CN113537288B

CN113537288B - Method and system for training target model based on modification of sample signal

Info

Publication number: CN113537288B
Application number: CN202110663854.8A
Authority: CN
Inventors: 郭春林; 郭尔富
Original assignee: Beijing Earth Cross High Technology Co ltd; North China Electric Power University
Current assignee: Beijing Earth Cross High Technology Co ltd; North China Electric Power University
Priority date: 2021-06-16
Filing date: 2021-06-16
Publication date: 2024-04-05
Anticipated expiration: 2041-06-16
Also published as: CN113537288A; WO2022262072A1

Abstract

The invention provides a method and a system for training a target model based on modification of a sample signal, wherein the method comprises the following steps: selecting a target model to be trained from a plurality of models, and acquiring a configuration file of the target model; determining an object device to which the target model relates based on a profile of the target model, and determining a plurality of sample signals associated with the object device; when the object equipment normally operates, signal acquisition or signal simulation is carried out on at least two sample signals in a plurality of sample signals, so that a normal sample set comprising the at least two sample signals is obtained; modifying at least one sample signal in the normal sample set according to a preset modification mode to obtain an abnormal sample set corresponding to the normal sample set; training the target model based on the normal sample set, the abnormal sample set and a preset training algorithm, so as to obtain a trained target model.

Description

Method and system for training target model based on modification of sample signal

Technical Field

The present invention relates to the field of artificial intelligence, and more particularly, to a method and system for training a target model based on modification of a sample signal.

Background

At present, with the development of artificial intelligence technology, a great number of machine learning algorithms are continuously emerging. Machine learning algorithms, and in particular deep learning, have achieved tremendous success in recent years, and data is the key factor that enables machine learning. The technician can implement machine learning using a simple algorithm, but without good data, the algorithm cannot be optimized.

In the model training based on machine learning, it is known that the data quality of the sample signal/sample data affects the training effect of the model. However, in actual cases, the number or rate of malfunctions occurring in operation of some types of devices is low, and thus the data amount of the sample signal/sample data of normal operation of this type of devices is relatively large, and the data amount of the sample signal/sample data at the time of abnormal operation or malfunction is small. In this case, the abnormal sample signal/sample data of the smaller data amount cannot meet the requirements of model training or testing.

Disclosure of Invention

In order to solve the above problems in the prior art, the present application proposes a model training method and system based on signal modification. The model training method and system based on signal modification are applicable to various preset algorithms, including machine learning algorithms, fitting algorithms and the like.

According to one aspect of the present invention, there is provided a method of training a target model based on modification of a sample signal, the method comprising:

selecting a target model to be trained from a plurality of models, and acquiring a configuration file of the target model;

determining an object device to which the target model relates based on a profile of the target model, and determining a plurality of sample signals associated with the object device;

when the object equipment normally operates, signal acquisition or signal simulation is carried out on at least two sample signals in a plurality of sample signals, so that a normal sample set comprising the at least two sample signals is obtained;

modifying at least one sample signal in the normal sample set according to a preset modification mode to obtain an abnormal sample set corresponding to the normal sample set;

training the target model based on the normal sample set, the abnormal sample set and a preset training algorithm, so as to obtain a trained target model.

The plurality of sample signals includes: vibration, sound, speed, displacement, stress, pressure, voltage, current, power, electric field strength, magnetic field strength, temperature, image, and brightness.

At least one of the at least two sample signals is a vibration/acoustic emission signal acquired by a sensor attached to the housing of the device.

At least one sample signal of the at least two sample signals is a sound signal collected outside the device.

According to yet another aspect of the present invention, there is provided a method of testing a target model based on modification of a sample signal, the method comprising:

determining a target model to be tested and obtaining a configuration file of the target model;

And testing the target model based on the normal sample set and the abnormal sample set, so as to determine the performance index of the target model based on the test result.

According to yet another aspect of the present invention, there is provided a system for training a target model based on modification of a sample signal, the system comprising:

the selecting device is used for selecting a target model to be trained from a plurality of models and acquiring a configuration file of the target model;

determining means for determining an object device to which the target model relates based on a profile of the target model, and determining a plurality of sample signals associated with the object device;

the acquisition device is used for carrying out signal acquisition or signal simulation on at least two sample signals in a plurality of sample signals when the object equipment normally operates, so as to acquire a normal sample set comprising the at least two sample signals;

the modification device is used for modifying at least one sample signal in the normal sample set according to a preset modification mode so as to obtain an abnormal sample set corresponding to the normal sample set;

the training device is used for training the target model based on the normal sample set, the abnormal sample set and a preset training algorithm, so that a trained target model is obtained.

According to yet another aspect of the present invention, there is provided a system for testing a target model based on modification of a sample signal, the system comprising:

the obtaining device is used for determining a target model to be tested and obtaining a configuration file of the target model;

and the testing device is used for testing the target model based on the normal sample set and the abnormal sample set, so that the performance index of the target model is determined based on the testing result.

