CN113610167A - Equipment risk detection method based on metric learning and visual perception - Google Patents
Equipment risk detection method based on metric learning and visual perception Download PDFInfo
- Publication number
- CN113610167A CN113610167A CN202110915531.3A CN202110915531A CN113610167A CN 113610167 A CN113610167 A CN 113610167A CN 202110915531 A CN202110915531 A CN 202110915531A CN 113610167 A CN113610167 A CN 113610167A
- Authority
- CN
- China
- Prior art keywords
- risk
- equipment
- real
- gaussian distribution
- ith
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to the technical field of metric learning and visual perception, in particular to an equipment risk detection method based on metric learning and visual perception. Inputting a first image of equipment to be tested and a risk monitoring index thereof into a corresponding equipment risk detection network to obtain a corresponding risk prediction result; the loss function of the equipment risk detection network is the sum of a cross entropy loss function and a mean square error loss function of the associated total variation, wherein the associated total variation is the similarity of the probabilities of the predicted risk levels of a plurality of pieces of equipment in the equipment images acquired at the same time period in different data; similarity is a weighted sum of probability ratios based on the predicted risk level between any two devices, the predicted risk probability, the risk fluctuation index, and the difference in risk level between the two devices. According to the invention, the risk prediction result of the equipment can be obtained in real time by utilizing the equipment image and the risk monitoring index to carry out network training, and the real-time performance and the accuracy of the equipment risk detection are improved.
Description
Technical Field
The invention relates to the technical field of metric learning, in particular to an equipment risk detection method based on metric learning and visual perception.
Background
At present, the economic level development is in the process of advancing from the middle-low end to the high end, but the safety cognition level and the safety concept at the enterprise level are relatively lagged behind, so that the safety management level is still at the bottom level, the accident prevention is still at the passive stage, and most of enterprises still belong to extensive safety management. Therefore, the detection of the equipment risk is enhanced, a prevention mechanism is adopted in advance, and a risk classification implementation management and control system is constructed.
At present, the risk detection method usually adopts artificial risk investigation to find out the equipment risk. The risk detection method is low in accuracy, and mutual influence and mutual correlation among the devices are not considered.
Disclosure of Invention
In order to solve the above technical problems, an object of the present invention is to provide a method for detecting equipment risk based on metric learning and visual perception, and the adopted technical solution is specifically as follows:
one embodiment of the invention provides a device risk detection method based on metric learning and visual perception, which comprises the following steps:
inputting a first image of equipment to be tested and a risk monitoring index thereof into a corresponding equipment risk detection network to obtain a corresponding risk prediction result;
the loss function of the equipment risk detection network is the sum of a cross entropy loss function and a mean square error loss function of the associated total variation, and the associated total variation is the difference value of the real-time associated total variation and the standard associated total variation; the standard association total variation is similarity according to probabilities of predicted risk levels of a plurality of devices in device images acquired at the same time period in historical data; the real-time correlation total variation is similarity according to probabilities of predicted risk levels of a plurality of devices in device images acquired at the same time period in real-time data; the similarity is a weighted sum of probability ratios of predicted risk levels between any two of the devices, and the probability of a predicted risk level is obtained from the predicted risk level of a device, the predicted risk probability, a risk fluctuation index, and a difference in risk level between the two devices.
Preferably, the method for acquiring a risk fluctuation indicator includes:
acquiring a risk level sequence corresponding to an equipment image of the equipment, and adopting Gaussian background modeling to acquire mixed Gaussian distribution and a mixed Gaussian distribution curve corresponding to the risk level sequence, wherein the abscissa of the mixed Gaussian distribution curve is a risk level, and the ordinate of the mixed Gaussian distribution curve is a risk probability;
averaging a plurality of the mixed Gaussian distribution curves to obtain a final Gaussian distribution curve; and obtaining the risk fluctuation index of the equipment from the final Gaussian distribution curve.
Preferably, the method for acquiring the difference of the risk levels between the two devices includes:
calculating a distance value between the final Gaussian distribution curves corresponding to any two devices; the distance value is taken as the difference in risk level between the two devices.
Preferably, the obtaining the risk fluctuation index of the equipment from the final gaussian distribution comprises:
the risk fluctuation index calculation formula is as follows:
wherein, BiThe risk fluctuation index for the ith said device; simiIs the gaussian distribution similarity of the ith said device.
Preferably, the method for obtaining the gaussian distribution similarity of the ith device includes:
the calculation formula of the gaussian distribution similarity of the ith device is as follows:
wherein f isi(ii) is the final gaussian distribution for the ith said device; argmax (f)i) Is fiTaking the corresponding risk grade when the maximum value is obtained; n (argmax (f)i) 1) taking argmax (f) as the mean valuei) Is normally distributed.
Preferably, the standard associated total variation is similarity of probabilities of predicted risk levels of a plurality of devices in device images acquired from the same period in the historical data, and includes:
the calculation formula of the standard correlation total variation is as follows:
wherein L is0Correlating the total variation for the criteria; e is the set of all the devices; i, j belongs to the ith device and the jth device of the E set; w is ai，jA difference in the risk level for an ith device and a jth device; giAttention of the ith device; gjAttention of the jth device; fiA risk probability sequence corresponding to a final Gaussian distribution function of the historical data of the ith device; fjA risk probability sequence corresponding to a final Gaussian distribution function of the history data of the jth device; max (F)i) The maximum risk probability in the risk probability sequence of the ith device; argmax (F)i) A risk level corresponding to the maximum risk probability on an ith device; max (F)j) The maximum risk probability in the risk probability sequence of the jth device; argmax (F)j) A risk level corresponding to the maximum risk probability on the jth device.
