Fault identification system and method

Abstract

A fault identification system and method. The system comprises an edge device and a server. The edge device can extract, from an original vibration signal, signal characteristics reflecting an abnormal vibration state of a device under test and report the signal characteristics to the server; and the server can identify a fault state of the device under test on the basis of the signal characteristics. The data volume of the signal characteristics reported in the process is small, thereby reducing the bandwidth requirement; and the signal characteristics can be used for accurate fault identification, thereby improving the accuracy of fault identification.

Description

Fault identification system and method

[0001] This application claims priority to the Chinese patent application filed with the State Intellectual Property Office of China on March 22, 2024, with application number 202410339811.8 and application name “Fault Identification System and Method”, the entire contents of which are incorporated by reference into this application. Technical Field

[0002] The embodiments of the present application relate to the technical field of equipment diagnosis, and in particular to a fault identification system and method. Background Art

[0003] Equipment condition monitoring and fault diagnosis have always been the core tasks of equipment maintenance. Although the equipment's vibration signal contains rich information about the equipment's health status, different equipment has different requirements for the frequency response coverage and sampling frequency of the vibration signal. In actual operation and maintenance, the health status information of industrial equipment is often not effectively captured due to insufficient sampling frequency, resulting in missed fault diagnosis and failure to achieve the purpose of predictive maintenance.

[0004] Therefore, the accuracy of current equipment fault identification methods is relatively low. Summary of the Invention

[0005] To solve the above technical problems, the present application provides a fault identification system and method. The method can improve the accuracy of equipment fault identification.

[0006] In one possible embodiment, the present application provides a fault identification system. The system includes a server and an edge device, wherein the edge device is connected to the device under test; the edge device is used to collect original vibration signals, wherein the original vibration signals can reflect the vibration state of the device under test; the edge device is also used to extract signal features from the original vibration signals, wherein the signal features include features that can reflect the abnormal vibration state of the device under test; the edge device is also used to send the signal features to the server; the server is used to receive the signal features from the edge device; the server is used to identify the fault state of the device under test based on the signal features.

[0007] The server can be an offline physical server or a cloud server (such as a server in a data center).

[0008] The features extracted from the vibration signal by the present invention can reflect the abnormal vibration state and effectively characterize the feature changes caused by equipment failure. In some scenarios, the feature can be unaffected by interference components and background noise, and can be effectively used to identify the fault state of the equipment. Then, the server uses the feature to identify the equipment fault, which can improve the accuracy of identifying the fault state of the equipment. In addition, in order to identify the fault state of the equipment, the edge device does not need to report a large amount of original vibration signals, but only needs to report the features extracted from the original vibration signals, which greatly reduces the amount of data and can reduce the bandwidth requirements and delay of data transmission. Moreover, the features extracted on the edge device side are not black box features, but features that can reflect the abnormal vibration state of the equipment, so that the method of the present application can be used for fault detection of a variety of equipment with similar abnormal vibration state characteristics.

[0009] In a possible implementation manner, the signal characteristics include: a frequency characteristic; the frequency characteristic includes a frequency reflecting at least one of a modulation component, a harmonic component, and a correlation component.

[0010] For example, the frequency feature may include a frequency reflecting at least one of the strongest modulation component, the strongest harmonic component, and the strongest correlation component.

[0011] In a possible implementation, the classification of the signal features includes at least one of the following: local signal features and global signal features, wherein the local signal features reflect local features of the original vibration signal, and the global signal features reflect global features of the original vibration signal.

[0012] That is, the edge device can extract local signal features and / or global signal features from the original vibration signal.

[0013] When extracting local signal features from the original vibration signal, the edge device may determine a local signal of the original vibration signal (e.g., at least one of a filtered signal and a decomposed signal), and then extract signal features from the local signal to obtain a local signal feature. The local signal feature may include the aforementioned frequency feature, specifically the frequency feature of the local signal (at least one of the filtered signal and the decomposed signal).

[0014] When extracting global signal features from the original vibration signal, the edge device may extract signal features from the original vibration signal (or the original vibration signal after processing (e.g., Hilbert transform)) to obtain global signal features. The global signal features may include the aforementioned frequency features, specifically the frequency features of the original vibration signal (or the processed original vibration signal).

[0015] In a possible implementation, the edge device is specifically configured to: determine a local signal of the original vibration signal, where the local signal includes at least one of a decomposed signal and a filtered signal; and extract signal features from the local signal as the local signal features.

[0016] In a possible implementation manner, the signal characteristics further include at least one of the following: an energy characteristic and an impact characteristic.

[0017] The energy feature may include but is not limited to at least one of the following: an autocorrelation coefficient of an envelope spectrum and an autocorrelation coefficient of an envelope signal.

[0018] For example, the calculation process of the autocorrelation coefficient of the envelope signal of the decomposed signal may be: performing an envelope operation on the decomposed signal to obtain the envelope signal; then, performing an autocorrelation operation on the envelope signal to obtain the autocorrelation coefficient of the signal.

[0019] For example, the calculation process of the autocorrelation coefficient of the envelope spectrum of the decomposed signal may be: calculating the envelope spectrum of the decomposed signal, and then performing an autocorrelation operation on the envelope spectrum to obtain the autocorrelation coefficient of the envelope spectrum.

[0020] The method of calculating the energy characteristics of the original vibration signal and the filtered signal is similar and will not be discussed here.

[0021] Impact characteristics may include, but are not limited to, kurtosis.

[0022] In one possible implementation, the server is further configured to send first information to the edge device when it identifies that the device under test is in a fault state, wherein the first information indicates that the device under test is in a fault state; and the edge device is further configured to send part of the original vibration signal to the server in response to the first information from the server.

[0023] The portion of the original vibration signal refers to a portion of the original vibration signal collected by the edge device, rather than all of the original vibration signal.

[0024] For example, this part of the original vibration signal may include the original vibration signal belonging to the signal feature used to identify the fault (referred to as the original vibration signal of the fault state), and optionally, it may also include the original vibration signals collected before and after the original vibration signal of the fault state.

[0025] Edge devices continuously collect raw vibration signals, but do not upload all of them to the server. When the server identifies a fault in the device under test based on signal characteristics, it can upload a portion of the raw vibration signal to the server for fault analysis, diagnosis, and storage. This ensures that sufficient information (such as signal characteristics) is provided for accurate fault identification without uploading high-frequency raw vibration signals that would cause bandwidth pressure.

[0026] In one possible implementation, the system further includes a client of the device under test; the server is further configured to send second information to the client when identifying that the device under test is in a fault state, the second information indicating fault information of the device under test; and the client is configured to output the fault information based on the second information.

[0027] The second information may be, for example, abnormality warning information. For example, the abnormality warning information may be a link to an interface that may include information indicating a device failure, information indicating the severity of the device failure (e.g., general failure, severe failure, etc.), which feature of the extracted signal features (also referred to as multi-dimensional features) is abnormal, the time point at which the abnormal feature occurs, a graph of the abnormal feature, and other information. This allows the device under test to output fault information, such as device abnormality data, based on the abnormality warning information.

[0028] In a possible implementation, the edge device is further configured to send part of the original vibration signal to the server when it is determined that the extracted signal feature meets a first alarm condition.

[0029] The first alarm condition may include alarm thresholds corresponding to the extracted various signal features, and the alarm thresholds may be customized by a user of the device under test.

[0030] In this way, the edge device can determine whether the device under test is in a fault state based on the extracted signal characteristics according to the alarm conditions set by the user. When the signal characteristics meet the alarm conditions, it means that the device under test may be in a fault state, thereby triggering the edge device to upload a part of the original vibration signal to the server to meet the alarm requirements set by the user.

[0031] In one possible embodiment, the signal characteristics include at least one of frequency characteristics, energy characteristics and impact characteristics; the edge device is specifically used to: determine that the signal characteristics meet the first alarm condition when it is determined that the extracted signal characteristics meet at least one of the following conditions: Condition 1, based on the frequency characteristics, determine that the signal to which the frequency characteristics belong (for example, a global signal, a local signal) has a stable periodicity; Condition 2, based on the energy characteristics, determine the proportion of the energy of the periodic modulation component of the signal to which the energy characteristics belong (for example, a global signal, a local signal) accounts for the energy of the signal, and the proportion exceeds a first threshold; Condition 3, the impact characteristic exceeds a second threshold.

[0032] In a possible implementation, the edge device is further configured to send part of the original vibration signal to the server when it is determined that the extracted signal feature meets a second alarm condition, wherein the threshold value of each signal feature in the second alarm condition is higher than the threshold value of the corresponding signal feature in the first alarm condition.

[0033] In this embodiment, it is possible to avoid the situation where the server fails to receive (or misses) some of the signal features extracted by the edge device due to poor communication between the edge device and the server, thereby making it impossible for the server to identify the device fault status based on the missed signal features and not triggering the server to send the above-mentioned first information to issue an alarm. In this embodiment, the edge device can use the higher threshold in the severe alarm condition (an example of the second alarm condition) based on the extracted signal features to implement an alarm for device anomalies, thereby avoiding the failure to detect device anomalies and improving system reliability.

[0034] In a possible implementation, the server is specifically configured to perform abnormality identification on the signal feature through a first artificial intelligence (AI) model to identify a fault state of the device under test.

[0035] For example, the first AI model may be a large model, which can effectively identify the fault status of the device under test.

[0036] In a possible implementation, the server is further configured to update the first AI model based on the original vibration signal from the edge device and the signal feature.

[0037] Since the server can not only receive the signal features extracted by the edge device, but also receive the original vibration signal when the device is in a faulty state, the server can use the accumulated data related to the device failure to expand and correct the samples to update the first AI model, so that the first AI model of the server is more adapted to the actual fault condition of the device under test, making it more accurate in fault identification.

[0038] In a possible implementation, the edge device is specifically configured to extract signal features from the original vibration signal using a second AI model.

[0039] Among them, the operation of the edge device for extracting signal features can also be implemented by a second AI model deployed on the edge device, such as a feature extraction model.

[0040] In one possible implementation, the server is further used to send third information to the edge device, where the third information includes parameters for updating the extracted signal features; the edge device is further used to receive the third information from the server and update the second AI model based on the third information, so that the type of signal features extracted by the second AI model from the original vibration signal is updated.

[0041] In an embodiment of the present application, the server can control the update of the AI ​​model used for feature extraction on the edge device side, so that the signal features extracted by it are updated. In this way, the signal features extracted by the edge device can better reflect the fault status of the device, while signal features that are poorly correlated with the fault status of the device or affect fault diagnosis do not need to be extracted or are given a smaller weight after extraction, so as to optimize the feature extraction model on the edge device side and improve the correlation between the extracted signal features and the device fault. This realizes the learning and application of the AI ​​model.

[0042] In one possible implementation, the present application also provides a fault identification method. The method includes: the edge device collecting an original vibration signal, wherein the original vibration signal can reflect the vibration state of the device under test; the edge device extracting signal features from the original vibration signal, wherein the signal features include features that can reflect the abnormal vibration state of the device under test; the edge device sending the signal features; the edge device receiving a fault state, wherein the fault state is the fault state of the device under test identified based on the signal features; and the edge device sending part of the original vibration signal in response to the fault state.

[0043] In an embodiment of the present application, the edge device can extract signal features that can reflect abnormal vibration states from the vibration signal, effectively characterize the feature changes caused by the equipment failure, and send the signal features for equipment fault identification, which can improve the accuracy of identifying the equipment's fault state. In addition, when the edge device receives the fault state, it can send a portion of the collected original vibration signal. In this way, the edge device does not need to report all the original vibration signals, but only needs to report a portion of the original vibration signal to assist in fault identification, and can reduce bandwidth pressure. Moreover, the features extracted on the edge device side are not black box features, but features that can reflect the abnormal vibration state of the equipment, so that the method of the present application can be used for fault detection of a variety of equipment with similar abnormal vibration state characteristics.

[0044] In a possible implementation manner, the signal characteristics include: a frequency characteristic; the frequency characteristic includes a frequency reflecting at least one of a modulation component, a harmonic component, and a correlation component.

[0045] In a possible implementation, the classification of the signal features includes at least one of the following: local signal features and global signal features, wherein the local signal features reflect local features of the original vibration signal, and the global signal features reflect global features of the original vibration signal.