According to still another aspect of the present invention, there is provided a model training method based on signal modification, which first generates a large number of normal samples and abnormal samples, and then trains a model using the samples and a set algorithm for diagnosing, discriminating or identifying the state of a subject device. Wherein,

the normal sample is generated by collecting or simulating more than two signals under the condition that the object equipment normally operates. (the types of sample signals include, but are not limited to, vibration, sound, velocity, displacement, stress, pressure, voltage, current, power, electric field strength, magnetic field strength, temperature, image, brightness, and combinations of signals)

The abnormal samples are generated by modifying a certain signal in the normal samples.

Wherein modifying a signal in a normal sample modifies the signal by amplitude, square amplitude, energy value, square energy value, or a ratio thereof.

The method comprises the steps of training, wherein in the training process, the amount of signal modification in an abnormal sample is reduced, so that a model with higher precision is obtained.

According to still another aspect of the present invention, there is provided a signal modification-based model evaluation method for diagnosing, discriminating, or identifying a state of a subject apparatus. Wherein:

and generating a normal sample by collecting or simulating more than two signals under the normal running condition of the object equipment.

An abnormal sample is generated by modifying a signal in the normal sample.

And testing the model by using the normal sample and the abnormal sample, and evaluating the performance of the model according to the test result.

Wherein modifying a signal in a normal sample is modifying the signal by amplitude, square amplitude, energy value, square energy value, or a ratio thereof.

And in the training process, the accuracy of the model is evaluated by reducing the signal modification amount in the abnormal sample.

According to the technical scheme, the abnormal sample or the negative sample can be effectively and accurately obtained, and the model is trained or tested by combining the obtained abnormal sample or negative sample with the normal sample or positive sample. By the model training and testing technology of the technical scheme, the model with better recognition accuracy can be obtained. The model obtained by the method or the system has good detection effect, low technical difficulty and low cost, and can be widely applied to schemes for identifying various devices.

Drawings

Exemplary embodiments of the present invention may be more completely understood in consideration of the following drawings:

FIG. 1 is a flow chart of a method of training a target model based on modification of a sample signal according to an embodiment of the invention;

FIG. 2 is a flow chart of a method of testing a target model based on modification of a sample signal according to an embodiment of the invention;

FIG. 3 is a schematic diagram of a system for training a target model based on modification of a sample signal in accordance with an embodiment of the present invention;

fig. 4 is a schematic diagram of a system for testing a target model based on modification of a sample signal according to an embodiment of the present invention.

Detailed Description

FIG. 1 is a flow chart of a method 100 of training a target model based on modification of a sample signal according to an embodiment of the invention. The method 100 begins at step 101.

In step 101, a target model to be trained is selected from a plurality of models, and a configuration file of the target model is obtained. In the actual context of industrial production or equipment operation, various types and/or sizes of equipment are widely used in various locations, production links, monitoring links, etc. For this reason, if it is necessary to determine the operation state of any one of the devices, or acquire parameters of any one of the devices, etc., it is necessary to acquire a model or a device model of each of the different types of devices. In general, a model or device model of each different type of device may be used to determine the operating state of the device, obtain operating parameters of the device, and so forth. For this reason, when a model of a specific device needs to be trained or tested, a target model that needs to be trained needs to be selected from models or device models associated with each of a plurality of different devices.

Typically, each model or device model has a configuration file, and the configuration file is used to describe various attributes of the model or device model. The various attributes are, for example: input parameters, output parameters, model type, model role, model accuracy, device type, device name, device identifier, etc. To this end, the object model has various attributes, and for example, a device type, a device name, a device identifier, and the like of the object model can be determined through a profile of the object model.

In step 102, an object device to which the target model relates is determined based on a profile of the target model, and a plurality of sample signals associated with the object device are determined. As described above, the device type, device name, device identifier, and the like of the object model can be determined by the configuration file of the object model. And further, the object device involved in the target model can be determined by parsing the configuration file of the target model. The object device may be any type of device. In addition, the configuration file of the target model may also include information of various sample signals associated with the object device. Alternatively, after determining the object device to which the target model relates, the device identifier or the device name of the object device may be used to retrieve information of a plurality of sample signals associated with the object device in the sample signal information base.

Wherein the plurality of sample signals includes one or more of: vibration, sound, speed, displacement, stress, pressure, voltage, current, power, electric field strength, magnetic field strength, temperature, image, and brightness. In general, the subject device may be characterized, trained, tested, described using one or more sample signals of vibration, sound, speed, displacement, stress, pressure, voltage, current, power, electric field strength, magnetic field strength, temperature, image, and brightness. It should be appreciated that the present application is described merely by way of example with respect to vibration, sound, speed, displacement, stress, pressure, voltage, current, power, electric field strength, magnetic field strength, temperature, image, and brightness, and that any reasonable sample signal may be used by those skilled in the art. In a practical scenario, various types of sensors may be used to acquire any of vibration, sound, speed, displacement, stress, pressure, voltage, current, power, electric field strength, magnetic field strength, temperature, image, and brightness.