Preferably, the method for acquiring the attention includes:
the calculation formula of the attention degree is as follows:
wherein, giThe attention of the ith said device; b isiThe risk fluctuation index for the ith said device; b isjThe risk fluctuation index for the jth of the devices; num is the number of all the devices.
Preferably, the real-time correlation total variation is a similarity of probabilities of predicted risk levels of a plurality of devices in a device image acquired at the same time interval in real-time data, and includes:
the calculation formula of the real-time total variation of the correlation diagram is as follows:
wherein L isuCorrelating the total variation for the criteria; e is the set of all the devices; i, j belongs to the ith device and the jth device of the E set; w is ai，jA difference in the risk level for an ith device and a jth device; giAttention of the ith device; gjAttention of the jth device; ziA real-time risk probability sequence corresponding to the real-time data of the ith device; zjA real-time risk probability sequence corresponding to the real-time data of the jth device; max (Z)i) The real-time maximum risk probability in the real-time risk probability sequence of the ith device; argmax (Z)i) A real-time risk level corresponding to the real-time maximum risk probability on an ith device; max (Z)j) The real-time maximum risk probability in the real-time risk probability sequence of the jth device; argmax (Z)j) And the real-time risk level corresponding to the real-time maximum risk probability on the jth device.
Preferably, the loss function of the equipment risk detection network is the sum of a cross entropy loss function and a mean square error loss function of the associated total variation, and includes:
the construction formula of the loss function is as follows:
wherein M is the number of training data in a training batch of the equipment;
predicting risk probability that the u-th training data output by the risk detection network corresponding to the kth equipment belong to the class c risk level;
belonging to the class c risk level for the u training data in the k equipment; l isuFor the u-th training data of the k-th device and other devicesThe training data at the same position as the u-th training data are input into the respective equipment risk detection network to obtain the real-time total variation of the association diagram; l is0The standard total variation is the correlation diagram.
The invention has the following beneficial effects:
according to the embodiment of the invention, based on metric learning and visual perception, various indexes such as risk fluctuation indexes, attention degrees, difference of risk levels, Gaussian mixture distribution, risk probability corresponding to the risk levels and the like are obtained according to an equipment image sequence, a risk monitoring index sequence and the corresponding risk levels, loss functions are constructed by the various indexes to train an equipment risk detection network, the equipment risk levels are predicted through the trained risk detection network, and artificial equipment risk level evaluation is not required; when the loss function is constructed, on one hand, the correlation degree between the devices is used as the difference of the risk levels between the nodes, and on the other hand, the accuracy of the risk level prediction of the devices to be tested is further improved through the influence of other devices on the devices to be tested; finally, the equipment image sequence and the risk monitoring index sequence are input into the equipment risk detection network to obtain the risk level of the equipment, artificial equipment risk level evaluation in fixed time is not needed, meanwhile, the fact that relevant relevance exists among the equipment in the same production scene is also considered, and accuracy and real-time performance of risk level prediction are further improved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions and advantages of the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.
Fig. 1 is a flowchart of a method for detecting a device risk based on metric learning and visual perception according to an embodiment of the present invention;
FIG. 2 is a diagram illustrating a case of multiple training data packets according to an embodiment of the present invention;
fig. 3 is a flowchart of a method for training a device risk detection network according to an embodiment of the present invention.
Detailed Description
To further illustrate the technical means and effects of the present invention for achieving the predetermined objects, the following detailed description of the method for detecting equipment risk based on metric learning and visual perception according to the present invention with reference to the accompanying drawings and preferred embodiments shows the following detailed descriptions. In the following description, different "one embodiment" or "another embodiment" refers to not necessarily the same embodiment. Furthermore, the particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
The embodiment of the invention provides a specific implementation method of an equipment risk detection method based on metric learning and visual perception, the method is suitable for risk detection of formaldehyde filtering equipment in a chemical plant, a camera in the embodiment is a monitoring RGB camera, and the camera acquires equipment images in a neighborhood range of equipment to be detected; and a plurality of gas sensors are arranged around the equipment to be tested, and image information and gas sensor reading information of the equipment to be tested are acquired in real time. According to the embodiment of the invention, the network training is carried out by constructing the loss function, the risk grade of the equipment to be tested is obtained through the trained equipment risk detection network, the equipment does not need to be artificially labeled with the risk grade, the relevance among the equipment is considered, and the risk grade obtained by the equipment risk detection network is more accurate than the artificially evaluated risk grade. The method and the device achieve the purposes of obtaining the risk level of the equipment in real time and improving the accuracy and the real-time performance of risk level prediction through the relevance between the equipment.
The following describes a specific scheme of the equipment risk detection method based on metric learning and visual perception in detail with reference to the accompanying drawings.
Referring to fig. 1, a flowchart of a method for detecting a device risk based on metric learning and visual perception according to an embodiment of the present invention is shown,
and S100, inputting the first image of the equipment to be tested and the risk monitoring index thereof into the corresponding equipment risk detection network to obtain a corresponding risk prediction result.
The method comprises the steps of collecting RGB images of equipment to be tested by using an RGB monitoring camera, and extracting neighborhood images of the equipment to be tested in the RGB images to be used as first images. In the embodiment of the invention, a target monitoring network is utilized to obtain a central point coordinate and a width and height dimension of an enclosure frame of a device to be detected, wherein the central coordinate of the enclosure frame is (x, y), the width of the enclosure frame is w, and the height of the enclosure frame is h, the enclosure frame information is formed by the central coordinate and the width and height dimension, the device to be detected is arranged in the enclosure frame, and the enclosure frame is utilized to cut the RGB image to obtain a neighborhood image of the device to be detected as a first image.