[0046] In a possible implementation, the edge device extracts signal features from the original vibration signal, including: the edge device determines a local signal of the original vibration signal, where the local signal includes at least one of a decomposed signal and a filtered signal; and the edge device extracts signal features from the local signal as the local signal features.

[0047] In a possible implementation manner, the signal characteristics further include at least one of the following: an energy characteristic and an impact characteristic.

[0048] In a possible implementation, the method further includes: when the edge device determines that the extracted signal feature meets a first alarm condition, the edge device sends part of the original vibration signal to the server.

[0049] In one possible embodiment, the signal feature includes at least one of a frequency feature, an energy feature, and an impact feature, and the method further includes: when the edge device determines that the extracted signal feature satisfies at least one of the following conditions, determining that the signal feature satisfies the first alarm condition: determining based on the frequency feature that the signal to which the frequency feature belongs has a stable periodicity; determining based on the energy feature that the energy of the periodic modulation component of the signal to which the energy feature belongs accounts for the proportion of the energy of the signal, wherein the proportion exceeds a first threshold; and the impact feature exceeds a second threshold.

[0050] In a possible implementation, the method further includes: when the edge device determines that the extracted signal feature meets the second alarm condition, the edge device sends part of the original vibration signal to the server, wherein the threshold value of each signal feature in the second alarm condition is higher than the threshold value of the corresponding signal feature in the first alarm condition.

[0051] In a possible implementation, the edge device extracts signal features from the original vibration signal, including: the edge device extracts signal features from the original vibration signal through an AI model.

[0052] In a possible implementation, the method further includes: the edge device receiving first information from the server, and updating the second AI model based on the first information so that the type of signal features extracted by the second AI model from the original vibration signal is updated, wherein the first information includes parameters for updating the extracted signal features.

[0053] The effects of the methods of the above embodiments are similar to the effects of the fault identification systems of the above embodiments, and are not described in detail here.

[0054] In one possible embodiment, the present application provides an edge device. The edge device is configured to collect an original vibration signal, wherein the original vibration signal can reflect the vibration state of a device under test; the edge device is configured to extract signal features from the original vibration signal, wherein the signal features include features that can reflect an abnormal vibration state of the device under test; the edge device is configured to transmit the signal features; the edge device is configured to receive a fault state, wherein the fault state is a fault state of the device under test identified based on the signal features; and the edge device is configured to transmit a portion of the original vibration signal in response to the fault state.

[0055] In one possible implementation, the present application provides a computer-readable storage medium, including a computer program. When the computer program runs on an edge device, the edge device executes the method in any one of the above implementations.

[0056] The effects of the computer-readable storage medium of this embodiment are similar to the effects of the fault identification methods of the above embodiments, and are not described in detail here.

[0057] In one possible implementation, the present application provides an edge device comprising a processor and a memory, wherein the memory is coupled to the processor, and the memory stores program instructions. When the program instructions are executed by the processor, the edge device executes the method in any one of the above implementations.

[0058] The effect of the edge device in this embodiment is similar to the effect of the fault identification method in the above embodiments, and will not be repeated here.

[0059] In one possible implementation, the present application provides a computer program product, which includes a software program, and when the software program is executed by an edge device, the method in any one of the above implementations is executed.

[0060] The effects of the computer program product of this embodiment are similar to the effects of the fault identification methods of the above embodiments, and are not described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] FIG1a is a schematic diagram showing the structure of an end-cloud collaborative system;

[0062] FIG1b is a schematic diagram illustrating a process of an end-cloud collaborative anomaly recognition method;

[0063] FIG1c is a schematic diagram illustrating a process of an end-cloud collaborative anomaly recognition method;

[0064] FIG1d is a schematic diagram illustrating a process of an end-cloud collaborative anomaly recognition method;

[0065] FIG1e is a schematic diagram illustrating a process of an end-cloud collaborative anomaly recognition method;

[0066] FIG2 is a schematic diagram showing an exemplary feature extraction process;

[0067] FIG3 is a schematic diagram of an exemplary signal spectrum;

[0068] FIG4 is a schematic diagram of an exemplary signal spectrum;

[0069] FIG5 is a schematic diagram of an exemplary signal spectrum;

[0070] FIG6 is a schematic diagram showing an exemplary classification result;

[0071] FIG7 is a schematic structural diagram of an edge device provided in an embodiment of the present application;

[0072] FIG8 is a schematic diagram of the structure of a chip provided in an embodiment of the present application. DETAILED DESCRIPTION

[0073] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are 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 are within the scope of protection of this application.

[0074] The term "and / or" in this article is merely a description of the association relationship between associated objects, indicating that three relationships may exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0075] In the description and claims of the embodiments of this application, the terms "first" and "second" are used to distinguish different objects, rather than to describe a specific order of objects. For example, the terms "first target object" and "second target object" are used to distinguish different objects, rather than to describe a specific order of objects.

[0076] In the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0077] In the description of the embodiments of this application, unless otherwise specified, "multiple" means two or more. For example, "multiple processing units" means two or more processing units; "multiple systems" means two or more systems.

[0078] Before explaining the technical solution of this application, the technical names involved in the technical solution of this application are first explained.

[0079] On the device side, it refers to vibration sensors that can be deployed on industrial equipment that requires anomaly detection.

[0080] Client refers to the client installed on industrial equipment (such as electromechanical equipment) that can interact with the cloud and present the interface.

[0081] High-frequency vibration signal, a raw vibration signal sampled at a high frequency (at a frequency of 26,000 Hz or above).

[0082] The cloud platform, which may also be referred to as a cloud management platform or cloud management platform in the embodiments of the present application, is a software system for cloud technology (also known as cloud computing technology) services provided by a cloud provider, which is used to provide an interface related to cloud services for tenants to remotely access cloud services. Tenants can log in to the cloud platform on the cloud service access page using a pre-registered account and password, and after a successful login, select and purchase the corresponding cloud service on the cloud service access page, such as object storage services, virtual machine services, container services, etc. Exemplarily, the cloud management platform can provide a unified entrance (also known as an operation and maintenance entrance), and users can access the cloud management platform through the unified entrance and manage the cloud resources (including hardware and software) of the corresponding cloud through the cloud management platform.

[0083] Cloud services include computing services, storage services, virtual machine services, network services, etc. Any device or function that a user device can access on the cloud platform can be considered a service provided by the cloud platform.

[0084] Infrastructure includes multiple cloud data centers located in multiple regions, with at least one cloud data center in each region. Cloud data centers may include compute nodes, storage nodes, and network nodes. Compute nodes may be compute servers, storage nodes may be storage servers, and network nodes may include network infrastructure such as switches. For example, the hardware resources corresponding to software resources such as computing services, storage services, or network services include compute resources, storage resources, or network resources. Computing resources may be deployed on compute nodes, storage resources may be deployed on storage nodes, and network resources may be deployed on network nodes.

[0085] The server can be a physical server under the cloud or a cloud server. The cloud server can be a cloud platform, the computing node mentioned above, or a computing node and a storage node.

[0086] An edge device is a device that is closer to the device being measured (DUT) than the cloud. It can quickly collect raw vibration signals that reflect the vibration state of the DUT. For example, this edge device can be a vibration sensor or other device.

[0087] Condition monitoring and fault diagnosis of industrial equipment (such as electromechanical equipment) have always been core tasks in industrial equipment maintenance. Predictive maintenance services based on high-frequency vibration signals, in particular, are widely used in equipment operation and maintenance. Although high-frequency vibration signals contain rich information about the equipment's health status, different equipment has different requirements for the frequency response coverage and sampling frequency of high-frequency vibration signals. In actual operation and maintenance, insufficient sampling frequency often prevents the effective capture of industrial equipment health information, leading to missed faults and failure to achieve the goal of predictive maintenance.

[0088] Common vibration sensors can be divided into wired sensors with signal transmission lines and wireless sensors without. Due to the complex wiring and potential safety risks associated with wired vibration sensors, wireless vibration sensors are often used in practical applications to collect vibration signals from electromechanical equipment. These signals are then transmitted wirelessly to the cloud via a wireless gateway for signal analysis and maintenance decision-making. While wireless vibration sensors significantly reduce the complexity and safety risks of wiring, they also place higher bandwidth requirements. Vibration sensors sample vibration signals with high precision and high frequency, resulting in a large amount of sampled vibration signal data. Currently, the bandwidth of most industrial equipment in service is limited, making it difficult to meet the requirements of high-frequency sampling. This is especially true for equipment in service, such as rail transit, steel mills, and coal mines, which have relatively low bandwidth, resulting in even greater bandwidth pressure. Therefore, vibration sensors currently installed in industrial environments often sample vibration signals at low frequencies, then compress the sampled vibration signals or replace the original vibration signals with their effective value to monitor vibration signals and identify any anomalies in the industrial equipment.

[0089] In the above solution, lowering the sampling frequency of vibration signals (for example, generally to 1000Hz to 2000Hz) can reduce the bandwidth pressure of vibration signals during wireless transmission. However, low-frequency sampling of vibration signals will result in the loss of some equipment health status information. In a series of retrospective investigations of missed faults, technicians found that low-frequency sampling of vibration signals did not effectively obtain the health status information of industrial equipment, and thus could not meet the information needs of equipment anomaly identification and fault diagnosis. Therefore, there is an urgent need to increase the sampling frequency of vibration signals.

[0090] Furthermore, while data compression in the aforementioned scheme can significantly reduce the storage space required for vibration signals, thereby significantly reducing the bandwidth pressure required to transmit the compressed vibration signals, data compression inherently removes noise and low-energy components from the signal. For vibration signals, the vibration information caused by faults is not significantly dominant. Especially in the presence of strong interference, data compression can lead to the loss of information about the actual health status of the equipment, potentially leading to missed diagnosis of industrial equipment faults.

[0091] In addition, in the above scheme, the method of replacing the original vibration signal with the effective value of vibration is a stronger data compression method. The vibration information is also completely compressed into one dimension of vibration intensity. Since the vibration intensity is obviously affected by factors such as working conditions, fault diagnosis based on the effective value of vibration often leads to misdiagnosis and missed diagnosis.

[0092] Therefore, how to fully utilize the rich information of high-frequency sampled vibration signals and effectively reduce the high bandwidth requirements of a large number of vibration signals, so as to achieve high-precision fault diagnosis of electromechanical equipment through full mining of vibration information and big data analysis within limited bandwidth, is the key to improving predictive maintenance of electromechanical equipment.

[0093] Furthermore, with the increasing intelligence of the industry, predictive maintenance is becoming increasingly demanding for unmanned inspections and automated O&M scheduling of electromechanical equipment. Currently, there is an urgent need for an anomaly detection method for electromechanical equipment based on end-to-end cloud collaboration to enable fault diagnosis and data analysis.

[0094] This application provides an end-cloud collaborative method and system for identifying abnormalities in a device under test to enable fault diagnosis and analysis of industrial equipment. The device under test can be any physical device that needs to detect faults, including industrial equipment (such as electromechanical equipment) or equipment in other scenarios. Furthermore, the device under test can also be a component on a physical device, using the same method.

[0095] The following uses the device to be tested as an electromechanical device as an example to illustrate that the method and system can realize the extraction of multi-dimensional features of the original vibration signal of the electromechanical device on the end side, and the multi-dimensional features are features that can reflect the abnormal vibration state of the electromechanical device. The multi-dimensional features may include at least one of frequency features, energy features, impact features, etc. In addition, the multi-dimensional features can be features extracted based on the original vibration signal, which can reflect the overall abnormality of the original vibration signal; the multi-dimensional features can also be features extracted based on the local signal of the original vibration signal (such as at least one of the filtered signal and the decomposition signal), which can reflect the local abnormality of the original vibration signal, so as to fully mine the information in the vibration information that can reflect the abnormal vibration state of the device, and lay the foundation for the end-cloud collaborative identification of device abnormalities.

[0096] In some embodiments, the present application can also trigger the end side to report the original vibration signal collected at high frequency when it is determined that the equipment has an abnormality based on the extracted multi-dimensional features. Since there is no need to report all the original vibration signals in real time, only the original vibration signal with the abnormality needs to be reported when the equipment abnormality is identified, which can greatly relieve the bandwidth pressure.