In step 103, when the object device is operating normally, signal acquisition or signal simulation is performed on at least two sample signals in a plurality of sample signals, so as to obtain a normal sample set including the at least two sample signals.

A normal sample set is a sample signal that is composed of sample signals or data, such as one or more of vibration, sound, speed, displacement, stress, pressure, voltage, current, power, electric field strength, magnetic field strength, temperature, image, and brightness, acquired by a sensor when the subject device is operating normally. In the present application, in order to make the model accuracy of the trained or tested target model higher, the present application performs signal acquisition or signal simulation on at least two sample signals of a plurality of sample signals, thereby obtaining a normal sample set including the at least two sample signals.

For example, the at least two sample signals are a vibration signal, a sound signal, and a voltage signal, a plurality of samples arranged in the sampling time sequence of the samples are included in the normal sample set, wherein each sample includes the vibration signal, the sound signal, and the voltage signal, and each sample has a sampling time. That is, each sample in the normal sample set is a signal group or set of signals having a sampling time and comprising each sample signal at the sampling time. When the sample data is stored, at least two sample subsets can be included in the normal sample set, and each sample subset is a vibration sample signal subset, an acoustic emission sample signal subset, a sound sample signal subset, or an electric field intensity sample signal subset, etc. It should be understood that the manner in which the signal subsets are divided is for data storage or data presentation only. In practice, each sample comprises each of at least two sample signals. Alternatively, a plurality of sample signal groups are included in the normal sample set, each sample signal group comprising a single vibration sample signal, a single acoustic emission sample signal and a single sound sample signal, e.g. each sample signal group being < vibration sample signal, acoustic emission sample signal, sound sample signal >. It should be appreciated that each set of sample signals may be considered one sample of a normal set of samples.

For example, at least one of the at least two sample signals is a vibration/acoustic emission signal acquired by a sensor attached to a device housing of the subject device. Alternatively, at least one sample signal among the at least two sample signals is a sound signal acquired outside the device of the object device. In practice, the sensor may be arranged to be in close proximity to the housing of the device or object device, to be arranged outside the device or object device, or to be arranged inside the device or object.

In step 104, at least one sample signal in the normal sample set is modified according to a preset modification manner, so as to obtain an abnormal sample set corresponding to the normal sample set. In actual cases, the equipment with stable operation state has fewer faults or lower fault rate in actual operation, so that the data volume of the sample signal/sample data of the normal operation of the equipment is larger, and the data volume of the sample signal/sample data in abnormal operation or fault is smaller. In this case, it is often difficult to obtain a sufficient abnormal sample signal. For this purpose, the present application modifies at least one sample signal in the normal sample set according to a preset modification manner.

The preset modification modes comprise: the corresponding modifier is increased or decreased for the amplitude, the amplitude square, the energy value or the energy value square, so that the value or the value interval to which the amplitude, the amplitude square, the energy value or the energy value square of the corresponding modifier belongs is changed. For example, the amplitude sample signal in the normal sample set of the object device is increased or decreased by the corresponding modifier k, the amplitude square sample signal in the normal sample set of the object device is increased or decreased by the corresponding modifier b, the energy value sample signal in the normal sample set of the object device is increased or decreased by the corresponding modifier c or the energy value square sample signal in the normal sample set of the object device is increased or decreased by the corresponding modifier d, so that the amplitude sample signal, the amplitude square sample signal, the energy value sample signal or the energy value square sample signal in the normal sample set of the object device becomes an abnormal operation or a failure operation sample signal. Thus, an abnormal sample set corresponding to the normal sample set is constituted from the amplitude sample signal, the amplitude square sample signal, the energy value sample signal, or the energy value square sample signal at the time of abnormal operation or malfunction operation of the object device.

The preset modification mode further comprises the following steps: the corresponding ratio is raised or lowered for the amplitude, the amplitude square, the energy value or the energy value square, so that the value or the value interval to which the amplitude, the amplitude square, the energy value or the energy value square of the corresponding ratio belongs is changed. For example, the amplitude sample signal in the normal sample set of the object device is raised or lowered by the corresponding ratio e, the amplitude square sample signal in the normal sample set of the object device is raised or lowered by the corresponding ratio f, the energy value sample signal in the normal sample set of the object device is raised or lowered by the corresponding ratio g or the energy value square sample signal in the normal sample set of the object device is raised or lowered by the corresponding ratio h, so that the amplitude sample signal, the amplitude square sample signal, the energy value sample signal or the energy value square sample signal in the normal sample set of the object device becomes the sample signal at the time of abnormal operation or malfunction. Thus, an abnormal sample set corresponding to the normal sample set is constituted from the amplitude sample signal, the amplitude square sample signal, the energy value sample signal, or the energy value square sample signal at the time of abnormal operation or malfunction operation of the object device.