As another embodiment, for the acquisition of the first image, an image of the device to be detected may be obtained by using edge detection, and then a neighborhood image of the device to be detected is obtained as the first image by performing a difference between the image of the device to be detected and the original RGB image.
And obtaining position information according to the coordinates of the central point of the surrounding frame of the equipment to be tested, determining the sensors deployed around the equipment to be tested, collecting all sensor data, and taking the sensor data as a risk monitoring index.
In an actual production scenario of a chemical plant, an administrator of an enterprise will perform maintenance on equipment at regular intervals of time t, and reset the safety risk of the equipment after the maintenance to 0. In order to facilitate the analysis of the risk fluctuation index of the subsequent equipment, a group of data segments are obtained when the acquisition time length is t during data acquisition.
In each group of data segments with the time length t, each elapsed time t0Obtaining a set of training data, that is, each set of data segment includes multiple sets of training data, each having a length of t0Training data ofThe first image sequence and the risk monitoring index sequence are contained in a group, and the condition of the groups of training data packets is shown in figure 2.
Wherein, will be at t0And performing a concatemate operation on each frame of first image acquired in time to obtain a first image sequence. Will be at t0And (4) carrying out the concatenate operation on the risk monitoring indexes acquired by time to obtain a risk monitoring index sequence.
And inputting the first image sequence and the risk monitoring index sequence into a corresponding equipment risk detection network to obtain a corresponding risk prediction result. The risk prediction result is the corresponding risk grade and risk probability of the device to be tested. In the embodiment of the present invention, the risk level is divided into 5 classes. The higher the grade is, the higher the risk of the device to be tested is, the more dangerous the device to be tested is, and the lower the grade is, the lower the risk of the device to be tested is, and the safer the device to be tested is.
Step S200, a loss function of the equipment risk detection network is the sum of a cross entropy loss function and a mean square error loss function of the associated total variation, and the associated total variation is the difference value of the real-time associated total variation and the standard associated total variation; the standard association total variation is similarity of probabilities of predicting risk levels of a plurality of devices in device images acquired at the same time period in historical data; the real-time correlation total variation is the similarity of the probabilities of the predicted risk levels of the plurality of devices in the device image acquired at the same time period in the real-time data; the similarity is a weighted sum of probability ratios of predicted risk levels between any two devices, and the probability of a predicted risk level is obtained from the predicted risk level of a device, the predicted risk probability, a risk fluctuation index, and the difference in risk levels between the two devices.
The equipment risk detection network training mainly comprises the following steps:
step S201, acquiring mode of risk fluctuation index.
And analyzing the training data of each device to obtain the risk fluctuation index of each device. The risk fluctuation index reflects the probability of the equipment risk level changing, and the risk fluctuation index of the equipment participates in the design of a subsequent loss function, so that the equipment risk detection network is more sensitive to equipment with a larger risk fluctuation index, and the accuracy of risk detection is improved.
Optionally selecting a device i as an example, and detailing the risk fluctuation index B of the device iiThe calculating method of (2):
(1) obtaining training data of a device i, for each data segment t1，t2，...，tnEach of length t0Training data of (2) is artificially labeled with risk level for each training data, i.e. each training data has a length of t0The equipment image sequence and the risk monitoring index sequence in the time period both correspond to a risk level label.
Data segment t1Of multiple lengths t0The risk grades corresponding to the time period are arranged according to the sequence of time acquisition to obtain a data segment t1Corresponding risk rank sequence, denoted D1＝{d1，d2，d3，...，dm}. Where m is the data segment t1Number of risk classes involved, i.e. m-t1/t0。
(2) Adopting a mixed Gaussian background modeling mode according to the risk grade sequence D1Constructing a data segment t1Corresponding mixed Gaussian distribution f1And a Gaussian mixture profile.
1) First of all with the risk level d1Obtaining an initial mixed gaussian distribution:
wherein the content of the first and second substances,
is a risk class d1The weight of (c);
is a risk class d1A corresponding gaussian distribution;
is a risk class d1Standard deviation of the data of (a); x is the current corresponding risk level;
is a risk class d1Mean of the data of (1).
2) Rank the risk d2And risk class d1Corresponding Gaussian distribution
And (6) matching. Judgment of
If yes, the risk level d is indicated2And Gaussian distribution
If the matching is successful, the risk grade d is used2For the Gaussian distribution with successful matching
Updating is carried out; the updated formula is as follows:
wherein the content of the first and second substances,
for the risk level d after updating1The data mean of (2);
for the risk level d after updating1Standard deviation of the data of (a); ρ is a forgetting coefficient, i.e. the risk level d2Substituting gaussian distribution
Resulting updated Gaussian distribution
A value of (i), i.e
Then by the updated data mean
And standard deviation of data
For Gaussian distribution
Updating to obtain the updated Gaussian distribution
The Gaussian mixture distribution at this time is
3) If not, the risk grade d is indicated2And Gaussian distribution
If the matching fails, a new Gaussian distribution is obtained
Wherein the content of the first and second substances,
is a risk class d2Standard deviation of the data of (a); x is the current corresponding risk level;
is a risk class d2Mean of the data of (1).
4) For multiple Gaussian distributions in mixed Gaussian distribution
Weights are assigned to obtain the final mixture gaussian distribution. The weight distribution method is as follows:
wherein the content of the first and second substances,
representing a Gaussian distribution
Updating the previous weight;
representing a Gaussian distribution
Updating the previous weights due to previous Gaussian distribution
Without weight, then this time
Is 0;
indicating a level of riskd2And Gaussian distribution
The matching result of (1);
denotes d2And Gaussian distribution
The matching result of (1); and if the matching is successful, the matching result value is 1, and if the matching is failed, the matching result value is 0.