[0097] In some embodiments, the cloud can perform high-precision identification and fault diagnosis of equipment anomalies based on the multi-dimensional features extracted on the terminal side, and can complete the update of at least one of the feature extraction model on the terminal side and the original vibration signal reporting model based on learning and using according to the analysis results of the equipment anomaly and the operation and maintenance results of the equipment anomaly on the cloud side, thereby quickly completing the scenario-based convergence of the model and achieving the purpose of quickly realizing high-precision anomaly identification of electromechanical equipment in different scenarios.

[0098] FIG1a is a schematic structural diagram of an end-cloud collaborative system of the present application.

[0099] As shown in FIG. 1 a , the system may include a mechatronic device 1000 , a gateway 200 , and a cloud 300 .

[0100] The electromechanical device 1000 may be deployed with a vibration sensor 100 , which is a wireless vibration sensor. The gateway 200 is a wireless gateway. The vibration sensor 100 may communicate wirelessly with the cloud 300 through the gateway 200 .

[0101] The connection method between the vibration sensor and the device under test (i.e., the electromechanical device 1000) can vary depending on the scenario and is not limited here. The following are some common connection methods:

[0102] 1. Direct contact installation:

[0103] Screw fixing: This is one of the most common mounting methods and is suitable for most situations. The sensor is fixed directly to the surface of the device under test with screws, ensuring good mechanical coupling for accurate transmission of vibration signals.

[0104] Magnetic mounting: For temporary measurements or where screw mounting is inconvenient, a vibration sensor with a magnetic base can be used. The magnetic base can be easily attached to the steel surface of the equipment, but may not be as stable as screw mounting.

[0105] 2. Indirect connection:

[0106] Handheld probe: For temporary testing or hard-to-reach locations, a handheld probe (handheld sensor) can be used. The operator holds the sensor close to the surface of the device under test. This method has lower accuracy and is susceptible to hand vibration.

[0107] Extension rods: In high-temperature, hazardous environments, or in inaccessible locations, extension rods can be used to mount the sensor at a distance from the equipment. This may introduce additional vibration transmission errors.

[0108] 3. Contactless:

[0109] Laser displacement sensors: For applications where direct contact is not possible, such as the shaft of rotating machinery, laser displacement sensors can be used to measure vibration from a distance. Laser sensors are sensitive to surface characteristics and are relatively expensive.

[0110] Regardless of the connection method chosen, good mechanical coupling between the sensor and the device under test is essential to minimize signal loss and interference during transmission and improve measurement accuracy. In practical applications, factors such as the device's operating environment, installation convenience, and cost must also be considered to determine the most appropriate connection method.

[0111] The cloud 300 may be a server on the cloud (eg, a computing server, also referred to as a computing node).

[0112] In existing technologies, device-side vibration sensors are wired sensors that connect to the gateway via a wired connection, which can lead to wiring difficulties such as tangling. Furthermore, wired vibration sensors require continuous signal transmission to collect high-frequency vibration signals and transmit them to the gateway. This results in high power consumption and a lifespan of approximately two years.

[0113] The vibration sensor 100 of the present application is a wireless vibration sensor that can be connected to the gateway 200 via a wireless connection (e.g., a wireless transmission protocol such as ZigBee, StarFlash, or Bluetooth) to solve the wiring difficulties of wired vibration sensors. In some scenarios, the vibration sensor 100 of the present application can communicate wirelessly with the gateway 200 via the StarFlash protocol. The StarFlash protocol is a low-power wireless communication protocol provided by the present application that can transmit large amounts of data, thereby extending the service life of the vibration sensor 100 of the present application, and generally the service life can be more than 2 years.

[0114] Of course, in some embodiments, when the wiring difficulty is relatively low, the vibration sensor 100 can also be replaced with a wired vibration sensor to perform wired data transmission through the gateway 100 .

[0115] In some embodiments, the method implemented by the cloud 300 can also be implemented by the device under test. For example, if the computing power of the device under test is high and cost is not a consideration, not only the vibration sensor of the present application but also the system of the present application can be deployed on each device under test, and the system can implement the method executed by the cloud 300 of the present application.

[0116] In some embodiments, the method implemented by the cloud 300 can also be deployed on a physical server under the cloud of the device under test, so as to perform fault diagnosis on the device under test through one or more physical servers under the cloud. The principle is the same and will not be repeated here.

[0117] In combination with FIG1a , FIG1b shows a process diagram of an anomaly identification method implemented by an end-cloud collaborative system of the present application.

[0118] As shown in Figure 1b, the vibration sensor 100 can collect high-frequency vibration signals from the electromechanical equipment 1000, and extract multi-dimensional features from the high-frequency collected vibration signals to obtain characteristic information of the high-frequency vibration signals; and transmit the characteristic information of the high-frequency vibration signals (for example, through the gateway 200 shown in Figure 1a) to the cloud 300.

[0119] The feature information may include at least one of a global feature and a local feature of the high-frequency vibration signal. Regardless of whether it is a global feature or a local feature, the feature information may be a multi-dimensional feature.

[0120] The multidimensional feature may include at least one of a frequency feature, an energy feature, and an impact feature of the vibration signal. The multidimensional feature can reflect the abnormal vibration state of the electromechanical device 1000. The energy feature reflects the energy of the abnormal vibration state, the impact feature reflects the impact of the abnormal vibration state, and the frequency feature reflects the frequency of the abnormal vibration state. The frequency feature can reflect the frequency characteristics of at least one of the following signal components: a modulation component, a harmonic component, and a correlation component.

[0121] Continuing to refer to FIG. 1 b , the cloud 300 may utilize feature information from the vibration sensor 100 to identify equipment anomalies.

[0122] In one possible implementation, when the cloud 300 identifies a device abnormality (e.g., an abnormal alarm exists) based on the characteristic information, it indicates that the characteristic information is an alarm characteristic, and the cloud 300 may send device abnormality information to the vibration sensor 100. The device abnormality information may indicate that there is an abnormality in the device (e.g., it may be "1" in binary).

[0123] In a possible implementation, upon receiving the device abnormality information, the vibration sensor 100 may send part of the original vibration signal to the cloud 300 , so that part of the original vibration signal can be reported when the device is abnormal.

[0124] The reported original vibration signal may include the original vibration signal to which the alarm feature belongs (eg, a group of original vibration signals).

[0125] In one possible implementation, within the range permitted by the bandwidth, the vibration sensor 100 can report a specified number of original vibration signals every day. Then, the part of the original vibration signals sent by the above-mentioned vibration sensor 100 to the cloud 300 is not limited to the original vibration signals belonging to the alarm characteristics, but may also include original vibration signals of other groups collected by the vibration sensor 100.

[0126] The term "group" of raw vibration signals refers to the fact that the vibration sensor can extract features from a group of raw vibration signals when extracting features. For example, the vibration sensor 100 may collect a group of raw vibration signals each time and then extract feature information from the group of raw vibration signals. However, this application does not limit whether the vibration sensor 100 extracts features based on a group of raw vibration signals or on a single or multiple raw vibration signals.

[0127] In one possible implementation, when the cloud 300 identifies a device anomaly based on the feature information, it may also send an anomaly warning message to the electromechanical device 1000 (e.g., a client of the electromechanical device 1000), so that the client can display device anomaly data based on the anomaly warning message. For example, the anomaly warning message may be a link to an interface that may include information indicating the device anomaly, information indicating the degree of the device anomaly (e.g., general anomaly, severe anomaly, etc.), which feature of the aforementioned multi-dimensional features is anomaly, at what point in time the anomaly occurred, and a graph of the anomaly features. This allows the electromechanical device 1000 to present these device anomaly data through the interface.

[0128] In the embodiment of the present application, the vibration sensor 100 still collects vibration signals at high frequency to avoid the problem of loss of equipment health status information caused by low-frequency vibration signals, ensuring the integrity of equipment health status information. Then, the vibration sensor 100 of the present application can extract multi-dimensional features that can reflect the abnormal vibration state of the equipment from the high-frequency vibration signal, realizing data conversion based on the high-frequency vibration signal. This not only retains rich vibration information caused by equipment failure, but also achieves significant compression of the large amount of high-frequency vibration signal data, for example, data compression ≥99%. When the characteristic information of the high-frequency vibration signal is transmitted to the cloud 300, the amount of transmitted data can be greatly reduced, thereby greatly reducing bandwidth pressure. In the wireless network scenario, high bandwidth is not required.

[0129] Furthermore, these multidimensional features are multidimensional, physically meaningful features strongly correlated with equipment failures, such as energy, impact, harmonics, modulation, and correlation of the original and local vibration signals. Cloud 300 can then utilize these received multidimensional features to effectively identify and diagnose equipment anomalies. When an equipment anomaly occurs, it can accurately and promptly identify the anomaly based on these multidimensional features. Furthermore, Cloud 300 can store these received multidimensional features to create a high-value feature library, facilitating anomaly identification.

[0130] Furthermore, since this multidimensional feature can reflect the abnormal vibration state of the equipment, and abnormal vibration states have similar characteristics, as long as industrial equipment with this characteristic is applicable to the abnormality detection method of this application, then the method of this application can be universally applied to abnormality identification of a class of equipment, that is, equipment for which the multidimensional feature can reflect the abnormal vibration state of the equipment when the equipment is abnormal. This makes the method and system of this application more portable and applicable not only to a specific industrial equipment, but also to a class of industrial equipment.

[0131] Furthermore, when the cloud 300 identifies a device abnormality based on the multi-dimensional features, the gateway 200 can notify the vibration sensor 100 and the electromechanical device 1000 of the device abnormality. The vibration sensor 100 can report part of the original vibration signal to the cloud 300 based on the device abnormality information. Among them, when the cloud 300 realizes abnormality recognition based on multi-dimensional features, it can be realized through an artificial intelligence (AI) model. Then, by receiving the original vibration signal reported by the vibration sensor 100, the cloud 300 can be used to train and update the AI ​​model (for example, before the AI ​​model is updated, a certain abnormality cannot be detected. After the AI ​​model is updated, the missed abnormality can be identified). On the other hand, it can also enable the cloud 300 to accumulate more abnormal asset data. Abnormal asset data refers to the original vibration signal collected when there is an equipment abnormality, because the original vibration signal when the equipment is normal is easy to obtain, but the original vibration signal when the industrial equipment has an abnormal failure (a kind of abnormal asset data) is difficult to obtain. Then, by reporting the original vibration signal of the industrial equipment when it is abnormal through the end side, it is convenient for the cloud 300 to accumulate more high-value abnormal asset data, so as to use the high-value abnormal asset data to further optimize the AI ​​model of the cloud 300.

[0132] In one possible implementation, as shown in FIG1b , within a period of time (e.g., half a year) after the vibration sensor 100 is deployed on the electromechanical device 1000, the AI ​​model (e.g., anomaly recognition model) of the cloud 300 can use the characteristic information of the high-frequency vibration signal reported by the end side to identify anomalies. When an equipment anomaly is identified, the end side is triggered to report part of the original vibration signal. Then, the cloud 300 can use the historical data accumulated within half a year after the deployment, specifically including the characteristic information and original vibration signal corresponding to the equipment anomaly, to train the AI ​​model to improve the accuracy of the AI ​​model of the cloud 300 in identifying anomalies. The AI ​​model may include an anomaly recognition model for identifying equipment anomalies based on the characteristic information. The anomaly recognition model can identify whether the equipment has an anomaly. Optionally, it may also include an anomaly type recognition model. The anomaly type recognition model can identify the type of anomaly to which the equipment anomaly belongs.

[0133] For example, if the AI ​​model deployed on the cloud 300 has not been trained, the cloud 300 can use the historical data about the electromechanical device 1000 received within half a year to generate training data. The historical data may include feature information and original vibration signals reported by the terminal side. Then, the training data is used to train the AI ​​model of the cloud 300. After the training converges, the AI ​​model can be applied to the reasoning process to identify abnormalities of the electromechanical device 1000. After the staff of the electromechanical device 1000 accepts the abnormality identification results of the cloud 300 (for example, the device abnormality information sent each time in Figure 1b), the AI ​​model after the training convergence can be put into use to realize abnormality identification reasoning of the electromechanical device 1000.