In addition, the preset modification mode further comprises the following steps: the corresponding steady state modifier is increased or decreased for amplitude, square amplitude, energy value, or square energy value. As described above, the amplitude, the square of the amplitude, the energy value, or the energy value square may have respective steady-state modifiers.

Wherein increasing or decreasing the corresponding steady state modifier by amplitude, square amplitude, energy value, or square energy value comprises: counting the maximum value Ymax and the minimum value Ymin of the corresponding parameters of the amplitude, the square amplitude, the energy value or the steady-state modifier DeltaY of the square energy value in the normal sample set; setting an adjustment target coefficient a, wherein 0< a <1; counting the value Ysignal of the corresponding parameter of the steady-state modifier DeltaY of the sample to be modified;

calculating steady-state modifier DeltaY=Ymin+a× (Ymax-Ymin) -Ysignal,

the corresponding steady-state modifier Δy is increased or decreased for amplitude, amplitude square, energy value, or energy value square.

Wherein the adjustment target coefficient a is obtained by adding a fixed value and a random value generated by a probability model, and satisfies 0< a <1. For example, a=m1+m2, where m1 is a fixed value and m2 is a random value generated by a probabilistic model.

In step 105, the target model is trained based on the normal sample set, the abnormal sample set, and a preset training algorithm, thereby obtaining a trained target model. The preset training algorithm can be any reasonable training algorithm in the artificial intelligence field, the deep learning field and the machine learning algorithm field. After the trained target model is obtained, an actual signal generated by the object device in actual operation is input into the trained target model, so that the operation state of the object device is diagnosed, judged or identified by using the trained target model, and the object device is determined to be in a normal state, an abnormal state or any other state. That is, the object model can diagnose, judge or recognize the operation state of the object device according to the collected sample signal, thereby determining that the object device is in a normal state, an abnormal state or any other state.

Inputting the pre-stored first test sample set into the trained target model, so that the trained target model diagnoses, discriminates or identifies the result state of the target device, and acquires the verification state corresponding to the pre-stored first test sample set. Wherein the validation state may correspond to each sample signal group or sample subset of the first set of test samples, respectively.

And determining the difference degree between the verification state and the result state, and determining that the trained target model meets the requirements when the difference degree is smaller than or equal to a preset threshold value. Since the resulting state of the object device for each sample signal group or sample subset in the first test sample set is known or predetermined, the degree of difference of the verification state from the resulting state may be used as a basis for determining the accuracy of the target model. For example, when a specific set of test sample signals is input to the trained target model, the resulting verification state is a ratio of 95:5 for the normal state to the abnormal state of the subject device, and the ratio of 99:1 for the normal state to the abnormal state of the subject device in the known or predetermined resultant state. From this, the degree of difference was 4% = (99-95)/(99+1). The predetermined threshold is a preset variance threshold, which may be considered as the minimum requirement for the accuracy of diagnosis, discrimination or identification of the target model. For example, when the predetermined threshold is 5%, the degree of difference is 4% and the degree of difference is less than the predetermined threshold, the trained target model is determined to be satisfactory. And when the preset threshold value is 2%, the difference degree is 4%, the difference degree is larger than the preset threshold value, and the trained target model is determined to be unsatisfactory.

Alternatively, after the normal sample set including the at least two sample signals is acquired, and the abnormal sample set corresponding to the normal sample set is acquired, the normal sample set is divided into a first normal sample subset and a second normal sample subset, and the abnormal sample set is divided into a first abnormal sample subset and a second abnormal sample subset. Wherein the first number ratio of sample signals in the first normal sample subset and the second normal sample subset (sample signals of normal samples or normal sample signals) is any reasonable ratio of 3:7,5:5,6:4, etc. Wherein the second number ratio of the sample signals (sample signals of the abnormal samples or abnormal sample signals) in the first abnormal sample subset and the second abnormal sample subset is any reasonable ratio of 3:7,5:5,6:4, etc. Wherein the first quantitative ratio and the second quantitative ratio may be equal or unequal.

And training the target model based on the first normal sample set, the first abnormal sample set and a preset training algorithm, so as to obtain a trained target model. The preset training algorithm can be any reasonable training algorithm in the artificial intelligence field, the deep learning field and the machine learning algorithm field.