Finally, the weight is weighted
And weight
Normalizing to obtain final weight
The Gaussian mixture distribution at this time is
5) Rank the risk d3Matching each Gaussian distribution in the mixed Gaussian distributions according to the step 2). E.g. if the risk rating d2And Gaussian distribution
The matching is successful to obtain the updated Gaussian distribution
Then the risk level d3And Gaussian distribution
Matching, if matching is successful, the Gaussian distribution is conducted again
Updating to obtain updated Gaussian distribution
The mixture gaussian distribution at this time
If the risk grade d of the previous step2And Gaussian distribution
The matching is not successful, and a new Gaussian distribution is obtained
Then the risk level d3Respectively with a Gaussian distribution
And Gaussian distribution
Matching is performed and steps 2) -4) are repeated again.
If the risk class d3If the matching with each Gaussian distribution in the mixed Gaussian distribution fails, a new Gaussian distribution is obtained
Finally, weights are distributed to the Gaussian distributions to obtain a new mixed Gaussian distribution f1I.e. a Gaussian distribution
Gaussian distribution
And Gaussian distribution
Assigning weights to obtain a new Gaussian mixture distribution f1。
Repeat traversal D1All values of (1), i.e. d1，d2，d3，...，dmTo obtain the final data segment t1Corresponding mixed Gaussian distribution f1. And obtaining a Gaussian mixture distribution curve according to the Gaussian mixture distribution, wherein the abscissa of the Gaussian mixture distribution curve is a risk level, the ordinate of the Gaussian mixture distribution curve is a risk probability, and the risk probability is the probability of occurrence of the corresponding risk level.
(3) And (3) repeatedly obtaining the Gaussian mixture distribution of each data segment and the corresponding Gaussian mixture distribution curve according to the method in the step (2), and marking the Gaussian mixture distribution as f1，f2，...，fn。
In order to avoid the influence of errors in the mixed Gaussian distribution obtained by a single data segment on subsequent processing, the embodiment of the invention calculates the average value of n risk probability values corresponding to the same abscissa in n mixed Gaussian distribution curves to obtain the final Gaussian distribution curve corresponding to the device i, and the final Gaussian distribution curve is marked as fi。
(4) As known from the priori knowledge, when the risk level of the equipment does not change, namely the risk level of the equipment is stable, namely the safety degree of the equipment is stable, the final Gaussian distribution curve fiThe normal distribution should be conformed, i.e. the more similar the final gaussian distribution curve is to the normal distribution curve, the more stable the risk level, and the larger the gaussian distribution similarity value, the more stable the risk level. The idea of metric-based learning consists of a final Gaussian distribution fiObtaining the similarity Sim with the normal distributioni。
Gaussian distribution similarity Sim of ith deviceiThe calculation formula of (2) is as follows:
wherein f isiThe final gaussian distribution for the ith device; argmax (f)i) Is fiTaking the corresponding risk grade when the maximum value is obtained; n (argmax (f)i) 1) taking argmax (f) as the mean valuei) Normal distribution of (2) to avoid Gauss distributionCloth curve fiAnd calculating the influence of translation on the similarity. It should be noted that the DTW algorithm is a known technique for calculating the similarity of curves, and is not described in detail.
Known from the prior knowledge, when the risk level of the equipment does not change, the final Gaussian distribution curve fiThe risk fluctuation index of the equipment can be measured by using the similarity with the normal distribution, namely, the closer the final Gaussian distribution curve corresponding to the equipment is to the normal distribution, the greater the similarity is, the smaller the fluctuation degree of the wind wave grade of the equipment is, and the more stable the risk grade is. And calculating the risk fluctuation index of the equipment according to the similarity obtained by the final Gaussian distribution.
Risk fluctuation index BiThe calculation formula of (2) is as follows:
wherein, BiA risk fluctuation index of the ith equipment; simiIs the gaussian distribution similarity of the ith device. It should be noted that the more stable the risk level of the equipment is, the higher the safety factor of the equipment is, and the risk fluctuation index BiThe smaller.
And step S202, carrying out equipment risk correlation analysis and constructing an equipment risk correlation diagram.
The production procedures of enterprises are fixed and unchangeable, so that certain risk relevance exists among all the devices, the risk relevance can provide auxiliary supervision information for the training of the subsequent device risk detection network, and the output accuracy of the device risk detection network is improved. And performing equipment risk correlation analysis according to the final Gaussian distribution curve of each equipment to construct an equipment risk correlation diagram, wherein the equipment risk correlation diagram can reflect the risk correlation existing between the equipment.
According to step S201, a final gaussian distribution curve for each device is obtained. And regarding each device as a node, storing the final Gaussian distribution curve corresponding to the device in each node, calculating a distance value between the final Gaussian distribution curves corresponding to any two devices by using a DTW algorithm, and taking the distance value as an edge weight value between the two devices to obtain a device risk association graph, wherein the edge weight value is also the difference of risk levels. The edge weight value reflects the deviation between Gaussian distribution curves of the two devices, namely reflects the difference of risk levels between the two devices, and each node and all other nodes in the device risk association graph have a connection relationship.
Because the risk fluctuation indexes of the devices are different, in order to ensure the safe operation of enterprise production, the device with the higher risk fluctuation index needs to be assigned with a higher attention, and the device with the higher risk fluctuation index is the device with the low risk level unstable safety degree. Therefore, the calculation of the attention of each device is obtained by fitting the risk fluctuation indexes corresponding to each device, and each device and each node correspond to one attention.
Degree of attention giThe calculation formula of (2) is as follows:
wherein, giAttention of the ith device; b isiA risk fluctuation index of the ith equipment; b isjA risk fluctuation index of the jth equipment; num is the number of all devices.
