration data corresponding to the abnormal operating status of the equipment, abnormal information is sent to a designated contact person through the communication module.

[0084] It can be understood that in this embodiment, the obtained equipment operation status identification results are stored in the database. On the industrial Internet of Things platform, the contact name and email address for receiving abnormal reminders are pre-configured, and a scheduled task with a frequency of 1 minute (the same frequency as the vibration reporting) is written to monitor whether the data is in an "abnormal" state. If an abnormal state exists, an email notification of the abnormality is sent to the designated contact.

Claims

1. The device status identification method based on ESOINN is characterized by: The following steps are involved: Step 1: collecting vibration data of the device according to a first preset period, and obtaining vibration data within a plurality of first preset periods according to a second preset period; Step 2: Incrementally update the ESOINN network based on the acquired vibration data, and determine the label type corresponding to each vibration data based on the ESOINN network; Step 3: determining the device operating state corresponding to each vibration data within each second preset period according to the tag type and the neuron node data corresponding to each tag type; The step 3 specifically includes: If the label type corresponding to the vibration data is a first-class label, the corresponding vibration data is determined to be outlier data, and the device operation state corresponding to the vibration data is determined to be an abnormal state; If, among all vibration data within the second preset period, except for vibration data with a tag type of the first type, the tag type corresponding to the remaining vibration data is one, it is determined that the device operating state corresponding to the remaining vibration data is the power-on state; If, among all vibration data within a second preset period, except for vibration data with a tag type of the first type, the remaining vibration data correspond to two tag types, the neuron means of all vibration data corresponding to the two types of tags are calculated respectively, and the device operation state of all vibration data corresponding to the tag type with a larger neuron node mean is determined to be the on state, and the device operation state of all vibration data corresponding to the tag type with a smaller neuron node mean is determined to be the off state; Among all the vibration data within the second preset period, if, except for the vibration data with the tag type of the first category tag, the tag types corresponding to the remaining vibration data are greater than two, then the weight values ​​corresponding to each type of tag except the first category tag are calculated respectively, and the device operation status of all the vibration data corresponding to the tag type with a positive and largest weight value is determined to be the power-on state, the device operation status of all the vibration data corresponding to the tag type with a positive and smallest weight value is determined to be the power-off state, and the device operation status of all the vibration data corresponding to the remaining tag types is determined to be the power-off state.

2. The device status identification method based on ESOINN according to claim 1, characterized in that: The step 2 specifically includes: Obtain historical vibration data of the device and pre-train the ESOINN network based on the historical vibration data; The pre-trained ESOINN network is updated according to the acquired vibration data, and the label type corresponding to each vibration data is determined based on the pre-trained ESOINN network.

3. The device status identification method based on ESOINN according to claim 1, characterized in that: The method further includes: If there is vibration data corresponding to the abnormal operation status of the equipment, the abnormal information will be sent to the designated contact through the communication module.

4. The device status identification method based on ESOINN according to claim 1, characterized in that: The weight values ​​corresponding to each type of label except the first type of label are calculated separately, specifically including: Determine the number of neuron nodes corresponding to each type of label except the first type of label, and determine the median corresponding to the number of neuron nodes; Eliminate the label types whose number of neuron nodes is less than the median and set their weights to 0 to obtain Types of remaining tags; The number of neuron nodes and the mean amplitude corresponding to each type of remaining label are determined respectively, and the weight value corresponding to each type of remaining label is calculated according to the mean amplitude and the number of neuron nodes.

5. The device status identification method based on ESOINN according to claim 4, characterized in that: The vibration data is the vibration data of the device. 、 and The amplitude data in the direction, the calculation formula of the amplitude mean is as follows: ； Where, Indicates the Class labels in 、 and The mean amplitude in the direction, Indicates the Class label Neuron nodes in The characteristic amplitude value in the direction, Indicates the Class label Neuron nodes in The characteristic amplitude value in the direction, Indicates the Class label Neuron nodes in The characteristic amplitude value in the direction, Indicates the The number of neuron nodes with class labels.

6. The device status identification method based on ESOINN according to claim 5, characterized in that: The calculation formula of the weight value is as follows: ； in, , ; Where, Indicates the The weight value of the class label, represents the average of the amplitude means of the remaining labels of each category, Represents the average number of neuron nodes for each type of remaining label; Indicates the number of remaining tag types.

Images