Preferably, a second normal sample subset and a second abnormal sample subset are utilized to form a second test sample set, the second test sample set is input into a trained target model, so that the trained target model diagnoses, discriminates or identifies the result state of the target device, and the verification state corresponding to the second test sample set stored in advance is obtained; and determining the difference degree between the verification state and the result state. When the degree of difference is smaller than or equal to a preset threshold value, determining that the trained target model meets the requirements; when the degree of difference is greater than a predetermined threshold, it is determined that the trained target model is unsatisfactory.

Preferably, the method further comprises the step of obtaining an adjusted abnormal sample set by reducing the modification quantity, and training the target model based on the normal sample set, the adjusted abnormal sample set and a preset training algorithm, so as to obtain a trained high-precision target model. For example, after increasing or decreasing the respective modifier by the magnitude, the magnitude squared, the energy value, or the energy value squared, such that the magnitude of the respective modifier is increased or decreased, an adjusted set of abnormal samples is obtained by decreasing the modifier.

The method further comprises the step of obtaining an adjusted abnormal sample set by reducing the ratio, and training the target model based on the normal sample set, the adjusted abnormal sample set and a preset training algorithm, so that a trained high-precision target model is obtained. For example, increasing or decreasing the corresponding ratio after being the magnitude, magnitude squared, energy value, or energy value squared further includes decreasing the ratio to obtain an adjusted set of abnormal samples.

FIG. 2 is a flow chart of a method 200 of testing a target model based on modification of a sample signal according to an embodiment of the invention. The method 200 begins at step 201.

In step 201, a target model to be tested is determined and a configuration file of the target model is obtained. In the actual context of industrial production or equipment operation, various types and/or sizes of equipment are widely used in various locations, production links, monitoring links, etc. For this reason, if it is necessary to determine the operation state of any one of the devices, or acquire parameters of any one of the devices, etc., it is necessary to acquire a model or a device model of each of the different types of devices. In general, a model or device model of each different type of device may be used to determine the operating state of the device, obtain operating parameters of the device, and so forth. For this purpose, a target model to be tested is determined before the test is performed.

In step 202, an object device to which the target model relates is determined based on a profile of the target model, and a plurality of sample signals associated with the object device are determined. As described above, the device type, device name, device identifier, and the like of the object model can be determined by the configuration file of the object model. And further, the object device involved in the target model can be determined by parsing the configuration file of the target model. The object device may be any type of device. In addition, the configuration file of the target model may also include information of various sample signals associated with the object device. Alternatively, after determining the object device to which the target model relates, the device identifier or the device name of the object device may be used to retrieve information of a plurality of sample signals associated with the object device in the sample signal information base.

Wherein the plurality of sample signals comprises: vibration, sound, speed, displacement, stress, pressure, voltage, current, power, electric field strength, magnetic field strength, temperature, image, and brightness. In general, the subject device may be characterized, trained, tested, described using one or more sample signals of vibration, sound, speed, displacement, stress, pressure, voltage, current, power, electric field strength, magnetic field strength, temperature, image, and brightness. It should be appreciated that the present application is described merely by way of example with respect to vibration, sound, speed, displacement, stress, pressure, voltage, current, power, electric field strength, magnetic field strength, temperature, image, and brightness, and that any reasonable sample signal may be used by those skilled in the art. In a practical scenario, various types of sensors may be used to acquire any of vibration, sound, speed, displacement, stress, pressure, voltage, current, power, electric field strength, magnetic field strength, temperature, image, and brightness.

In step 203, when the object device is operating normally, signal acquisition or signal simulation is performed on at least two sample signals of a plurality of sample signals, so as to obtain a normal sample set including the at least two sample signals.

For example, at least one of the at least two sample signals is a vibration/acoustic emission signal acquired by a sensor attached to the housing of the device. Alternatively, at least one sample signal of the at least two sample signals is a sound signal collected outside the device. In practice, the sensor may be arranged to be in close proximity to the housing of the device or object device, to be arranged outside the device or object device, or to be arranged inside the device or object.

In step 204, at least one sample signal in the normal sample set is modified according to a preset modification manner, so as to obtain an abnormal sample set corresponding to the normal sample set. In actual cases, the equipment with stable operation state has fewer faults or lower fault rate in actual operation, so that the data volume of the sample signal/sample data of the normal operation of the equipment is larger, and the data volume of the sample signal/sample data in abnormal operation or fault is smaller. In this case, it is often difficult to obtain a sufficient abnormal sample signal. For this purpose, the present application modifies at least one sample signal in the normal sample set according to a preset modification manner.