In step S203, the standard association total variation is a similarity of probabilities of predicted risk levels of the plurality of devices in the device image acquired from the same period in the history data. The similarity is a weighted sum of probability ratios of predicted risk levels between any two devices, and the probability of a predicted risk level is obtained from the predicted risk level of a device, the predicted risk probability, a risk fluctuation index, and the difference in risk levels between the two devices.
The degree of association between devices is characterized by the attention, risk level and risk probability between devices. Because the relationship between the parameters is complex, the standard association total variation between the devices is obtained by fitting the attention degree, the difference of the risk level, the risk level and the risk level probability between the devices in a mathematical modeling mode.
Wherein the training data of the device is historical training data.
The calculation method of the standard correlation total variation is as follows:
wherein L is0The standard correlation total variation; e is the set of all devices; i, j belongs to the ith device and the jth device of the E set; w is ai，jIs the difference in risk level for the ith device and the jth device; giAttention of the ith device; gjAttention of the jth device; fiA risk probability sequence corresponding to a final Gaussian distribution function of the historical data of the ith device; fjA risk probability sequence corresponding to a final Gaussian distribution function of the history data of the jth device; max (F)i) The maximum risk probability in the risk probability sequence of the ith device; argmax (F)i) A risk level corresponding to the maximum risk probability on the ith device; max (F)j) The maximum risk probability in the risk probability sequence of the jth device; argmax (F)j) The risk level corresponding to the maximum risk probability on the jth device.
It should be noted that i and j respectively represent the numbers of the nodes of the device, and are manually specified according to a certain rule. In order to distinguish different nodes when calculating the total variation of the association, the number corresponding to the node of each device is different,
step S204, the real-time correlation total variation is the similarity of the probabilities of the predicted risk levels of the plurality of devices in the device image acquired in the same time period in the real-time data. The similarity is a weighted sum of probability ratios of predicted risk levels between any two devices, and the probability of a predicted risk level is obtained from the predicted risk level of a device, the predicted risk probability, a risk fluctuation index, and the difference in risk levels between the two devices.
The degree of association between devices is characterized by the attention, risk level and risk probability between devices. Due to the fact that the relation among the parameters is complex, the real-time correlation total variation among the devices is obtained by means of mathematical modeling through the attention degree, the difference of the risk levels, the risk levels and the risk level probability among the devices.
Wherein the training data of the device is real-time training data.
The calculation formula of the real-time total variation of the correlation diagram is as follows:
wherein L isuThe standard correlation total variation; e is the set of all devices; i, j belongs to the ith device and the jth device of the E set; w is ai，jIs the difference in risk level for the ith device and the jth device; giAttention of the ith device; gjAttention of the jth device; ziA real-time risk probability sequence corresponding to the real-time data of the ith device; zjA real-time risk probability sequence corresponding to the real-time data of the jth device; max (Z)i) The real-time maximum risk probability in the real-time risk probability sequence of the ith device; argmax (Z)i) A real-time risk level corresponding to a real-time maximum risk probability on the ith device; max (Z)j) The real-time maximum risk probability in the real-time risk probability sequence of the jth equipment; argmax (Z)j) The real-time risk level corresponding to the real-time maximum risk probability on the jth device.
And S205, constructing a loss function of the equipment risk detection network, and training the equipment risk detection network.
Each device corresponds to a device risk detection network. And training the network corresponding to each device, and taking the trained network as a device risk detection network. The equipment risk detection network inputs training data of each equipment and outputs the training data as risk levels, and the training data are equipment images and risk monitoring indexes.
Firstly, a loss function of the equipment risk detection network is established, the loss function needs to be minimized as much as possible when the loss function is established, and the standard associated total variation and the real-time associated total variation are differentiated to enable the real-time associated total variable to be close to the standard total variation as much as possible. Due to the fact that the comprehensive relation among the standard correlation total variation, the real-time correlation total variation and the risk levels and risk probabilities corresponding to the training data of different time periods of each device is complex, the functional relation among the standard correlation total variation, the real-time correlation total variation and the risk levels and risk probabilities corresponding to the training data of different time periods of each device is fitted through a mathematical modeling method, and the loss function of the device risk detection network is obtained.
The loss function is constructed by the formula:
wherein M is the number of training data in a training batch of equipment;
predicting risk probability that the u-th training data output by the risk detection network corresponding to the kth equipment belong to the class c risk level;
belonging to the class c risk level for the u training data in the k equipment; l isuInputting the association graphs obtained by the respective device risk detection networks for the u-th training data of the kth device and the training data of other devices at the same position as the u-th training data to obtain real-time total variation; l is0The correlation diagram standard total variation.
The loss function provides supervision information by using the relevance of risks among the devices, and the accuracy of network output is ensured.
In summary, in the embodiments of the present invention, metric learning and visual perception are utilized, a risk fluctuation index, an attention degree, a mixed gaussian distribution and a risk probability corresponding to a risk level are obtained by utilizing an apparatus image sequence, a risk monitoring index sequence and a corresponding risk level, a difference in risk level between any two apparatuses is obtained according to the mixed gaussian distribution, a standard associated total variation and a real-time associated total variation are obtained by utilizing the mixed gaussian distribution, the risk level, the risk probability and the attention degree, a loss function is constructed by utilizing a sum of the standard associated total variation, the real-time associated total variation and a difference in risk level between any two apparatuses, and an apparatus risk detection network is trained by utilizing the loss function. Finally, the risk grade of the equipment can be obtained by inputting the equipment image sequence and the risk monitoring index sequence into the equipment risk detection network, the equipment risk grade evaluation is not required to be carried out manually, meanwhile, the correlation among the equipment in the same production scene is also considered, and the accuracy and the real-time performance of the risk grade prediction are further improved.