[0134] Alternatively, if the AI ​​model deployed on the cloud 300 is an AI model that has been trained in other scenarios (e.g., electromechanical equipment in another factory), then when the AI ​​model is put into use in this scenario (e.g., the scenario of electromechanical equipment 1000 shown in FIG1b ), the AI ​​model can be fine-tuned using historical data within the first six months of deploying the AI ​​model on the cloud 300 to update the AI ​​model so that the updated AI model can adapt to the identification of device anomalies in the application scenario of electromechanical equipment 1000. Then, after the staff of the electromechanical equipment 1000 have accepted the abnormality identification results of the cloud 300 (e.g., the device abnormality information sent each time in FIG1b ), the AI ​​model after the training convergence can be put into use to implement abnormality identification reasoning for the electromechanical equipment 1000.

[0135] Alternatively, the factory of the electromechanical equipment 1000 shown in Figure 1b can provide training data for the above-mentioned AI model. Then the AI ​​model deployed on the cloud 300 can use the training data for offline training. After the AI ​​model is accepted by the factory staff, it can be put online for use to realize the process shown in Figure 1b.

[0136] Among them, in each of the above scenarios, during the training process of the AI ​​model in the cloud 300, the AI ​​model can output device abnormality information when an abnormality is identified to instruct the terminal side to report the original vibration signal.

[0137] In addition, an AI model may also be deployed on the vibration sensor 100 side. The AI ​​model may include a feature extraction model and a reporting model for the original vibration signal. The feature extraction model is used to extract multidimensional features from the original vibration signal, and the reporting model is used to report the original vibration signal (to the cloud 300). Among them, the AI ​​model deployed on the vibration sensor 100 side may be a solidified AI model after training. The AI ​​model is a high-performance model. Generally, when the adaptability of the AI ​​model in the deployment scenario (such as electromechanical equipment in the high-speed rail scenario) is poor, the AI ​​model on the terminal side may also be optimized during actual application. The specific process will be described in Example 4 below.

[0138] In some embodiments, the vibration sensor 100 of the present application may include but is not limited to: an acquisition chip, an edge computing chip, a reporting chip, a transmission chip, a battery, etc.

[0139] The acquisition chip can be used to collect high-frequency raw vibration signals. The edge computing chip can be used to extract multidimensional features from the collected raw vibration signals. The reporting chip is used to report the raw vibration signals based on the reporting strategy. The transmission chip can be used to transmit the extracted multidimensional features and the raw vibration signals. The battery is used to provide power.

[0140] Example 1

[0141] The following describes the training process of the AI ​​model (e.g., the anomaly recognition model) of Cloud 300.

[0142] In combination with the above-mentioned Figures 1a and 1b, Figure 1c is a schematic diagram illustrating an exemplary workflow of the end-cloud collaboration system of the present application.

[0143] As shown in FIG. 1 c , the vibration sensor 100 may execute but is not limited to the following steps: S101 to S105 .

[0144] The cloud 300 may execute but is not limited to the following steps: S201 to S203.

[0145] Take the example of an abnormality recognition model pre-embedded in Cloud 300.

[0146] The following describes the specific implementation process of the above steps:

[0147] Optionally, S101 , collecting original vibration signals and equipment operating parameters.

[0148] S102: extracting multi-dimensional features from the original vibration signal based on the equipment operating parameters.

[0149] The multi-dimensional features may include original signal features, filtered signal features, and decomposed signal features.

[0150] In other embodiments, the device operating parameters may not be referenced when extracting multi-dimensional features.

[0151] S103: Report the extracted multi-dimensional features to the cloud 300.

[0152] S201, the cloud 300 trains an anomaly recognition model based on the received multi-dimensional features.

[0153] During the training of the anomaly recognition model, whether a device is abnormal can be identified based on multi-dimensional features.

[0154] Then, when the abnormality recognition model recognizes an abnormality based on the received multi-dimensional features, S202 may be executed.

[0155] S202, the cloud 300 may send device abnormality information.

[0156] S106 , the vibration sensor 100 reports part of the original vibration signal.

[0157] In this embodiment, when implementing S106 , the vibration sensor 100 may report part of the original vibration signal to the cloud 300 based on the device abnormality information from the cloud 300 .

[0158] Optionally, when implementing S106 , the vibration sensor 100 may determine whether to actively report the original vibration signal based on the multi-dimensional features extracted in S102 . If so, part of the original vibration signal may be reported to the cloud 300 .

[0159] In step S203, the cloud 300 can archive data based on the original vibration signal reported by the terminal side. In addition, the multi-dimensional features when anomalies are identified during the model training process in step S201 can also be archived.

[0160] Optionally, in the above S201, the cloud 300 may also use the original vibration signal and multi-dimensional features archived in S203 when the device is abnormal to train the abnormality recognition model.

[0161] The training process of the AI ​​model in the cloud 300 can adopt conventional model training methods, which are not limited here. The following example describes the training process of the AI ​​model.

[0162] Cloud 300 can add labels to the reported multidimensional features and then input the labeled multidimensional features into the AI ​​model (here, the classifier) ​​for classification. Figure 6 shows six classification results for different types of faults and normal conditions, including five types of faults (i.e., abnormalities) and normal conditions.

[0163] As shown in Figure 6, the multi-dimensional features extracted from expert knowledge can be used to improve the diagnostic accuracy of the AI ​​model on a small number of samples, thereby reducing the dependence of deep learning on the amount of data.

[0164] For example, when the target scene has only a few samples and labels, it is difficult to determine the cluster center. In this case, a feature adaptation migration diagnosis method can be used. In this method, the cluster center can be learned by using multi-dimensional features with labels from other scenes, or feature data of labels in other scenes of the target scene.

[0165] First, obtain the training set (multi-dimensional features with labels) in other scenarios as source domain data to train an AI model; then, use the training data in the target scenario (for example, the scenario of this electromechanical equipment 1000, such as the high-speed rail scenario) as the target domain data. The target domain data contains a small number of labeled samples (normal (represented by 0) or abnormal (represented by 1)) and unlabeled samples to be diagnosed, where the samples are multi-dimensional features. The labeled samples in the target domain data can be used as "adaptation anchors", and the classification function (Softmax) is used to map the target domain data to the label space (the high-dimensional space obtained by the classification function, one is the high-dimensional space corresponding to "1", and the other is the high-dimensional space corresponding to "0"); then, the samples to be diagnosed are clustered in the high-dimensional space. Specifically, the k-means clustering algorithm can be used in the feature space to make the samples to be diagnosed in the target domain data cluster towards similar adaptation key points; finally, the samples in the target domain data and the samples in the source domain data are adapted to the feature edge distribution (for example, one of the samples is multiplied by a coefficient to adapt). This adaptation can include boundary adaptation and mapping from the target domain to the source domain's label space. Furthermore, based on the label correspondence between the "adaptation anchor" and the source domain samples (matching the two sets of labels through operations such as coefficient multiplication), it guides the directional adaptation of the characteristic conditional distributions of the target and source domain samples, thereby achieving the transfer of diagnostic knowledge from the source domain to the target domain.

[0166] In this way, in the embodiment of the present application, within a period of time (e.g., half a year) after the vibration sensor 100 is deployed on the electromechanical device 1000, the AI ​​model (e.g., anomaly recognition model) on the cloud side 300 can use the characteristic information of the high-frequency vibration signal reported by the end side to identify anomalies, and when the device anomaly is identified, trigger the end side to report part of the original vibration signal. The cloud side 300 can use the historical data accumulated within half a year after the deployment of the AI ​​model (e.g., anomaly recognition model), specifically including the multi-dimensional and original vibration signals corresponding to the device anomaly, to train the AI ​​model, so as to improve the accuracy of the AI ​​model on the cloud side 300 in anomaly recognition and the matching degree of the IA model in anomaly recognition in the application scenario of the electromechanical device 1000.

[0167] The implementation principles of each process in Figure 1c in the above Example 1 are the same as the implementation principles of each process in Figure 1d in the following Example 2. Therefore, the above steps are not specifically explained and illustrated in Example 1. Please refer to the relevant instructions in Example 2 for details.

[0168] Example 2

[0169] After the AI ​​model in the cloud 300 is trained and converged through the process shown in Example 1, model inference can be performed to identify abnormalities in the electromechanical device 1000.

[0170] The following example 2 introduces the process of anomaly identification in the end-cloud collaborative system.

[0171] In the above-mentioned Example 1 and Example 2, the execution process on the vibration sensor 100 side is completely consistent in the two examples. Therefore, for S101 to S103 and S106 in Example 1, reference can be made to the detailed description of the same steps in Example 2.

[0172] In combination with Figures 1a, 1b, and 1c, Figure 1d exemplarily shows a schematic diagram of the process of anomaly identification in the end-cloud collaborative system of the present application in a reasoning scenario.

[0173] As shown in FIG1d , the process may include the following steps:

[0174] S101, the vibration sensor 100 collects original vibration signals and, optionally, equipment operating parameters.

[0175] For example, the vibration sensor 100 may collect original vibration signals at certain time intervals, for example, collecting a set of original vibration signals every 2 seconds.

[0176] This application does not impose any restrictions on the period and number of raw vibration signals collected by the vibration sensor 100.

[0177] For example, the spectrum of the original vibration signal can be seen in FIG3 .

[0178] S102 : The vibration sensor 100 extracts multi-dimensional features from the original vibration signal based on the equipment operating parameters.

[0179] The vibration sensor 100 may calculate a target frequency range based on the device operating parameters, which may reflect the frequency range in which the electromechanical device 1000 may experience a fault, that is, the electromechanical device 1000 may be prone to abnormalities and faults only within the target frequency range.

[0180] The device operating parameters and the device structural parameters are both needed to determine the target frequency range. The device structural parameters are fixed and can be input through the configuration file.

[0181] For example, device operating parameters may include operating parameters of various devices within electromechanical device 1000, such as the device's rated speed and gear position. Device structural parameters are considered manufacturing design parameters. For example, gear structural parameters may include, but are not limited to, the number of teeth, reduction ratio, input shaft speed, and output shaft speed.

[0182] Another example is a bearing. The structural parameters of a bearing may include but are not limited to: the pitch diameter of the bearing, the number of rollers, the roller diameter, etc.

[0183] The multi-dimensional features may include but are not limited to: original signal features, filtered signal features, and decomposed signal features.

[0184] The original signal feature may include multi-dimensional features (at least one of frequency features, energy features, and impact features) extracted from the original vibration signal.

[0185] The filtered signal feature is a multi-dimensional feature (at least one of a frequency feature, an energy feature, and an impact feature) extracted from the filtered signal of the original vibration signal.

[0186] The decomposition signal feature is a multi-dimensional feature (at least one of a frequency feature, an energy feature, and an impact feature) extracted from the decomposition signal of the original vibration signal.

[0187] In this embodiment, the multidimensional features reported by the client to the cloud 300 do not include the target frequency range. Based on this target frequency range, the client can extract multidimensional features from the original vibration signal. Regardless of which of the above signals (original vibration signal, filtered signal, or decomposed signal) (referred to here as the signal to be processed) is being used for multidimensional feature extraction, this can be achieved using either Method 1 or Method 2 below.

[0188] Method 1:

[0189] First, the frequency characteristics of the signal to be processed can be calculated; then, if it is determined that the frequency characteristics are within the constraint of the target frequency range, it means that the signal to be processed can reflect the fault of the electromechanical device 1000, and then at least one of the energy characteristics and the impact characteristics of the signal to be processed will be calculated (the calculation process is described in the relevant description of Figure 2). Conversely, if it is determined that the frequency characteristics are not within the constraint of the target frequency range, it means that the signal to be processed cannot reflect the fault of the electromechanical device 1000, and then the frequency characteristics and energy characteristics of the signal to be processed can be set to zero or there is no need to calculate the frequency characteristics and energy characteristics of the signal to be processed.

[0190] The frequency characteristics described in this article are expressed under the constraints of the target frequency range: at least one frequency in the frequency characteristics is within the target frequency range, and there is a frequency doubling relationship between the other frequencies in the frequency characteristics and the above-mentioned at least one frequency. The frequency doubling relationship can be a multiple relationship or a multiple relationship within a certain error.

[0191] For example, when a rotary device (such as a bearing) fails, the failure is periodic, and the derivative of the frequency can reflect the period.