The preset modification modes comprise: the corresponding ratio is raised or lowered for the amplitude, the amplitude square, the energy value or the energy value square, so that the value or the value interval to which the amplitude, the amplitude square, the energy value or the energy value square of the corresponding ratio belongs is changed. For example, the amplitude sample signal in the normal sample set of the object device is raised or lowered by the corresponding ratio e, the amplitude square sample signal in the normal sample set of the object device is raised or lowered by the corresponding ratio f, the energy value sample signal in the normal sample set of the object device is raised or lowered by the corresponding ratio g or the energy value square sample signal in the normal sample set of the object device is raised or lowered by the corresponding ratio h, so that the amplitude sample signal, the amplitude square sample signal, the energy value sample signal or the energy value square sample signal in the normal sample set of the object device becomes the sample signal at the time of abnormal operation or malfunction. Thus, an abnormal sample set corresponding to the normal sample set is constituted from the amplitude sample signal, the amplitude square sample signal, the energy value sample signal, or the energy value square sample signal at the time of abnormal operation or malfunction operation of the object device.

Wherein increasing or decreasing the corresponding steady state modifier by amplitude, square amplitude, energy value, or square energy value comprises:

counting the maximum value Ymax and the minimum value Ymin of the corresponding parameters of the steady-state modifier DeltaY of the amplitude, the amplitude square, the energy value or the energy value square in the normal sample set;

setting an adjustment target coefficient a, wherein 0< a <1;

counting the value Ysignal of the corresponding parameter of the steady-state modifier DeltaY of the sample to be modified;

calculating steady-state modifier DeltaY=Ymin+a× (Ymax-Ymin) -Ysignal,

The adjustment target coefficient a is obtained by adding a fixed value and a random value generated by a probability model, and satisfies 0< a <1. For example, a=m1+m2, where m1 is a fixed value and m2 is a random value generated by a probabilistic model.

In step 205, the target model is tested based on the normal sample set and the abnormal sample set, thereby determining a performance index of the target model based on the test results. The object model can diagnose, judge or identify the running state of the object equipment according to the collected sample signals, so as to determine that the object equipment is in a normal state, an abnormal state or any other state. Wherein the performance index of the target model may include a diagnostic accuracy, a discrimination accuracy, or an identification accuracy of the target model.

Testing the target model based on the normal sample set and the abnormal sample set, so that determining the performance index of the target model based on the test result comprises: and respectively or sequentially inputting the normal sample set and the abnormal sample set into the trained target model, so that the trained target model diagnoses, judges or identifies the result state of the target device, and acquires the verification state corresponding to the normal sample set and/or the abnormal sample set. A correct ratio of the result state is determined based on the verification state, a test result is determined based on the correct ratio, and a performance index of the target model is determined based on the test result. For example, after the normal sample set and the abnormal sample set are inputted into the trained target model separately or sequentially, the trained target model diagnoses, discriminates or recognizes the result state of the target device 100 times, determines that the number of times of diagnosing, discriminating or recognizing the result state is correct among 100 times based on the verification state is 99 times, determines that the correct ratio of the result state is 99/100=99% based on the verification state, and then the test result is 99%.

According to one embodiment, the performance index of the target model is determined to be high accuracy when the test result is greater than or equal to 97%, medium accuracy when the test result is less than 97% and greater than or equal to 90%, and low accuracy when the test result is less than 90%.

Preferably, the method further comprises obtaining an adjusted abnormal sample set by reducing the ratio, and testing the target model based on the normal sample set and the adjusted abnormal sample set, thereby determining the diagnosis precision, the discrimination precision or the identification precision of the target model. For example, increasing or decreasing the corresponding ratio after being the magnitude, magnitude squared, energy value, or energy value squared further includes decreasing the ratio to obtain an adjusted set of abnormal samples.

Preferably, the method further comprises the step of obtaining an adjusted abnormal sample set by reducing the modification quantity, and testing the target model based on the normal sample set and the adjusted abnormal sample set, so as to determine the diagnosis precision, the discrimination precision or the identification precision of the target model. For example, after increasing or decreasing the respective modifier by the magnitude, the magnitude squared, the energy value, or the energy value squared, such that the magnitude of the respective modifier is increased or decreased, an adjusted set of abnormal samples is obtained by decreasing the modifier.

FIG. 3 is a schematic diagram of a system 300 for training a target model based on modification of a sample signal according to an embodiment of the invention. The system 300 includes: selection means 301, determination means 302, acquisition means 303, modification means 304, training means 305 and identification means 306.

And the selecting device 301 is configured to select a target model to be trained from multiple models, and obtain a configuration file of the target model. In the actual context of industrial production or equipment operation, various types and/or sizes of equipment are widely used in various locations, production links, monitoring links, etc. For this reason, if it is necessary to determine the operation state of any one of the devices, or acquire parameters of any one of the devices, etc., it is necessary to acquire a model or a device model of each of the different types of devices. In general, a model or device model of each different type of device may be used to determine the operating state of the device, obtain operating parameters of the device, and so forth. For this reason, when a model of a specific device needs to be trained or tested, a target model that needs to be trained needs to be selected from models or device models associated with each of a plurality of different devices.