It should be noted that: the precedence order of the above embodiments of the present invention is only for description, and does not represent the merits of the embodiments. And specific embodiments thereof have been described above. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing may also be possible or may be advantageous.
The embodiments in the present specification are described in a progressive manner, and the same and similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.
Claims (9)
1. The equipment risk detection method based on metric learning and visual perception is characterized by comprising the following steps of:
inputting a first image of equipment to be tested and a risk monitoring index thereof into a corresponding equipment risk detection network to obtain a corresponding risk prediction result;
the loss function of the equipment risk detection network is the sum of a cross entropy loss function and a mean square error loss function of the associated total variation, and the associated total variation is the difference value of the real-time associated total variation and the standard associated total variation; the standard association total variation is similarity according to probabilities of predicted risk levels of a plurality of devices in device images acquired at the same time period in historical data; the real-time correlation total variation is similarity according to probabilities of predicted risk levels of a plurality of devices in device images acquired at the same time period in real-time data; the similarity is a weighted sum of probability ratios of predicted risk levels between any two of the devices, and the probability of a predicted risk level is obtained from the predicted risk level of a device, the predicted risk probability, a risk fluctuation index, and a difference in risk level between the two devices.
2. The apparatus risk detection method based on metric learning and visual perception according to claim 1, wherein the method for obtaining the risk fluctuation index comprises:
acquiring a risk level sequence corresponding to an equipment image of the equipment, and adopting Gaussian background modeling to acquire mixed Gaussian distribution and a mixed Gaussian distribution curve corresponding to the risk level sequence, wherein the abscissa of the mixed Gaussian distribution curve is a risk level, and the ordinate of the mixed Gaussian distribution curve is a risk probability;
averaging a plurality of the mixed Gaussian distribution curves to obtain a final Gaussian distribution curve; and obtaining the risk fluctuation index of the equipment from the final Gaussian distribution curve.
3. The device risk detection method based on metric learning and visual perception according to claim 2, wherein the method for obtaining the difference of risk level between the two devices comprises:
calculating a distance value between the final Gaussian distribution curves corresponding to any two devices; the distance value is taken as the difference in risk level between the two devices.
4. The method of claim 2, wherein the deriving the risk fluctuation index of the device from the final Gaussian distribution comprises:
the risk fluctuation index calculation formula is as follows:
wherein, BiThe risk fluctuation index for the ith said device; simiIs the gaussian distribution similarity of the ith said device.
5. The method for detecting device risk based on metric learning and visual perception according to claim 4, wherein the method for obtaining the Gaussian distribution similarity of the ith device comprises:
the calculation formula of the gaussian distribution similarity of the ith device is as follows:
wherein f isi(ii) is the final gaussian distribution for the ith said device; argmax (f)i) Is fiTaking the corresponding risk grade when the maximum value is obtained; n (argmax (f)i) 1) taking argmax (f) as the mean valuei) Is normally distributed.
6. The method of claim 1, wherein the standard association total variation is similarity of probabilities of predicted risk levels of a plurality of devices in device images acquired from a same time period in historical data, comprising:
the calculation formula of the standard correlation total variation is as follows:
wherein L is0Correlating the total variation for the criteria; e is the set of all the devices; i, j belongs to the ith device and the jth device of the E set; w is ai，jA difference in the risk level for an ith device and a jth device; giAttention of the ith device; gjAttention of the jth device; fiA risk probability sequence corresponding to a final Gaussian distribution function of the historical data of the ith device; fjA risk probability sequence corresponding to a final Gaussian distribution function of the history data of the jth device; max (F)i) The maximum risk probability in the risk probability sequence of the ith device; argmax (F)i) A risk level corresponding to the maximum risk probability on an ith device; max (F)j) The maximum risk probability in the risk probability sequence of the jth device; argmax (F)j) A risk level corresponding to the maximum risk probability on the jth device.
7. The device risk detection method based on metric learning and visual perception according to claim 6, wherein the attention obtaining method comprises:
the calculation formula of the attention degree is as follows:
wherein, giThe attention of the ith said device; b isiThe risk fluctuation index for the ith said device; b isjThe risk fluctuation index for the jth of the devices; num is the number of all the devices.
8. The method of claim 1, wherein the real-time associated total variation is a similarity of probabilities of predicted risk levels for a plurality of devices in a device image captured from a same time period in real-time data, comprising:
the calculation formula of the real-time total variation of the correlation diagram is as follows:
wherein L isuCorrelating the total variation for the criteria; e is the set of all the devices; i, j belongs to the ith device and the jth device of the E set; w is ai，jA difference in the risk level for an ith device and a jth device; giAttention of the ith device; gjAttention of the jth device; ziA real-time risk probability sequence corresponding to the real-time data of the ith device; zjA real-time risk probability sequence corresponding to the real-time data of the jth device; max (Z)i) The real-time maximum risk probability in the real-time risk probability sequence of the ith device; argmax (Z)i) A real-time risk level corresponding to the real-time maximum risk probability on an ith device; max (Z)j) The real-time maximum risk probability in the real-time risk probability sequence of the jth device; argmax (Z)j) And the real-time risk level corresponding to the real-time maximum risk probability on the jth device.