[0192] Method 2:

[0193] First, the frequency, energy, and impact characteristics of the signal to be processed can be calculated separately. Then, it is determined whether the frequency characteristics are within the target frequency range. If it is determined that the frequency characteristics are within the target frequency range, the frequency, energy, and impact characteristics of the signal to be processed can all be used to identify device anomalies. Conversely, if it is determined that the frequency characteristics are not within the target frequency range, the frequency and energy characteristics of the signal to be processed can be adjusted to zero.

[0194] In either Method 1 or Method 2 above, in one possible implementation, to conserve bandwidth, if the client side determines that the frequency feature is not within the target frequency range, the client side may not report the extracted frequency feature, energy feature, or impact feature. In other words, the client side may report at least one of the extracted frequency feature, energy feature, and impact feature only if it determines that the frequency feature is within the target frequency range.

[0195] Regardless of the above-mentioned method 1 or method 2, in another possible implementation, when the bandwidth is large enough, the end side may report the extracted multi-dimensional features to the cloud 300 regardless of whether the frequency feature is within the constraints of the target frequency range, and the AI ​​model of the cloud 300 will perform abnormality identification based on the received multi-dimensional features. Then, for the multi-dimensional features in the scenario where the frequency feature is not within the constraints of the target frequency range, the AI ​​model of the cloud 300 will generally identify them as normal when identifying abnormalities, that is, it will not determine that the device has a fault. Only when the multi-dimensional features in the scenario where the frequency feature is within the constraints of the target frequency range are identified as abnormal, the AI ​​model of the cloud 300 may identify the device abnormality when identifying abnormalities, thereby triggering S302.

[0196] In another possible implementation, the multidimensional features reported by the end side to the cloud 300 may include not only at least one of the above-mentioned original signal features, filtered signal features, and decomposed signal features, but also a target frequency range (or information indicating the target frequency range, such as device operating parameters). The end side may then report the above-mentioned original signal features, filtered signal features, decomposed signal features, and target frequency range together to the cloud 300. Alternatively, the vibration sensor 100 may report the target frequency range separately to the cloud 300, which is not limited here.

[0197] Then, when the cloud 300 executes S301 as shown in Figure 1d, it can perform an inference process of abnormality identification on at least one of the original signal features, filtered signal features, and decomposed signal features based on the target frequency range to determine whether there is an abnormality in the electromechanical equipment 1000, or further include the type of abnormality.

[0198] When the AI ​​model in cloud 300 uses the target frequency range to identify whether the multidimensional features are abnormal, the AI ​​model in cloud 300 will generally identify the multidimensional features in scenarios where the frequency features are not within the target frequency range as normal, i.e., it will not determine that the device is faulty. However, only when the multidimensional features are within the target frequency range will the AI ​​model in cloud 300 identify the device as abnormal, triggering S302.

[0199] The following describes the specific implementation process of feature extraction on the client side with reference to FIG2 .

[0200] FIG2 is a schematic diagram showing an exemplary process of the vibration sensor 100 of the present application performing frequency characteristics, energy characteristics, impact characteristics, and optionally, frequency distribution characteristics and sound quality characteristics on the original vibration signal collected by the electromechanical device 1000 based on the device operating parameters.

[0201] As shown in Figure 2, the process may include the following steps:

[0202] S300: Preprocess the original vibration signal to obtain a de-averaged signal of a stable working condition.

[0203] In a possible implementation, when implementing S300, it can be implemented through the following solution, but is not limited to the following implementation solution:

[0204] The mean value x can be calculated for a set of original vibration signals (also expressed as vibration acceleration signals x(t)) collected by the vibration sensor 100. mea and root mean square (also expressed as rms x ); Then, subtract the mean value x from a set of original vibration signalsmea , to perform de-meaning processing and obtain a set of de-meaning signals.

[0205] Optionally, the vibration severity sev calculated from the de-averaged signal can be calculated based on the vibration acceleration severity calculation formula. x The vibration intensity is an energy characteristic and can also be used as one of the multi-dimensional characteristics reported to Cloud 300.

[0206] Then, according to the above-mentioned mean square amplitude rms of the electromechanical device 1000 x and its shutdown threshold value, to determine whether the electromechanical device 1000 is shut down. x When the mean square amplitude is less than the shutdown threshold, it is determined that the electromechanical device 1000 is shut down, and the corresponding set of de-averaged signals can be marked as shutdown signals; on the contrary, when the mean square amplitude rms x When it is greater than or equal to the shutdown threshold, it is determined that the electromechanical device 1000 is in the power-on state, and the corresponding group of de-averaged signals can be marked as power-on signals.

[0207] The de-meaned signal marked as the power-on signal is segmented to obtain segmented signals. The mean square amplitudes of the segmented signals are then calculated. The degree of operating stability of the electromechanical device 1000 during the sampling period is determined based on the distribution of the mean square amplitudes of the multiple segmented signals. If the distribution of the mean square amplitudes of the multiple segmented signals falls within a stable range, the de-meaned signal marked as the power-on signal can be further marked as a stable operating condition, thereby obtaining the pre-processed de-meaned signal of the stable operating condition shown in FIG2 .

[0208] On the contrary, when the distribution of the mean square amplitude of the plurality of segmented signals is not within the stable range, the de-averaged signal marked as the power-on signal may continue to be marked as an unstable working condition. The present application may not process the de-averaged signal marked as an unstable working condition.

[0209] The sampling period is the sampling period for a set of original vibration signals. The sampling period can be the time range of the sampling time or the duration of the sampling time. For example, in the example of S101 above, if a set of original vibration signals is collected every 2 seconds, the sampling period can be 2 seconds, or the sampling period for collecting the above set of original vibration signals (which also has a duration of 2 seconds).

[0210] In the embodiment of the present application, a set of original vibration signals may be sampled every 2 seconds, and multi-dimensional features may be extracted from the set of original vibration signals.

[0211] As shown in Figure 2, for the de-averaged signal under steady working conditions, the present application can perform processing in parallel on at least one of the following six aspects. Among them, when the present application performs the above six aspects of processing on the de-averaged signal under steady working conditions, it can be implemented in parallel or serially, and there is no restriction here.

[0212] Aspect 1: Extract local multi-dimensional features, where local refers to the filtered signal.

[0213] Aspect 1 may include but is not limited to: S1001 to S1004 as shown in FIG. 2 .

[0214] S1001: Filter the de-meaning signal of the steady working condition to obtain a filtered signal.

[0215] For example, n fixed-bandwidth filters may be used to perform n independent and parallel filtering on the de-averaged signal of a stationary working condition to obtain n filtered signals.

[0216] For example, a graph of the filtered signal can be seen in FIG4 .

[0217] S1002 to S1004 may be performed for each of the n filtered signals to extract multi-dimensional features of the filtered signal.

[0218] S1002: Calculate frequency characteristics of the filtered signal.

[0219] The frequency feature may include at least one of the following frequencies: the strongest harmonic frequency, the strongest modulation frequency in the envelope spectrum, and the strongest correlation frequency in the envelope signal.

[0220] Among them, the strongest harmonic frequency indicates: the frequency of the strongest harmonic component in the filtered signal;

[0221] The strongest modulation frequency indication in the envelope spectrum: After calculating the envelope spectrum of the filtered signal, the frequency of the strongest modulation component in the calculated envelope spectrum signal.

[0222] When calculating the strongest correlation frequency in the envelope signal of the filtered signal, the envelope of the filtered signal can be obtained by taking the envelope of the filtered signal and then calculating the frequency corresponding to the strongest component of the autocorrelation coefficient of the envelope signal. The strongest correlation frequency in the envelope signal indicates the frequency of the most strongly correlated component in the filtered signal.

[0223] The calculation of the autocorrelation coefficient of the envelope signal is described in S1003 and will not be repeated here.

[0224] Optionally, when executing S1002 , the frequency center of gravity of the filtered signal may be calculated to represent a frequency feature of the filtered signal.

[0225] As shown in FIG2 , the frequency characteristics of the filtered signal can be used to determine whether the filtered signal has stable periodicity.

[0226] S1003: Calculate energy characteristics of the filtered signal.

[0227] The energy feature may include at least one of the following: an autocorrelation coefficient of the envelope signal and an autocorrelation coefficient of the envelope spectrum.

[0228] For example, an envelope operation may be performed on the filtered signal to obtain an envelope signal; then, an autocorrelation operation may be performed on the envelope signal to obtain an autocorrelation coefficient of the signal.

[0229] For example, the envelope spectrum of the filtered signal is calculated, and then an autocorrelation operation is performed on the envelope spectrum, thereby obtaining an energy index of the envelope signal, which is the autocorrelation coefficient of the envelope spectrum here.

[0230] As shown in FIG2 , the energy characteristics of the filtered signal can be used to determine a first energy ratio of the periodic modulation component to the filtered signal.

[0231] Optionally, when executing S1003, the energy ratio of the frequency band energy of the filtered signal to the energy of the signal before filtering (i.e., the de-averaged signal before filtering) can also be calculated to obtain the energy ratio of the filtered signal to the energy ratio of the signal before filtering as an energy feature of the filtered signal.

[0232] The energy ratio of the filtered signal to the signal before filtering can be used to determine the strength of the energy ratio of the filtered signal to the total signal.

[0233] S1004: Calculate the impact characteristics of the filtered signal.

[0234] The kurtosis of the filtered signal may be calculated to determine the impact of the filtered signal.

[0235] Aspect 2: Extract local multi-dimensional features, where local refers to decomposed signals.

[0236] Aspect 2 may include but is not limited to: S2000 to S2004 as shown in FIG. 2 .

[0237] S2000: normalize the de-meaned signal of the steady working condition to obtain a normalized signal.

[0238] Among them, the de-mean signal of the stable working condition is normalized to obtain the normalized signal

[0239] S2001, performing singular value decomposition on the normalized signal to obtain multiple decomposition signals.

[0240] Among them, the normalized signal can be And perform singular value decomposition to obtain multiple decomposition signals n represents the dimension of the decomposed signal.

[0241] For example, the upper figure of FIG5 shows a schematic diagram of a decomposed signal, and the lower figure of FIG5 shows a schematic diagram of an envelope spectrum of the decomposed signal.

[0242] For example, S2002 to S2004 may be executed for each decomposed signal or part of the decomposed signals in the plurality of decomposed signals, and there is no limitation on the execution order of S2002 to S2004.

[0243] S2002, calculating the frequency characteristics of the decomposed signal.

[0244] The frequency feature may include at least one of the following frequencies: the strongest harmonic frequency, the strongest modulation frequency in the envelope spectrum, and the strongest correlation frequency in the envelope signal.

[0245] Among them, the strongest harmonic frequency indicates: the frequency of the strongest harmonic component in the decomposed signal;

[0246] The strongest modulation frequency indication in the envelope spectrum: After calculating the envelope spectrum of the decomposed signal, the frequency of the strongest modulated component in the calculated envelope spectrum signal.

[0247] When calculating the strongest correlation frequency in the envelope signal of the decomposed signal, the envelope of the decomposed signal can be taken to obtain its envelope signal. Then, the frequency corresponding to the strongest component of the envelope signal's autocorrelation coefficient is calculated. The strongest correlation frequency in the envelope signal indicates the frequency of the most strongly correlated component in the decomposed signal.

[0248] As shown in FIG2 , the frequency characteristics of the decomposed signal can be used to determine whether the decomposed signal has stable periodicity.

[0249] S2003: Calculate energy characteristics of the decomposed signal.

[0250] The energy feature may include at least one of the following: an autocorrelation coefficient of the envelope signal and an autocorrelation coefficient of the envelope spectrum.

[0251] For example, an envelope operation may be performed on the decomposed signal to obtain an envelope signal; then, an autocorrelation operation may be performed on the envelope signal to obtain an autocorrelation coefficient of the envelope signal.

[0252] For example, the envelope spectrum of the decomposed signal is calculated, and then an autocorrelation operation is performed on the envelope spectrum to obtain the autocorrelation coefficient of the envelope spectrum.

[0253] As shown in FIG2 , the energy characteristics of the decomposed signal can be used to determine a second energy ratio of the periodic modulation component to the decomposed signal.

[0254] Optionally, when executing S2003, the singular values ​​of the decomposed signal and the singular values ​​of the normalized signal before decomposition can also be calculated, and the ratio can be calculated to obtain the energy ratio of the decomposed signal to the signal before decomposition, as an energy feature of the decomposed signal.

[0255] The energy ratio of the decomposed signal to the signal before decomposition can be used to determine the strength of the energy ratio of the decomposed signal to the total signal, and can indicate local energy.

[0256] S2004: Calculate the impact characteristics of the decomposed signal.

[0257] The kurtosis of the decomposed signal may be calculated to determine the impact of the decomposed signal.

[0258] Aspect 3: Extracting global multi-dimensional features.

[0259] Aspect 3 may include but is not limited to: S3001 to S3004 as shown in FIG. 2 .

[0260] S3001, performing Hilbert transform on the mean-averaged signal of the steady working condition to obtain an envelope signal and an envelope spectrum signal.

[0261] S3002: Calculate frequency characteristics of the envelope signal and the envelope spectrum signal.

[0262] The frequency characteristics may include: the strongest harmonic frequency in the envelope spectrum, the strongest modulation frequency in the envelope spectrum, and the strongest correlation frequency in the envelope signal.

[0263] Among them, the strongest harmonic frequency is the frequency of the strongest harmonic component in the envelope spectrum;

[0264] The strongest modulation frequency indication in the envelope spectrum: the frequency of the strongest modulation component in the envelope spectrum signal obtained by S3001.

[0265] The strongest correlation frequency in the envelope signal is used to indicate the frequency corresponding to the strongest component of the autocorrelation coefficient of the envelope signal obtained in S3001. The strongest correlation frequency in the envelope signal indicates the frequency of the strongest correlation component in the original vibration signal.

[0266] As shown in FIG. 2 , the frequency characteristics obtained in S3002 can be used to determine whether the original vibration signal has stable periodicity.

[0267] S3003: Calculate energy characteristics of the envelope signal and the envelope spectrum signal.

[0268] The energy feature may include at least one of the following: an autocorrelation coefficient of the envelope signal and an autocorrelation coefficient of the envelope spectrum.

[0269] For example, an autocorrelation operation may be performed on the envelope signal to obtain an autocorrelation coefficient of the signal.

[0270] For example, an autocorrelation operation may be performed on the envelope spectrum (signal) to obtain the autocorrelation coefficient of the envelope spectrum.

[0271] As shown in FIG2 , the energy characteristics of the envelope signal and the envelope spectrum signal can be used to determine the third energy ratio of the periodic modulation component to the original vibration signal.

[0272] S3004: Calculate the impact characteristics of the envelope spectrum signal.

[0273] The kurtosis of the envelope spectrum signal may be calculated to obtain the kurtosis of the envelope spectrum, which is used to determine the impact of the original vibration signal.

[0274] Aspect 4: Extracting global impact features.

[0275] As shown in FIG. 2 , aspect 4 can be implemented through S4001 .

[0276] S4001 can calculate the impact characteristics of the de-averaged signal under stable working conditions.

[0277] The impact characteristics may include but are not limited to: peak-to-peak value and kurtosis value, which can be used to determine the impact of the original vibration signal.

[0278] Aspect 5: Extract global frequency distribution features.

[0279] As shown in FIG. 2 , aspect 5 can be implemented through S4003 .

[0280] S4002, the frequency distribution characteristics of the mean-valued signal under the steady working condition can be calculated.

[0281] The frequency distribution characteristics may include, but are not limited to, frequency band energy and signal skewness, and may be used to determine the energy distribution of the original vibration signal.

[0282] Aspect 6: Extracting global sound quality features.

[0283] As shown in FIG. 2 , aspect 6 can be implemented through S4003 .

[0284] S4003, the characteristics of the sound dimension can be calculated for the de-averaged signal under the steady working condition.

[0285] The characteristics of the sound dimension may include but are not limited to: loudness, roughness, and sharpness, which can be used to determine the sound quality characteristics of the original vibration signal.

[0286] In the embodiment of Figure 2, the frequency characteristics extracted by the present application for the original vibration signal, the decomposition signal of the original vibration signal, and the filtered signal of the original vibration signal can reflect at least one of the frequency of the strongest harmonic component of the corresponding signal, the frequency of the component with the strongest autocorrelation coefficient in the envelope signal of the corresponding signal, and the frequency of the component with the strongest modulation in the envelope spectrum signal of the corresponding signal. Among them, a certain frequency and its multiples, under the above-mentioned constraints of the target frequency range, generally indicate that the rotating equipment corresponding to the frequency component has an abnormal increase in vibration energy. Therefore, these frequencies found are generally related to faults. Based on this, by finding the frequency characteristics when the equipment fails abnormally, it is easy to find the abnormal equipment.

[0287] Returning to FIG. 1 d , after S102 , S103 may be executed.

[0288] S103 , the vibration sensor 100 reports the extracted multi-dimensional features to the cloud 300 .

[0289] In a possible implementation, in combination with FIG. 2 , the reported multi-dimensional features may include various features obtained as shown in FIG. 2 .

[0290] In the embodiment of FIG. 1 d , the vibration sensor 100 uploads the various multi-dimensional features obtained in FIG. 2 only when the frequency features meet the conditional constraints of the target frequency range.

[0291] With reference to the example of FIG2 , when at least one of the strongest harmonic frequency of the same filtered signal, the strongest modulation frequency in the envelope spectrum, and the strongest correlation frequency in the envelope signal is within the target frequency range, and the other frequencies are in a multiple frequency relationship with the target frequency range, it indicates that the filtered signal has stable periodicity; otherwise, it indicates that the filtered signal does not have stable periodicity.

[0292] With reference to the example of FIG2 , when at least one of the strongest harmonic frequency of the same decomposition signal, the strongest modulation frequency in the envelope spectrum, and the strongest correlation frequency in the envelope signal is within the target frequency range, and the other frequencies are in a multiple frequency relationship with the target frequency range, it indicates that the decomposition signal has stable periodicity; otherwise, it indicates that the decomposition signal does not have stable periodicity.

[0293] With reference to the example of FIG2 , when at least one of the strongest harmonic frequency obtained in S3002 corresponding to the same original vibration signal, the strongest modulation frequency in the envelope spectrum, and the strongest correlation frequency in the envelope signal is within the target frequency range, and the other frequencies are in a multiple frequency relationship with the target frequency range, it indicates that the original vibration signal has stable periodicity; otherwise, it indicates that the original vibration signal does not have stable periodicity.

[0294] S301, the cloud 300 uses the trained AI model to perform inference based on the received multi-dimensional features to identify abnormalities of the electromechanical device 1000.

[0295] For example, the AI ​​model is an abnormality recognition model, which can identify whether the device is abnormal based on the received multi-dimensional features. When it is identified that the device is abnormal, S302 is executed.

[0296] Among them, this application does not limit the model type of the abnormality recognition model of the cloud 300. Since it is trained using historical data, it is more accurate than the solution of using the alarm conditions or serious alarm conditions shown in Figure 1d on the end side to identify device abnormalities.

[0297] S302 , when the AI ​​model in the cloud 300 identifies an abnormality, device abnormality information may be sent to the vibration sensor 100 .

[0298] Among them, the specific implementation and related introduction of S302 can refer to the related introduction of the cloud 300 sending device abnormality information in Figure 1b, which will not be repeated here.

[0299] S106 , the vibration sensor 100 reports part of the original vibration signal to the cloud 300 .

[0300] The scenario that triggers S106 may be at least one of the following scenarios:

[0301] Scenario 1: Cloud 300 issues an abnormal alarm.

[0302] During S106, when the vibration sensor 100 receives the device abnormality information from the cloud 300, indicating that the cloud 300 has issued an abnormality alarm, the vibration sensor 100 may report a portion of the original vibration signal to the cloud 300. The portion of the original vibration signal may include, but is not limited to, the original vibration signal belonging to the multi-dimensional features based on which the cloud 300 identified the device abnormality.

[0303] In scenario 2, the multi-dimensional features extracted on the client side meet the user-specified alarm conditions.

[0304] For example, as shown in FIG1 d , after S102 , S104 may be executed. When the vibration sensor 100 determines that the extracted multi-dimensional features meet the alarm condition specified by the user, the vibration sensor 100 may report part of the original vibration signal to the cloud 300 .

[0305] The user may specify an alarm condition through the client, and the alarm condition may include a threshold value of at least one of the original signal feature, the filtered signal feature, and the decomposed signal feature in the multidimensional feature.

[0306] In one possible implementation, combined with S1002 to S1004 shown in Figure 2, taking the filtered signal shown in Figure 2 as an example, when it is determined that the same filtered signal has stable periodicity (for the specific judgment method, please refer to the above introduction and will not be repeated here), the first energy ratio of the periodic modulation component to the filtered signal exceeds the first threshold, the impact of the filtered signal exceeds the second threshold, and the energy ratio of the filtered signal to the signal before filtering exceeds the third threshold, if at least one of these conditions is met, it can be determined that the extracted multidimensional features meet the alarm conditions specified by the user, and the vibration sensor 100 can report part of the original vibration signal to the cloud 300.

[0307] In combination with any of the above embodiments, in a possible embodiment, combined with S2002 to S2004 shown in Figure 2, taking the decomposed signal shown in Figure 2 as an example, when it is determined that the same decomposed signal has stable periodicity (for the specific judgment method, please refer to the above introduction and will not be repeated here), the second energy ratio of the periodic modulation component to the decomposed signal exceeds the fourth threshold, the impact of the filtered signal exceeds the fifth threshold, and the energy ratio of the decomposed signal to the signal before decomposition exceeds the sixth threshold, when at least one of these conditions is met, it can be determined that the extracted multidimensional features meet the alarm conditions specified by the user, and the vibration sensor 100 can report part of the original vibration signal to the cloud 300.

[0308] In combination with any of the above embodiments, in one possible embodiment, in combination with S3002 to S3004 and S4003 shown in Figure 2, taking the original vibration signal shown in Figure 2 as an example, when it is determined that the same original vibration signal has stable periodicity (for the specific judgment method, please refer to the above introduction and will not be repeated here), the third energy ratio of the periodic modulation component to the original vibration signal exceeds the seventh threshold, the impact of the original vibration signal exceeds the eighth threshold, and the frequency distribution feature exceeds the frequency distribution feature threshold (for example, at least one of the frequency energy threshold, the signal skewness threshold, etc.), when at least one of these conditions is met, it can be determined that the extracted multidimensional feature meets the alarm condition specified by the user, and the vibration sensor 100 can report part of the original vibration signal to the cloud 300.

[0309] In some embodiments, to meet the needs of different scenarios, the alarm conditions on the terminal side may also include, but are not limited to: the above-mentioned vibration intensity threshold of the original signal, and the threshold of the sound dimension feature as shown in Figure 2 (for example, at least one of the loudness threshold, roughness threshold, and sharpness threshold). In this way, in specific scenarios, the terminal side can send an acoustic alarm to the cloud 300 as auxiliary information to supplement the fault alarm triggered by the terminal side. When the corresponding feature exceeds the threshold of the feature, an alarm is triggered to report the original vibration signal.

[0310] Optionally, in scenario 3, the multi-dimensional features extracted on the terminal side meet serious alarm conditions.

[0311] For example, as shown in FIG1 d , after S102 , S105 may be executed. When the vibration sensor 100 determines that the extracted multi-dimensional features meet the serious alarm condition, the vibration sensor 100 may report part of the original vibration signal to the cloud 300 .

[0312] Among them, the serious alarm condition may include a threshold value of at least one feature among the original signal feature, the filtered signal feature, and the decomposed signal feature in the multi-dimensional feature.

[0313] The principles of the implementation process of this scenario 3 are the same as those of scenario 2. Please refer to the description of scenario 2 for details. The difference is that for the same feature, the threshold of the corresponding feature in the severe alarm condition is greater than the threshold of the feature in the above-mentioned user-guided alarm condition.

[0314] The target frequency range for determining the periodicity of signal stability is the same, but the thresholds used to determine other features are different.

[0315] For example, the threshold for evaluating the first energy ratio of the periodic modulation component to the filtered signal is higher than the first threshold in the severe alarm condition. The other thresholds are similar and will not be described in detail here.

[0316] In this scenario 3, poor end-cloud communication can prevent the cloud 300 from not receiving the multi-dimensional features related to the fault extracted by the end-side, making it impossible to identify whether the device is abnormal on the cloud 300 side and thus not triggering an alarm on the cloud 300 (for example, sending device abnormality information through S302). Therefore, when end-cloud communication is poor, the end-side can use the extracted multi-dimensional features and a higher threshold in the severe alarm condition to issue an alarm for device abnormalities, thereby preventing device abnormalities from going undetected and improving system reliability.

[0317] The functions of Scene 3 can be enabled or disabled, and can be flexibly enabled or disabled based on user needs.

[0318] Optionally, in scenario 4, the client side of the electromechanical device 1000 receives an input from the user to adjust the number of original vibration signals reported by the client side.

[0319] The vibration sensor 100 may adjust the number of reported raw vibration signals based on a user's input for adjusting the number of reported raw vibration signals.

[0320] For example, the current system is configured to collect 100 sets of raw vibration signals per day on the client side and report 10 sets of raw vibration signals per day. If the user determines that these 10 sets of raw vibration signals are insufficient or excessive, they can increase or decrease this number, for example, to 20 sets. In this way, the client side can report raw vibration signals in addition to scenarios 1 to 3, so that the total number of raw vibration signals reported per day is 20 sets. For example, if scenarios 1 to 3 report only 15 sets of raw vibration signals on a given day, the client side can continue to report 5 sets of raw vibration signals even if scenarios 1 to 3 are not triggered.

[0321] For example, in a coal mine scenario, the preset reporting of raw vibration signals is two groups. The user adjusts the input to indicate that not only the abnormal raw vibration signal should be reported, but also at least one group of raw vibration signals received by the end-side before or after the abnormal raw vibration signal, and specifies the number of groups of raw vibration signals to be reported (a total of five groups). In this way, the end-side can adjust the reported raw vibration signals according to the user's adjustment input. When the reporting of raw vibration signals is implemented by a reporting model, the reporting model can be updated based on the user's adjustment input.

[0322] In addition, the time when the vibration sensor 100 reports the original vibration signal can be different from the time when it reports the extracted multidimensional features. For example, the original vibration signal can be reported in the time gap between the reporting time of the multidimensional features to reduce bandwidth usage. Of course, if the bandwidth is sufficient, the original vibration signal can also be uploaded simultaneously with the multidimensional features. This is not a restriction here.

[0323] In step S303, the cloud 300 can archive data based on the original vibration signal reported by the terminal. In addition, the multi-dimensional features of the abnormality identified during the model inference process in step S301 can also be archived.

[0324] The specific implementation of data archiving may be to save the data, for example, in the form of a file on a disk, etc. There is no restriction on the saving method.

[0325] Among them, the cloud 300 can archive the multi-dimensional features reported by the end side. In addition, it can also archive the original vibration signal reported by the end side.

[0326] In addition, the cloud 300 can also add labels to the multi-dimensional features of the abnormalities associated with the reported original vibration signal.

[0327] Optionally, the cloud 300 may also store data distribution of original vibration signals with alarms within a period of time and data distribution of multi-dimensional features.

[0328] In some scenarios, a device may have multiple vibration sensors installed at different locations. In order to facilitate tracing the location (also referred to as the installation measurement point) to which the data stored in the cloud 300 belongs, the vibration sensor may report the location information to the cloud 300, so that the operation and maintenance personnel can observe the changing trends of the historical characteristics of the detected equipment (such as the above-mentioned electromechanical equipment 1000) at the installation measurement points of each sensor, which is convenient for customers to trace back.

[0329] Optionally, the operation and maintenance personnel may correct the archived data, for example, correct the labels of missed reports (i.e., missed alarms) and false reports (i.e., false alarms), so that the archived data in the cloud 300 is updated.

[0330] For example, if the cloud 300 identifies a multidimensional feature as abnormal, or the end side issues an alarm for certain multidimensional feature detections (for example, the above-mentioned alarm conditions or severe alarm conditions are met), which is an abnormal situation, the cloud 300 can archive the multidimensional features with abnormalities and add abnormal labels to the multidimensional features. However, due to the possible problems of false positives and missed positives, the multidimensional features and the abnormal labels are not necessarily compatible. Therefore, the labels of the multidimensional features of the false positives can be corrected, for example, from abnormal labels to normal labels, or from a certain abnormal label to another abnormal label (that is, the abnormal problems corresponding to the two abnormal labels are different). In addition, abnormal labels can be added to the multidimensional features corresponding to the abnormalities with missed positives.

[0331] In the embodiment of the present application, the data archived in the cloud 300 may include multidimensional features, the AI ​​model of the cloud 300 and its model parameters, the multidimensional features of false alarms and their labels and their corresponding original vibration signals, the multidimensional features of missed alarms and their labels and their corresponding original vibration signals, the data distribution of the original vibration signals with the above-mentioned abnormalities (also expressed as alarms), the data distribution of the multidimensional features with the above-mentioned abnormalities (also expressed as alarms), etc.

[0332] The embodiment of the present application realizes the calculation and reporting of the multi-dimensional features of the original vibration signal through the vibration sensor side (an example of an edge device), and the cloud can obtain rich device status information without the need to transmit a large amount of original signals. The multi-dimensional features may include the overall features of the high-frequency vibration signal (the overall energy, impact, frequency of the original vibration signal, and optionally, sound quality, frequency distribution and other features) and local features (local energy, impact, and frequency of the decomposed signal under fixed bandwidth filtering and decomposition noise reduction), thereby replacing the transmission of the original signal and reducing the bandwidth pressure of the acquisition and transmission of high-frequency vibration signals. In addition, the multi-dimensional features may contain rich device health information, which can replace the original signal to ensure analysis accuracy and greatly reduce bandwidth. And the multi-dimensional features may include the overall energy, impact, harmonics, modulation, sound quality, frequency distribution and other features of the original signal, and the local energy, impact, harmonics, modulation, frequency distribution and other features under fixed bandwidth filtering and signal decomposition noise reduction, which have stronger robustness.

[0333] In addition, the end side can selectively report the original vibration signal in the event of an alarm or severe alarm, as well as selectively report the original vibration signal when an anomaly is identified in the cloud. By controlling the selective reporting of the original vibration signal through end-cloud collaboration, a large amount of data features is transmitted, and part of the original signal is reported for error tracing and model updating.

[0334] Moreover, the cloud-based anomaly recognition and fault diagnosis functions are implemented based on high-value multi-dimensional features. The model input is not the original vibration signal, nor the features extracted by the black box, but the specified multi-dimensional features. This makes the cloud-based AI model have strong migration capabilities and the ability to quickly trace back diagnostic errors, achieving scenario-based high-precision convergence of the model and applicable to many scenarios.

[0335] Example 3

[0336] FIG1e exemplarily shows a schematic diagram of the learning-while-using process of the AI ​​model of the present application.

[0337] Compared to Figures 1d and 1e, Example 3 and Example 2 share most of the same content. The difference is that Example 3 adds that the cloud 300 can use archived data to update the AI ​​model on the cloud 300 side. In addition, after the AI ​​model on the cloud 300 side is updated, the cloud 300 can optionally update the feature extraction model used for feature extraction on the device side. Optionally, the cloud 300 can also update the reporting model used for reporting raw vibration signals on the device side.

[0338] In this embodiment, not only can the training and reasoning of the AI ​​model of the cloud 300 be completed through the cloud 300, but the AI ​​model of the cloud 300 can also be optimized based on the data archived by the cloud 300. Optionally, the feature extraction model on the end side and the reporting model of the original vibration signal can also be updated to ensure that high-value data (such as abnormal multi-dimensional features and original vibration signals) are maximized when bandwidth permits.

[0339] The implementation process of S101 to S106 and S301 to S303 shown in FIG1e in Example 3 is the same as the description of FIG1e in the above Example 2, and will not be repeated here.

[0340] As shown in FIG. 1e , the process may include S101 to S106 and S301 to S303 . For detailed description, please refer to Example 2 and will not be repeated here.

[0341] As shown in FIG. 1e , after S303 , S304 may be further included.

[0342] S304, the cloud 300 uses the archived data to update the AI ​​model of the cloud 300.

[0343] Specifically, taking the AI ​​model including the anomaly recognition model as an example, the cloud 300 can use the archived data after the AI ​​model has run for a period of time to update the AI ​​model, wherein the archived data can refer to the specific introduction of S303 described in Figure 1e above.

[0344] For example, the cloud 300 can use the AI ​​model to identify the multi-dimensional features of the anomaly and the corresponding original vibration signal reported by the end side, the original vibration signal reported by the end side due to an alarm or a serious alarm and the corresponding multi-dimensional features to add labels to the data with anomalies within a period of time; in addition, the abnormal data can be increased or decreased and the labels can be corrected in combination with the false positives and missed positives. In this way, historical data (multi-dimensional features, original vibration signals) that can reflect the abnormality of the equipment can be obtained. Then, the abnormal historical data with labels is used to update the abnormal recognition model of the cloud 300, so that the abnormal recognition model can more accurately identify the abnormality of the electromechanical equipment 1000 after optimization, making it more suitable for the abnormal recognition of the electromechanical equipment 1000.

[0345] In addition, the AI ​​model may also include a diagnostic model for identifying the type of abnormality. For example, the above-mentioned abnormality recognition model can only identify whether there is an abnormality, while the diagnostic model can identify the type of abnormality.

[0346] After a large number of multi-dimensional features and original vibration signals are archived in the cloud 300, the corrected historical data reflecting device anomalies can be used to update the diagnostic model. This allows the updated diagnostic model to achieve scenario-based convergence in the context of the electromechanical device 1000, making it more suitable for identifying anomaly types in the electromechanical device 1000. Furthermore, the anomaly types identified by the diagnostic model and their corresponding historical data can also be fully displayed and archived in S303 for the output display of high-value data and efficient data backtracking.

[0347] In addition, in some embodiments, as described above, the features reported by the device side may include a target frequency range determined based on device operating parameters. However, this target frequency range is not necessarily very accurate. Therefore, the AI ​​model in the cloud 300 can be optimized based on the accumulated historical data, so that the target frequency range used by the AI ​​model in the cloud 300 for anomaly identification can be adjusted to match the frequency range where actual device anomalies occur.

[0348] In addition, in some embodiments, the cloud 300 may also analyze the main source of error of the AI ​​model of the cloud 300 based on the information of false positives and false negatives. The main source may be a multi-dimensional feature or a model parameter of the AI ​​model.

[0349] For example, if the label of a multidimensional feature is normal, but the result obtained by the abnormality recognition model when analyzing the multidimensional feature is abnormal, then it is possible to analyze which target feature in the multidimensional feature causes the error of the AI ​​model, so that when the multidimensional feature is input into the AI ​​model, the target feature is removed (that is, the multidimensional feature input of the AI ​​model no longer includes the target feature), or a smaller weight is set for the target feature in the multidimensional feature, and then it is input into the AI ​​model for recognition. In this way, the update of the AI ​​model can lead to changes in the input features of the AI ​​model. For example, when a new feature is introduced, the feature extraction model on the end side needs to be updated based on this (see the following introduction).

[0350] For another example, if the main source is the model parameters of the AI ​​model, the model parameters of the AI ​​model can be optimized based on the corrected historical data.

[0351] In an embodiment of the present application, the diagnostic accuracy of different application sites can be fed back to analyze the main sources of errors. By reproducing the feature weights under different results, the degree of overfitting of the classifier (an example of an AI model) and the contribution rate of different features to equipment abnormality identification and fault type can be analyzed. For features or classifiers that are prone to misdiagnosis or missed diagnosis, experts can develop new fault characterization features or AI models based on reported fault data or falsely reported normal data, and form a high-value knowledge graph that is learned and used at the same time. The update of the feature extraction model on the terminal side is triggered by the cloud 300 to achieve rapid adaptation and high-precision monitoring in different scenarios.

[0352] Optionally, the feature information reported by the terminal side to the cloud 300 may also include device installation parameters of the electromechanical device 1000 .

[0353] The device installation parameters may include height from the ground, fixed installation or magnetic installation, and other installation method information, etc., which are related to the device installation situation.

[0354] The user can input the device installation parameters through the client on the electromechanical device 1000 side.

[0355] The client can report the device installation parameters to the cloud, for example, by uploading them when reporting features. There is no restriction here.

[0356] Then the cloud 300 can archive the device installation parameters in the above S303

[0357] When the AI ​​model in cloud 300 identifies an anomaly, it can associate and store the historical data associated with the device installation parameters. When the administrator of electromechanical device 1000 logs into cloud 300 using their account to access the data archived there, they can input the device installation parameters, causing cloud 300 to display the archived data associated with those parameters, thereby tracing the source of the anomaly.

[0358] In addition, in the event of abnormal false alarms and missed alarms, the cloud 300 can also optimize and adjust the AI ​​model of the cloud 300 based on the device installation parameters when executing S304.

[0359] Optionally, after S304, S305 may be further included.

[0360] S305: The cloud 300 updates the feature extraction model on the terminal side based on the update of the AI ​​model.

[0361] Optionally, after S304, S306 may be further included.

[0362] S306: The cloud 300 updates the reporting model on the terminal side based on the update of the AI ​​model.

[0363] The following is an introduction to the principle of S305.

[0364] After the AI ​​model on the cloud 300 is updated, the updated AI model needs to introduce new features. In other words, the updated AI model needs to further identify anomalies based on the new features. If the multidimensional features on which the AI ​​model identifies anomalies are increased, the cloud 300 can instruct the feature extraction model on the end side to extract new features, so as to not only report the original multidimensional features but also add new features to the reported multidimensional features to meet the anomaly identification requirements of the AI ​​model on the cloud 300.

[0365] The principle of S306 is introduced below.

[0366] When the multi-dimensional features extracted by the feature extraction model on the terminal side change, the reporting mechanism of the original data of the reporting model also needs to be updated. Therefore, the reporting model also needs to be updated.

[0367] For example, if the multidimensional features only include filtered signal features and decomposed signal features, but not original signal features, then there is no need to set alarm conditions and severe alarm conditions regarding the original signal features in the reporting model, so that the reporting model can be updated.

[0368] Regarding the way in which the cloud 300 updates the two models on the end side, the updated model parameters can be sent to the end side, or an updated feature extraction model and an updated reporting model can be generated, and the two updated models can be sent to the end side for replacement. There is no restriction here.

[0369] On the contrary, if the update of the AI ​​model of the cloud 300 does not introduce new features and does not cause changes in the reporting mechanism of the original vibration signal, there is no need to trigger S305 and S306 to update the terminal side model.

[0370] In an embodiment of the present application, multi-dimensional high-value features can be used in the cloud 300 to complete the training of a transferable anomaly recognition model and the operation and maintenance analysis of the anomaly recognition results, and based on the corrected historical data of the cloud 300, at least one of the end-side feature extraction model and the original vibration signal reporting model can be updated.

[0371] Based on the multi-dimensional features and historical data reported by the end-side, the training and inference of the portable AI model are completed in the cloud, and the abnormal alarm results are output. According to the cloud-side alarm results, partial reporting of the original vibration signal on the end-side is triggered to accumulate more data assets and complete further optimization of the AI ​​model on the cloud. The cloud side not only has the communication capabilities of device anomaly identification, end-side model update, and original data reporting, but also can design a reasonable partial original data reporting mechanism according to the needs of cloud-side analysis to support model updates and error tracing.

[0372] The cloud 300 and end-side models of this application can be learned and used at the same time, which can realize the update of the end-side feature extraction model and the reporting model of high-frequency vibration signals. Moreover, by training the AI ​​model of the cloud 300 in different scenarios, intelligent alarm and diagnosis can be realized, and the scenario-based convergence and accumulation of the model can be completed. In addition, the cloud 300 can archive key (such as abnormal) multi-dimensional features for high-value presentation and efficient backtracking of data. It can quickly adapt to different devices and form high-value knowledge graphs for different faults in different scenarios. It has strong migration capabilities and is applicable to many scenarios.

[0373] The following describes an edge device provided by an embodiment of the present application. As shown in Figure 7:

[0374] FIG7 is a schematic diagram of the structure of an edge device provided in an embodiment of the present application. As shown in FIG7 , the edge device 500 may include: a processor 501 , optionally a transceiver 505 , and optionally a memory 502 .

[0375] The transceiver 505 may be referred to as a transceiver unit, a transceiver, or a transceiver circuit, etc., and is configured to implement transceiver functions. The transceiver 505 may include a receiver and a transmitter. The receiver may be referred to as a receiver or a receiving circuit, etc., and is configured to implement a receiving function; the transmitter may be referred to as a transmitter or a transmitting circuit, etc., and is configured to implement a transmitting function.

[0376] The transceiver 505 may also be a communication interface.

[0377] The memory 502 may store a computer program or software code or instruction 504, which may also be referred to as firmware. The processor 501 may implement the fault identification method provided in each embodiment of the present application by running the computer program or software code or instruction 503 therein, or by calling the computer program or software code or instruction 504 stored in the memory 502. The processor 501 may be a central processing unit (CPU), and the memory 502 may be, for example, a read-only memory (ROM) or a random access memory (RAM).

[0378] The processor 501 and transceiver 505 described in this application can be implemented on an integrated circuit (IC), an analog IC, a radio frequency integrated circuit RFIC, a mixed signal IC, an application specific integrated circuit (ASIC), a printed circuit board (PCB), an electronic device, etc.

[0379] The modules included in the edge device 500 are only examples and are not limited in this application.

[0380] For example, the structure of the edge device may not be limited to FIG7. The edge device may be an independent device or may be part of a larger device. For example, the implementation form of the edge device may be:

[0381] (1) An independent integrated circuit IC, or chip, or chip system or subsystem; (2) A collection of one or more ICs, optionally including a storage component for storing data and instructions; (3) A module that can be embedded in other devices; (4) In-vehicle equipment, etc.; (5) Others, etc.

[0382] For edge devices implemented as chips or chip systems, see the schematic diagram of the chip structure shown in Figure 8. The chip shown in Figure 8 includes a processor 601 and an interface 602. There may be one or more processors 601, and there may be multiple interfaces 602. Optionally, the chip or chip system may include a memory 603. The processor 601 is configured to call and run instructions from the interface 602. When the processor 601 executes the instructions, the steps of the above-described method embodiment may be performed.

[0383] Among them, all relevant contents of each step involved in the above method embodiment can be referred to the functional description of the corresponding functional module and will not be repeated here.

[0384] Based on the same technical concept, an embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. The computer program includes at least one section of code, which can be executed by an edge device to control the edge device to implement the above method embodiment.

[0385] Based on the same technical concept, an embodiment of the present application also provides a computer program, which, when executed, is used to implement the above method embodiment.

[0386] The program may be stored in whole or in part on a storage medium packaged with the processor, or may be stored in whole or in part on a memory not packaged with the processor.

[0387] Based on the same technical concept, an embodiment of the present application further provides a chip including a processor. The processor can implement the above method embodiment.

[0388] The steps of the method or algorithm described in conjunction with the disclosure of the embodiments of the present application can be implemented in a hardware manner, or can be implemented by a processor executing a software instruction. The software instruction can be composed of corresponding software modules, and the software module can be stored in a random access memory (Random Access Memory, RAM), a flash memory, a read-only memory (Read Only Memory, ROM), an erasable programmable read-only memory (Erasable Programmable ROM, EPROM), an electrically erasable programmable read-only memory (Electrically EPROM, EEPROM), a register, a hard disk, a mobile hard disk, a read-only compact disc (CD-ROM) or any other form of storage medium well known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and can write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an ASIC.

[0389] Those skilled in the art will appreciate that in one or more of the above examples, the functions described in the embodiments of the present application can be implemented using hardware, software, firmware, or any combination thereof. When implemented using software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any media that facilitates the transmission of computer programs from one place to another. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0390] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the protection scope of the technical solutions of the embodiments of the present application.

Claims

1. A fault identification system, characterized in that: The system includes a server and an edge device, wherein the edge device is connected to the device under test; The edge device is used to collect an original vibration signal, wherein the original vibration signal can reflect the vibration state of the device under test; The edge device is further configured to extract signal features from the original vibration signal, wherein the signal features include features that can reflect an abnormal vibration state of the device under test; The edge device is further configured to send the signal characteristics to the server; The server is configured to receive the signal characteristics from the edge device; The server is configured to identify a fault state of the device under test based on the signal characteristics.

2. The system according to claim 1, wherein: The signal characteristics include: frequency characteristics; The frequency characteristics include a frequency reflecting at least one of a modulation component, a harmonic component, and a correlation component.

3. The system according to claim 2, characterized in that The classification of the signal features includes at least one of the following: local signal features and global signal features, wherein the local signal features reflect local features of the original vibration signal, and the global signal features reflect global features of the original vibration signal.

4. The system according to claim 3, characterized in that The edge device is specifically used for: determining a local signal of the original vibration signal, the local signal comprising at least one of a decomposed signal and a filtered signal; A signal feature is extracted from the local signal to serve as the local signal feature.

5. The system according to any one of claims 1 to 4, characterized in that The signal characteristics further include at least one of the following: energy characteristics and impact characteristics.

6. The system according to any one of claims 1 to 5, characterized in that The server is further configured to send first information to the edge device when identifying that the device under test is in a fault state, wherein the first information indicates that the device under test is in a fault state; The edge device is further configured to send part of the original vibration signal to the server in response to the first information from the server.

7. The system according to any one of claims 1 to 6, characterized in that The system further includes a client of the device under test; The server is further configured to send second information to the client when identifying that the device under test is in a fault state, where the second information indicates fault information of the device under test; The client is configured to output the fault information based on the second information.

8. The system according to any one of claims 1 to 7, characterized in that The edge device is further configured to send part of the original vibration signal to the server when it is determined that the extracted signal feature meets a first alarm condition.

9. The system according to claim 8, characterized in that The signal characteristics include at least one of frequency characteristics, energy characteristics and impact characteristics; The edge device is specifically used for: When it is determined that the extracted signal feature satisfies at least one of the following conditions, it is determined that the signal feature satisfies the first alarm condition: Determining, based on the frequency feature, that the signal to which the frequency feature belongs has stable periodicity; determining, based on the energy feature, a ratio of the energy of the periodic modulation component of the signal to which the energy feature belongs to the energy of the signal, wherein the ratio exceeds a first threshold; The impact characteristic exceeds a second threshold.

10. The system according to claim 8 or 9, characterized in that The edge device is also used to send part of the original vibration signal to the server when it is determined that the extracted signal feature meets the second alarm condition, wherein the threshold value of each signal feature in the second alarm condition is higher than the threshold value of the corresponding signal feature in the first alarm condition.

11. The system according to any one of claims 6 to 10, characterized in that The server is specifically used to identify abnormalities of the signal characteristics through a first artificial intelligence (AI) model to identify the fault status of the device under test.

12. The system according to claim 11, wherein: The server is further configured to update the first AI model based on the original vibration signal from the edge device and the signal feature.

13. The system according to claim 12, wherein: The edge device is specifically used to extract signal features from the original vibration signal through a second AI model.

14. The system according to claim 13, wherein: The server is further configured to send third information to the edge device, where the third information includes parameters for updating the extracted signal features; The edge device is further configured to receive the third information from the server and update the second AI model based on the third information, so that the type of signal features extracted by the second AI model from the original vibration signal is updated.

15. A fault identification method, characterized in that: The method comprises: The edge device collects the original vibration signal, wherein the original vibration signal can reflect the vibration state of the device under test; The edge device extracts signal features from the original vibration signal, where the signal features include features that can reflect an abnormal vibration state of the device under test; The edge device sends the signal characteristic; The edge device receives a fault status, where the fault status is the fault status of the device under test identified based on the signal feature; The edge device transmits a portion of the original vibration signal in response to the fault condition.

16. A computer-readable storage medium, characterized in that The method comprises a computer program, which, when running on an edge device, causes the edge device to perform the method according to claim 15 .

17. An edge device, characterized in that: The edge device comprises a processor and a memory, wherein the memory is coupled to the processor and stores program instructions. When the program instructions are executed by the processor, the edge device executes the method according to claim 15.

18. A computer program product, characterized in that The computer program product comprises a software program which, when executed by an edge device, causes the steps of the method of claim 15 to be performed.

Images