Determining means 302 for determining an object device to which the target model relates based on a profile of the target model, and determining a plurality of sample signals associated with the object device. As described above, the device type, device name, device identifier, and the like of the object model can be determined by the configuration file of the object model. And further, the object device involved in the target model can be determined by parsing the configuration file of the target model. The object device may be any type of device. In addition, the configuration file of the target model may also include information of various sample signals associated with the object device. Alternatively, after determining the object device to which the target model relates, the device identifier or the device name of the object device may be used to retrieve information of a plurality of sample signals associated with the object device in the sample signal information base.

And the acquiring device 303 is configured to perform signal acquisition or signal simulation on at least two sample signals in a plurality of sample signals when the object device is operating normally, so as to acquire a normal sample set including the at least two sample signals.

And the modifying device 304 is configured to modify at least one sample signal in the normal sample set according to a preset modification manner, so as to obtain an abnormal sample set corresponding to the normal sample set. In actual cases, the equipment with stable operation state has fewer faults or lower fault rate in actual operation, so that the data volume of the sample signal/sample data of the normal operation of the equipment is larger, and the data volume of the sample signal/sample data in abnormal operation or fault is smaller. In this case, it is often difficult to obtain a sufficient abnormal sample signal. For this purpose, the present application modifies at least one sample signal in the normal sample set according to a preset modification manner.

calculating steady-state modifier DeltaY=Ymin+a× (Ymax-Ymin) -Ysignal,

Training means 305 for training the target model based on the normal sample set, the abnormal sample set and a preset training algorithm, thereby obtaining a trained target model. The preset training algorithm can be any reasonable training algorithm in the artificial intelligence field, the deep learning field and the machine learning algorithm field. After the trained target model is obtained, an actual signal generated by the object device in actual operation is input into the trained target model, so that the operation state of the object device is diagnosed, judged or identified by using the trained target model, and the object device is determined to be in a normal state, an abnormal state or any other state. That is, the object model can diagnose, judge or recognize the operation state of the object device according to the collected sample signal, thereby determining that the object device is in a normal state, an abnormal state or any other state.

And the recognition device 306 is used for inputting an actual signal generated by the object device in actual operation into the trained target model after the trained target model is obtained, so that the operation state of the object device is diagnosed, judged or recognized by using the trained target model, and the object device is determined to be in a normal state or an abnormal state.

Fig. 4 is a schematic diagram of a system 400 for testing a target model based on modification of a sample signal according to an embodiment of the present invention. The system 400 includes: acquisition means 401, determination means 402, acquisition means 403, modification means 404, and test means 405.

The obtaining device 401 is configured to determine a target model that needs to be tested and obtain a configuration file of the target model. In the actual context of industrial production or equipment operation, various types and/or sizes of equipment are widely used in various locations, production links, monitoring links, etc. For this reason, if it is necessary to determine the operation state of any one of the devices, or acquire parameters of any one of the devices, etc., it is necessary to acquire a model or a device model of each of the different types of devices. In general, a model or device model of each different type of device may be used to determine the operating state of the device, obtain operating parameters of the device, and so forth. For this purpose, a target model to be tested is determined before the test is performed.

Determining means 402 for determining an object device to which the target model relates based on a profile of the target model, and determining a plurality of sample signals associated with the object device. As described above, the device type, device name, device identifier, and the like of the object model can be determined by the configuration file of the object model. And further, the object device involved in the target model can be determined by parsing the configuration file of the target model. The object device may be any type of device. In addition, the configuration file of the target model may also include information of various sample signals associated with the object device. Alternatively, after determining the object device to which the target model relates, the device identifier or the device name of the object device may be used to retrieve information of a plurality of sample signals associated with the object device in the sample signal information base.

And the obtaining device 403 is configured to perform signal acquisition or signal simulation on at least two sample signals in a plurality of sample signals when the object device is operating normally, so as to obtain a normal sample set including the at least two sample signals.

And the modifying device 404 is configured to modify at least one sample signal in the normal sample set according to a preset modification manner, so as to obtain an abnormal sample set corresponding to the normal sample set.

In actual cases, the equipment with stable operation state has fewer faults or lower fault rate in actual operation, so that the data volume of the sample signal/sample data of the normal operation of the equipment is larger, and the data volume of the sample signal/sample data in abnormal operation or fault is smaller. In this case, it is often difficult to obtain a sufficient abnormal sample signal. For this purpose, the present application modifies at least one sample signal in the normal sample set according to a preset modification manner.

setting an adjustment target coefficient a, wherein 0< a <1;

calculating steady-state modifier DeltaY=Ymin+a× (Ymax-Ymin) -Ysignal,

And the testing device 405 is used for testing the target model based on the normal sample set and the abnormal sample set, so as to determine the performance index of the target model based on the test result. The object model can diagnose, judge or identify the running state of the object equipment according to the collected sample signals, so as to determine that the object equipment is in a normal state, an abnormal state or any other state. Wherein the performance index of the target model may include a diagnostic accuracy, a discrimination accuracy, or an identification accuracy of the target model.

Claims

1. A method of training a target model based on modification of a sample signal, the method comprising:

training the target model based on the normal sample set, the abnormal sample set and a preset training algorithm, thereby obtaining a trained target model, comprising: dividing the normal sample set into a first normal sample subset and a second normal sample subset, and dividing the abnormal sample set into a first abnormal sample subset and a second abnormal sample subset;

training the target model based on the first normal sample set, the first abnormal sample set and a preset training algorithm, so as to obtain a trained target model;

A second test sample set is formed by using the second normal sample set and the second abnormal sample set, the second test sample set is input into a trained target model, so that the trained target model diagnoses, judges or identifies the result state of the target device, and the verification state corresponding to the pre-stored second test sample set is obtained;

determining the difference between the verification state and the result state, and determining that the trained target model meets the requirements when the difference is smaller than or equal to a preset threshold value;

when the difference degree is larger than a preset threshold value, determining that the trained target model does not meet the requirements;

the preset modification mode comprises the following steps: the corresponding modifier is increased or decreased for the amplitude, the amplitude square, the energy value or the energy value square, so that the value or the value interval to which the amplitude, the amplitude square, the energy value or the energy value square of the corresponding modifier belongs is changed;

the preset modification mode comprises the following steps: increasing or decreasing the corresponding steady state modifier for amplitude, square amplitude, energy value, or square energy value;

setting an adjustment target coefficient a, wherein 0< a <1;

calculating steady-state modifier DeltaY=Ymin+a× (Ymax-Ymin) -Ysignal,

increasing or decreasing the corresponding steady-state modifier Δy for amplitude, square amplitude, energy value, or square energy value;

the adjustment target coefficient a is obtained by adding a fixed value and a random value generated by a probability model, and satisfies 0< a <1.

2. The method of claim 1, the plurality of sample signals comprising: vibration, sound, speed, displacement, stress, pressure, voltage, current, power, electric field strength, magnetic field strength, temperature, image, and brightness.

3. The method of claim 1, at least one of the at least two sample signals being a vibration/acoustic emission signal acquired by a sensor attached to a housing of the device.

4. The method of claim 1, at least one of the at least two sample signals being a sound signal acquired external to the device.

5. A method of testing a target model based on modification of a sample signal, the method comprising:

testing the target model based on the normal sample set and the abnormal sample set, so as to determine the performance index of the target model based on the test result;

the method further comprises training the target model based on the normal sample set, the abnormal sample set and a preset training algorithm, so as to obtain a trained target model, and comprises the following steps: dividing the normal sample set into a first normal sample subset and a second normal sample subset, and dividing the abnormal sample set into a first abnormal sample subset and a second abnormal sample subset;

setting an adjustment target coefficient a, wherein 0< a <1;

calculating steady-state modifier DeltaY=Ymin+a× (Ymax-Ymin) -Ysignal,

6. A system for training a target model based on modification of a sample signal, the system comprising:

the training device is used for training the target model based on the normal sample set, the abnormal sample set and a preset training algorithm, so that a trained target model is obtained;

the modifying device is used for dividing the normal sample set into a first normal sample subset and a second normal sample subset, and dividing the abnormal sample set into a first abnormal sample subset and a second abnormal sample subset;

the training device utilizes a second normal sample subset and a second abnormal sample subset to form a second test sample set, the second test sample set is input into a trained target model, so that the trained target model diagnoses, judges or identifies the result state of the target equipment, and the verification state corresponding to the second test sample set stored in advance is obtained;

setting an adjustment target coefficient a, wherein 0< a <1;

calculating steady-state modifier DeltaY=Ymin+a× (Ymax-Ymin) -Ysignal,

7. The system of claim 5, the plurality of sample signals comprising: vibration, sound, speed, displacement, stress, pressure, voltage, current, power, electric field strength, magnetic field strength, temperature, image, and brightness.

8. The system of claim 5, at least one of the at least two sample signals being a vibration/acoustic emission signal acquired by a sensor attached to the device housing.

9. The system of claim 5, at least one of the at least two sample signals being a sound signal acquired external to the device.

10. A system for testing a target model based on modification of a sample signal, the system comprising:

the testing device is used for testing the target model based on the normal sample set and the abnormal sample set, so that the performance index of the target model is determined based on the testing result;

the training device is used for forming a second test sample set by utilizing the second normal sample set and the second abnormal sample set, inputting the second test sample set into the trained target model, so that the trained target model diagnoses, judges or identifies the result state of the target equipment, and acquires the verification state corresponding to the prestored second test sample set;

setting an adjustment target coefficient a, wherein 0< a <1;

calculating steady-state modifier DeltaY=Ymin+a× (Ymax-Ymin) -Ysignal,