9. The apparatus risk detection method based on metric learning and visual perception according to claim 1, wherein the loss function of the apparatus risk detection network is a sum of a cross entropy loss function and a mean square error loss function of an associated total variation, comprising:
the construction formula of the loss function is as follows:
wherein M is the number of training data in a training batch of the equipment;
predicting risk probability that the u-th training data output by the risk detection network corresponding to the kth equipment belong to the class c risk level;
belonging to the class c risk level for the u training data in the k equipment; l isuInputting the association graphs obtained by the device risk detection networks into the u training data of the kth device and the training data of other devices at the same position as the u training data to obtain real-time total variation of the association graphs; l is0The standard total variation is the correlation diagram.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110915531.3A CN113610167B (en) | 2021-08-10 | 2021-08-10 | Equipment risk detection method based on metric learning and visual perception |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110915531.3A CN113610167B (en) | 2021-08-10 | 2021-08-10 | Equipment risk detection method based on metric learning and visual perception |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113610167A true CN113610167A (en) | 2021-11-05 |
| CN113610167B CN113610167B (en) | 2022-06-24 |
Family
ID=78340142
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110915531.3A Active CN113610167B (en) | 2021-08-10 | 2021-08-10 | Equipment risk detection method based on metric learning and visual perception |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113610167B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116468928A (en) * | 2022-12-29 | 2023-07-21 | 长春理工大学 | Thermal infrared small target detection method based on visual perception correlator |
Citations (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20140106883A (en) * | 2013-02-27 | 2014-09-04 | 한국전자통신연구원 | Apparatus and method for detecting a risk situation by analyzing a relation of object |
| CN104595170A (en) * | 2014-12-18 | 2015-05-06 | 中国矿业大学 | Air compressor monitoring diagnosis system and method adopting adaptive kernel Gaussian hybrid model |
| CN106462152A (en) * | 2014-04-01 | 2017-02-22 | Tlv有限公司 | Risk evaluation system for process system, risk evaluation program, and risk evaluation method |
| CN107358366A (en) * | 2017-07-20 | 2017-11-17 | 国网辽宁省电力有限公司 | A kind of distribution transformer failure risk monitoring method and system |
| CN107808204A (en) * | 2017-12-14 | 2018-03-16 | 广东电网有限责任公司清远供电局 | A kind of risk management and control system and method to appraisal of equipment data automatic identification |
| CN108921452A (en) * | 2018-07-27 | 2018-11-30 | 国网河北能源技术服务有限公司 | A kind of compound method for early warning of transmission line of electricity risk assessment based on fuzzy algorithmic approach |
| CN110132966A (en) * | 2019-05-14 | 2019-08-16 | 生态环境部卫星环境应用中心 | A kind of cure of Soil pollution spatial position risk evaluating method and system |
| CN110516950A (en) * | 2019-08-21 | 2019-11-29 | 西北工业大学 | A kind of risk analysis method of entity-oriented parsing task |
| CN110677430A (en) * | 2019-10-14 | 2020-01-10 | 西安交通大学 | User risk degree evaluation method and system based on log data of network security equipment |
| CN110751170A (en) * | 2019-09-06 | 2020-02-04 | 武汉精立电子技术有限公司 | Panel quality detection method, system, terminal device and computer readable medium |
| CN110782333A (en) * | 2019-08-26 | 2020-02-11 | 腾讯科技（深圳）有限公司 | Equipment risk control method, device, equipment and medium |
| CN111095345A (en) * | 2018-02-14 | 2020-05-01 | 松下知识产权经营株式会社 | Risk evaluation system and risk evaluation method |
| CN111652496A (en) * | 2020-05-28 | 2020-09-11 | 中国能源建设集团广东省电力设计研究院有限公司 | Operation risk assessment method and device based on network security situation awareness system |
| CN111950942A (en) * | 2020-10-19 | 2020-11-17 | 平安国际智慧城市科技股份有限公司 | Model-based water pollution risk assessment method and device and computer equipment |
| CN112016743A (en) * | 2020-08-24 | 2020-12-01 | 广东电网有限责任公司 | Power grid equipment maintenance prediction method and device, computer equipment and storage medium |
| BR102020018380A2 (en) * | 2020-09-09 | 2021-03-02 | Kingspan Isoeste Construtivos Isotermicos S/A | structural improvement in isothermal panel, facilitator of intervention for installation or maintenance of / in the electrical networks, telephony and hydraulics of a residential building and its manufacturing process |
| CN112446637A (en) * | 2020-12-09 | 2021-03-05 | 张龙 | Building construction quality safety online risk detection method and system |
| CN112906243A (en) * | 2021-03-18 | 2021-06-04 | 长江大学 | Multipoint geostatistical modeling parameter optimization method based on variation function |
| CN113037745A (en) * | 2021-03-06 | 2021-06-25 | 国网河北省电力有限公司信息通信分公司 | Intelligent substation risk early warning system and method based on security situation awareness |
-
2021
- 2021-08-10 CN CN202110915531.3A patent/CN113610167B/en active Active
Patent Citations (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20140106883A (en) * | 2013-02-27 | 2014-09-04 | 한국전자통신연구원 | Apparatus and method for detecting a risk situation by analyzing a relation of object |
| CN106462152A (en) * | 2014-04-01 | 2017-02-22 | Tlv有限公司 | Risk evaluation system for process system, risk evaluation program, and risk evaluation method |
| CN104595170A (en) * | 2014-12-18 | 2015-05-06 | 中国矿业大学 | Air compressor monitoring diagnosis system and method adopting adaptive kernel Gaussian hybrid model |
| CN107358366A (en) * | 2017-07-20 | 2017-11-17 | 国网辽宁省电力有限公司 | A kind of distribution transformer failure risk monitoring method and system |
| CN107808204A (en) * | 2017-12-14 | 2018-03-16 | 广东电网有限责任公司清远供电局 | A kind of risk management and control system and method to appraisal of equipment data automatic identification |
| CN111095345A (en) * | 2018-02-14 | 2020-05-01 | 松下知识产权经营株式会社 | Risk evaluation system and risk evaluation method |
| CN108921452A (en) * | 2018-07-27 | 2018-11-30 | 国网河北能源技术服务有限公司 | A kind of compound method for early warning of transmission line of electricity risk assessment based on fuzzy algorithmic approach |
| CN110132966A (en) * | 2019-05-14 | 2019-08-16 | 生态环境部卫星环境应用中心 | A kind of cure of Soil pollution spatial position risk evaluating method and system |
| CN110516950A (en) * | 2019-08-21 | 2019-11-29 | 西北工业大学 | A kind of risk analysis method of entity-oriented parsing task |
| CN110782333A (en) * | 2019-08-26 | 2020-02-11 | 腾讯科技（深圳）有限公司 | Equipment risk control method, device, equipment and medium |
| CN110751170A (en) * | 2019-09-06 | 2020-02-04 | 武汉精立电子技术有限公司 | Panel quality detection method, system, terminal device and computer readable medium |
| CN110677430A (en) * | 2019-10-14 | 2020-01-10 | 西安交通大学 | User risk degree evaluation method and system based on log data of network security equipment |
| CN111652496A (en) * | 2020-05-28 | 2020-09-11 | 中国能源建设集团广东省电力设计研究院有限公司 | Operation risk assessment method and device based on network security situation awareness system |
| CN112016743A (en) * | 2020-08-24 | 2020-12-01 | 广东电网有限责任公司 | Power grid equipment maintenance prediction method and device, computer equipment and storage medium |
| BR102020018380A2 (en) * | 2020-09-09 | 2021-03-02 | Kingspan Isoeste Construtivos Isotermicos S/A | structural improvement in isothermal panel, facilitator of intervention for installation or maintenance of / in the electrical networks, telephony and hydraulics of a residential building and its manufacturing process |
| CN111950942A (en) * | 2020-10-19 | 2020-11-17 | 平安国际智慧城市科技股份有限公司 | Model-based water pollution risk assessment method and device and computer equipment |
| CN112446637A (en) * | 2020-12-09 | 2021-03-05 | 张龙 | Building construction quality safety online risk detection method and system |
| CN113037745A (en) * | 2021-03-06 | 2021-06-25 | 国网河北省电力有限公司信息通信分公司 | Intelligent substation risk early warning system and method based on security situation awareness |
| CN112906243A (en) * | 2021-03-18 | 2021-06-04 | 长江大学 | Multipoint geostatistical modeling parameter optimization method based on variation function |
Non-Patent Citations (5)
| Title |
|---|
| F A ALAW 等: "Development of risk assessment model for biomass plant boiler using bayesian network", 《ICCEIB 2020》 * |
| 战希臣 等: "基于BP神经网络的维修检测设备研制项目风险评估与仿真", 《中国管理信息化》 * |
| 芦铭: "基于风险分析的医疗设备检查评分系统", 《中国医疗设备》 * |
| 赵煜: "甲醛生产危险因素及预防措施", 《云南化工》 * |
| 黄军凯 等: "基于深度学习的电网一次设备风险智能评估方法", 《电力大数据》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116468928A (en) * | 2022-12-29 | 2023-07-21 | 长春理工大学 | Thermal infrared small target detection method based on visual perception correlator |
| CN116468928B (en) * | 2022-12-29 | 2023-12-19 | 长春理工大学 | Thermal infrared small target detection method based on visual perception correlator |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113610167B (en) | 2022-06-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104200288B (en) | A kind of equipment fault Forecasting Methodology based on dependency relation identification between factor and event | |
| CN115578015B (en) | Sewage treatment whole process supervision method, system and storage medium based on Internet of things | |
| CN112910859B (en) | Internet of things equipment monitoring and early warning method based on C5.0 decision tree and time sequence analysis | |
| CN111680870B (en) | Comprehensive evaluation method for quality of target motion trail | |
| CN109298225B (en) | Automatic identification model system and method for abnormal state of voltage measurement data | |
| CN111813644B (en) | Evaluation method and device for system performance, electronic equipment and computer readable medium | |
| CN116612098B (en) | Insulator RTV spraying quality evaluation method and device based on image processing | |
| CN115580445A (en) | Unknown attack intrusion detection method, device and computer readable storage medium | |
| CN111737826B (en) | Rail transit automatic simulation modeling method and device based on reinforcement learning | |
| CN113610167B (en) | Equipment risk detection method based on metric learning and visual perception | |
| CN116485049B (en) | Electric energy metering error prediction and optimization system based on artificial intelligence | |
| CN116563841B (en) | Detection method and detection device for power distribution network equipment identification plate and electronic equipment | |
| CN113344406A (en) | Power quality monitoring reliability assessment method for intelligent fusion terminal in distribution network area | |
| CN112949735A (en) | Liquid hazardous chemical substance volatile concentration abnormity discovery method based on outlier data mining | |
| CN115616470B (en) | Method, system, equipment and medium for predicting metering error state of current transformer | |
| CN116884192A (en) | Power production operation risk early warning method, system and equipment | |
| CN115618743B (en) | State evaluation method and state evaluation system of sighting telescope system | |
| CN110909774A (en) | Power transformation equipment heating defect reason distinguishing method based on Bayesian classification | |
| CN115147349A (en) | Method and device for determining smelting end point of converter, electronic equipment and storage medium | |
| CN116126807A (en) | Log analysis method and related device | |
| CN105824871B (en) | A kind of picture detection method and equipment | |
| CN112508946B (en) | Cable tunnel anomaly detection method based on antagonistic neural network | |
| CN114492199A (en) | Method and system for analyzing performance of voltage transformer, electronic equipment and medium | |
| CN111835541B (en) | Method, device, equipment and system for detecting aging of flow identification model | |
| CN113958463A (en) | Online monitoring method, system and device for fan blade |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |