Disclosure of Invention
The application aims to provide an early warning method and system for monitoring the running state of target equipment, so as to solve or alleviate the problems in the prior art.
In order to achieve the above purpose, the present application provides the following technical solutions:
an early warning method for monitoring the running state of target equipment comprises the following steps:
step S10, taking each monitoring index of the target equipment at the current moment as input, and calculating to obtain an influence factor of the dynamic optimization BP neural network;
step S20, dynamically optimizing the weight of each neuron of the hidden layer of the BP neural network by adopting the influence factors to obtain an equipment running state early warning model of the BP neural network with dynamically optimized parameters;
step S30, based on the device running state early warning model, according to the acquired input data set of the target device, early warning is carried out on the running state of the target device at the next moment;
the input data set comprises various monitoring indexes of a current moment and a plurality of previous moments, and each monitoring index comprises a monitoring index array value of target equipment.
In the method for monitoring the early warning of the operating state of the target device, optionally, step S10 includes: :
step S11: taking each monitoring index of the target equipment at the current moment as an input variable, carrying out normalization processing on the input variable to obtain an input vector, and calculating membership function values corresponding to each component in the input vector;
step S12, calculating the similarity between the input vector and m data subclasses according to the membership function value, and according to the formula:
the similarity is normalized to obtain the influence factor
In the formula, m is the number of data subclasses in the input data set of the device operation state early warning model, and alpha
j(k) Representing the similarity of the input vector at time k to the ith data subclass of the m data subclasses,
representing the sum of the similarity of the input vector at the moment k and m data subclasses, i and j belonging to [1, m ∈]And m is a positive integer greater than 1.
Optionally, in the method for monitoring the operation state of the target device in the early warning step S11, the input variable is x (k) ═ x (k)1,x(k)2,……,x(k)n]N represents the number of input variables, according to the formula:
the input vector after the normalization processing is obtained as follows:
wherein i is (1, 2, … n), ximaxIs the maximum value of the range of the ith monitoring index, ximinThe minimum value of the monitoring index range of the ith item;
according to a membership function model:
calculating membership function values corresponding to all components in the input vector;
in the formula (I), the compound is shown in the specification,
respectively representing the input vector at the k-th time, the k-1 th time and after the ith normalization processing
And v, c and b are parameters of a membership function model.
Optionally, in step S12, according to the membership function value corresponding to each component in the input vector, the method according to the formula:
obtaining an input vector
Similarity to each data subclass.
The early warning method for monitoring the operating state of the target device optionally includes, according to a formula:
obtaining an input vector
Dynamically optimizing the input of each neuron of a hidden layer of the BP neural network in parameters; in the formula, w
jrRepresenting connection weights input to an r-th neuron belonging to a j-th data subclass in the hidden layer;
representing a bias value of an r-th neuron of a BP neural network hidden layer belonging to a j-th data subclass, wherein r represents the number of neurons belonging to each data subclass in the hidden layer, and r is (1, 2, 3);
for the influence factor, p represents the number of monitoring indexes of the target device at the time k.
In the method for monitoring the early warning of the operating state of the target device, optionally, step S30 includes:
step S31, based on the device running state early warning model, obtaining a running state evaluation predicted value of the target device at the next moment according to the obtained input data set of the target device;
step S32, comparing the operation state evaluation predicted value with a preset operation state threshold value, responding to the fact that the operation state evaluation predicted value is larger than the preset operation state threshold value, and early warning the operation state of the target equipment at the next moment;
in step S31, according to the obtained input data set of the target device, according to the formula:
obtaining the output of a hidden layer of the parameter dynamic optimization BP neural network;
in the formula (I), the compound is shown in the specification,
for input vectors in the input data set
Dynamically optimizing the input of each neuron of a hidden layer of the BP neural network in parameters; f () is a Sigmoid function;
according to the formula:
yo(k)=f(yi(k))
obtaining an estimated value of the running state of the target equipment at the next moment, wherein yi (k) represents the input of the BP neural network output layer, b0Representing the BP neural network output layer bias value.
Optionally, in step S20, before dynamically optimizing the weight of each neuron in the hidden layer of the BP neural network by using the influence factor, the method further includes training, according to a pre-obtained training set and based on an adam optimization algorithm, an equipment operation state early warning model for dynamically optimizing parameters of the BP neural network to obtain the equipment operation state early warning model, and specifically includes:
acquiring a clustering center of the training set based on a K-means method so as to divide the training set into a plurality of data subclasses;
and optimizing parameters of the equipment running state early warning model according to the clustering center of the training set based on the adam optimization algorithm to obtain the equipment running state early warning model.
Optionally, the early warning method for monitoring the operating state of the target device obtains the clustering center of the training set based on the K-means method, so as to divide the training set into a plurality of data subclasses, specifically:
randomly selecting monitoring indexes of m target devices from the training set as initial clustering centers of the training set; the training set comprises a plurality of time sequence sample data, each time sequence sample data comprises a time sequence sample number series value of the target equipment and a sample state evaluation value corresponding to the time sequence sample number series value, and the sample state evaluation value is obtained by manual labeling;
dividing the sample data set into m classes according to the distance from each sample data in each sample data set in the training set to m initial clustering centers respectively, and according to a formula:
calculating the clustering center of each data subclass; in the formula, CtRepresents the clustering center of the t-th subclass of data, t is the [1, m ]]M is a positive integer greater than 1, X represents a set of input vectors in the time series sample data;
according to each time sequence sample data in the training set to each clustering center CtRe-dividing the training set into m classes;
performing loop iteration on the division of the training set until the clustering center CtNo change is made, resulting in m data subclasses.
Optionally, the early warning method for monitoring the operating state of the target device optimizes parameters of the device operating state early warning model according to the clustering center of the training set based on the adam optimization algorithm to obtain the device operating state early warning model, and specifically includes:
an AdamaOptizer optimizer based on a TensorFlow calculation framework performs iterative optimization training on parameters of the equipment running state model according to a data training set and a preset loss function to obtain the equipment running state model;
wherein the loss function is:
wherein l is the number of samples of the training set,i denotes the ith sample in the training set,
is the running state of the target device at the next moment,
indicating that the target device is abnormal,
the target equipment works normally, P is the prediction probability of the sample class, and theta generally refers to the parameter needing to be trained.
The application also provides an early warning system for monitoring the running state of the target equipment, the early warning system comprises an intelligent sensing and control subsystem and a data visual management platform, and the intelligent sensing and control subsystem comprises:
the calculation unit is configured to take various monitoring indexes of the target equipment at the current moment as input and calculate to obtain an influence factor;
the building unit is configured to dynamically optimize the weight of each neuron of the hidden layer of the BP neural network by adopting the influence factors to obtain an equipment running state early warning model of the BP neural network with dynamically optimized parameters;
the early warning unit is configured to perform early warning on the running state of the target equipment at the next moment according to the acquired input data set of the target equipment based on the equipment running state early warning model;
the input data set comprises various monitoring indexes of a current moment and a plurality of previous moments, and each monitoring index comprises a monitoring index array value of target equipment.
Compared with the closest prior art, the technical scheme of the embodiment of the application has the following beneficial effects:
according to the technical scheme, various monitoring indexes of the target equipment at the current moment are used as input, influence factors of a dynamic optimization BP neural network are obtained through calculation, the weight of each neuron of a hidden layer of the BP neural network is dynamically optimized through the influence factors, an equipment running state early warning model of the parameter dynamic optimization BP neural network is obtained, then the equipment running state early warning model of the BP neural network is dynamically optimized based on the parameters, and early warning is conducted on the running state of the target equipment at the next moment according to various monitoring indexes of the target equipment which are collected in advance. According to the technical scheme, the active prediction of the equipment fault condition can be realized, the equipment fault is avoided in time, the operation and maintenance operation of the field equipment can be guided, and a scientific basis is provided for the overhaul of the operation state of the target equipment. In addition, because the influence factors are introduced into the hidden layer of the BP neural network, part of neurons are dynamically inhibited from participating in the operation in the calculation process, so that network parameters are dynamically reduced in the calculation process, the network generalization is improved, the effect of accurately predicting the operation trend of large target equipment in advance is achieved in practical application, and the influence on production and the waste of manpower, material resources and financial resources caused by equipment failure are avoided.
Detailed Description
The technical solutions in the embodiments of the present application will be described clearly and completely below, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all embodiments. All other embodiments that can be derived from the embodiments given herein by a person of ordinary skill in the art are intended to be within the scope of the present disclosure.
The present application will be described in detail with reference to examples. It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict.
Fig. 1 is a signal framework diagram of an early warning system for monitoring an operation state of a target device according to some embodiments of the present disclosure; FIG. 2 is a schematic block diagram of a smart sensing and control subsystem provided in accordance with some embodiments of the present application; fig. 3 is a schematic diagram of a front side configuration of a smart sensing and control subsystem cabinet provided in accordance with some embodiments of the present application; fig. 4 is a schematic view of a cabinet interior layout provided in accordance with some embodiments of the present application; FIG. 5 is a block diagram of a data processing and control unit according to some embodiments of the present application; as shown in fig. 1 to 5, the early warning system for monitoring the operating state of the target device includes an intelligent sensing and control subsystem 100 and a data visualization management platform 300, where a plurality of early warning systems for monitoring the operating state of the device are respectively deployed in each factory building of a factory area or an area where the device needs to be monitored, and the deployment environment may be outdoor or indoor, so that each monitoring index of the target device 400 can be acquired in real time and the operating state of the device can be predicted.
Because the data transmission protocols of different target devices are different, monitoring index data of different target devices cannot be directly uploaded, the PLC is used as a relay in the application, the PLC is in communication connection with a plurality of different target devices through different industrial data transmission protocols, and then the PLC uploads the received monitoring index data of different target devices to the data processing and Control unit through an OPC (object Linking and embedding) for Process Control) ua (unified protocol, so that simultaneous monitoring of a plurality of different target devices is realized.
As shown in fig. 1 to 4, the intelligent sensing and control subsystem 100 includes a cabinet 101, a data processing and control unit 103, a PLC106, and an industrial switch 104, and the data processing and control unit 103, the PLC106, and the industrial switch 104 are all disposed in the cabinet.
Wherein, the data processing and controlling unit 103 is in communication connection with the data visualization management platform 300; the PLC106 is in communication connection with the data processing and control unit 103 through the industrial switch 104; the PLC106 is in communication connection with a plurality of target devices 400 through different industrial data transmission protocols; the PLC106 collects monitoring index data of each target device 400, and uploads the collected monitoring index data to the data processing and control unit 103, and the data processing and control unit 103 determines whether to generate warning information according to the received monitoring index data, and sends the warning information to the data visualization management platform 300 after generating the warning information.
The early warning system can realize active prediction of the fault condition of the target equipment, avoid equipment faults in time, and enable a user to timely and accurately master the working state of the target equipment so as to carry out different adjustments on the target equipment and ensure efficient and safe working of the target equipment 400.
It should be noted that, the PLC106 uploads the collected monitoring index data to the data processing and control unit 103 in a unified manner according to the OPCUA protocol, and the data processing and control unit 103 may also download a control command to the PLC106 according to the OPCUA protocol, and then sends the control command to the field target device 400 through the PLC 106.
The data processing and control unit 103 predicts whether the target device 400 needs to be warned by a preset warning technology, uploads the data to the data visualization management platform 300 deployed on the server 200 in a wireless or wired network through a Message Queue Telemetry Transport (MQTT) standard data format, and notifies a relevant responsible person of intelligent warning of a device fault by short Message through the 4G module 1034 of the E9510 when a monitored value exceeds a preset safety threshold value. The preset early warning technique may be a threshold method in the prior art. In other implementations, the data processing and control unit 103 may further dynamically optimize a device operating state early warning model of a BP (Back prediction, error reversal) neural network through preset parameters, predict the probability of a future failure of the target device 400, upload the predicted probability data and the device parameter monitoring data to the data visualization management platform 300 deployed on the server 200 through a message queue telemetry transmission standard data format in a wireless or wired network as two different subjects of real-time data and early warning information, and notify the relevant responsible person of the device failure intelligent early warning through a short message through the 4G module 1034 of the E9510 when the device failure prediction probability value exceeds a preset value (e.g., 0.5).
The data visualization management platform 300 can be deployed in the factory local area network server 200 or on the public network cloud server 200, the data visualization management platform 300 adopts a B/S architecture, and the main functions are as follows: (1) monitoring in real time; (2) inquiring a report form; (3) managing multiple users; (4) user management & rights management; (5) and managing equipment fault warning.
As shown in fig. 4, in an alternative embodiment of the present application, a partition 1011 is disposed inside the cabinet 101, the partition 1011 extends along the front-back direction of the cabinet 101, and divides the cabinet 101 into an upper cabinet and a lower cabinet, and the data processing and control unit 103, the PLC106 and the industrial switch 104 are disposed inside the upper cabinet.
It should be noted that the cabinet body 101 surface adopts pickling, bonderizing, plastic-blasting, high temperature to toast key processing technology and guarantees that the box can not rust, and the bottom installation supporting legs, it is convenient to remove, settles firmly.
In an optional embodiment of the present application, the intelligent sensing and control subsystem 100 further includes an Uninterruptible Power Supply (UPS) 109 and a dc Power Supply 102, the UPS109 is disposed in the lower cabinet 101, and the UPS109 is externally connected to a 220V ac Power Supply and connected to the dc Power Supply 102; the dc power supply 102 is connected to the data processing and control unit 103, the PLC106, and the industrial switch 104, and supplies a dc 24V power to the data processing and control unit 103, the PLC106, and the industrial switch 104. In this embodiment, the UPS109 functions to convert ac and dc, and may convert 220V ac power to 24V dc power.
In an optional embodiment of the present application, the PLC106 is mounted with a PLC communication module 107, and is configured to access monitoring index data of target devices 400 with different communication protocols;
in an optional embodiment of the present application, the intelligent sensing and control subsystem 100 further includes a connection terminal 108, the connection terminal 108 is disposed in the upper cabinet, and the signal of the target device 400 is accessed to the PLC communication module 107 through the connection terminal 108.
As shown in fig. 4, in an alternative embodiment of the present application, a back plate 1012 is fixed on an inner wall surface of the upper cabinet, the PLC106, the industrial switch 104, the dc power supply 102 and the connection terminal 108 are all fixed on a surface of the back plate 1012 facing away from the inner wall surface of the upper cabinet, a wire trough 1013 for routing communication and power lines is further formed on the surface of the back plate 1012 facing away from the inner wall surface of the upper cabinet, and the data processing and control unit 103 is disposed on the partition 1011. The direct current power supply 102, the industrial switch 104, the PLC106 and the wiring terminal 108 are all hung on a back panel 1012; an industrial switch 104 and a PLC106 are respectively arranged on two sides of the direct current power supply 102, and a wiring terminal 108 is arranged under the direct current power supply 102, the industrial switch 104 and the PLC 106; cabling channels 1013 are routed between and around the dc power source 102, the industrial switch 104, the PLC106, and the wiring terminals 108. The data processing and control unit 103 is disposed on the upper surface of the partition 1011 and directly below the back plate 1012. Therefore, not only the layout of all parts in the upper cabinet body is reasonable, but also the internal space is fully utilized; meanwhile, the strong current part and the weak current part are subjected to isolation control, so that the safety performance of the equipment is effectively improved.
In an optional embodiment of the present application, the smart sensing and control subsystem 100 further includes an industrial touch screen 105, and the industrial touch screen 105 is in communication connection with the data processing and control unit 103 through the industrial switch 104, and in communication connection with the PLC106 through an ethernet protocol, and is configured to display the monitoring variable of the target device 400 or control the control variable in the PLC 106. Specifically, the monitoring variables of the target device 400 can be displayed in the configuration picture of the industrial touch screen 105 in real time, and the start-stop button in the configuration picture of the industrial touch screen 105 is clicked to control the control variables in the PLC106, so that the target device 400 is started and stopped on the field side.
In an optional embodiment of the present application, the cabinet body 101 is a rectangular parallelepiped, the cabinet body 101 has one side opening (i.e. a front opening), the rotating door is assembled at the front opening, that is, the rotating door is rotatably connected to the opening side of the cabinet body 101, and can cover or open the opening, the surface of the rotating door facing away from the opening is provided with a mounting groove, the size of the mounting groove is matched with the size of the industrial touch screen 105, the industrial touch screen 105 is embedded in the mounting groove, and thus, the installation operation of the industrial touch screen 105 can be realized.
In the specific embodiment of the application, the rotating door is provided with a connecting side and a free side which are opposite, the connecting side is rotatably connected with one opening side of the cabinet body, and the free side and the rotating door are abutted against the other opening side of the cabinet body when being covered; the surface of the rotating door back to the opening is also provided with a door lock, and the door lock is positioned at the lower part of the industrial touch screen and is arranged close to the free side of the rotating door. It should be noted that the connection side of the revolving door and the cabinet body may be connected by a hinge, a shaft hole or other reasonable and effective connection means. In addition, the door lock may be a mechanical lock or an electronic lock, which is not limited herein and is within the protection scope of the present invention.
In the optional embodiment of this application, the cabinet body 101 is the cuboid form, and the bottom of cabinet body 101 is equipped with a plurality of supporting legss, and a plurality of supporting legss are at the bottom of cabinet body 101 along circumference equipartition. Specifically, the supporting legs has 4, and 4 supporting legs correspond respectively and install in the four corners of the cabinet body, and the bottom of each supporting leg all includes universal wheel, locating part and adjusting part, and the adjusting part adjusts the height of every supporting leg through spiral pivoted mode, makes a plurality of supporting legs be in same horizontal position, and then, the slope of wall cabinet body 101 improves cabinet body 101 stability. The locating part is used for restricting the rotation of universal wheel through modes such as joint after the cabinet body 101 is in the horizontal position to it is fixed with cabinet 101, avoids the continuation of the cabinet body 101 to remove. It should be noted that, the limiting element may be a conventional limiting element, and the adjusting element may be a conventional adjusting element, which is not limited herein and is within the protection scope of the present application.
In an optional embodiment of the present application, the data processing and controlling unit 103 includes a microcontroller 1032, a communication module 1033, and a storage module 1031, which are in communication connection, and the communication module 1033 performs information interaction with the data visualization management platform 300, and can send out warning information to the data visualization management platform 300.
The microcontroller 1032 adopts a processor based on an Intel Haswell micro-architecture, the design power consumption is only 15 watts, and the excellent balance of performance, power consumption and volume is achieved. Microcontroller 1032 does not have fan passive heat dissipation, can run wide temperature, the operating temperature is-20 deg.C-70 deg.C, the applicable frequency of mechanical vibration is 10 Hz-150 Hz.
In an optional embodiment of the present application, the communication module 1033 includes a 4G module 1034, a WIFI module 1035, and an ethernet port 1036, and the data processing and control unit 103 can send out the warning information through the 4G module 1034; the industrial switch 104 accesses the ethernet port 1036 to enable information interaction between the data processing and control unit 103 and the PLC 106. Wherein the 4G module 1034 has a model number E9510,
further, the intelligent sensing and control subsystem 100 further includes an antenna 110, the antenna 110 extends out of the top end of the cabinet 101, and the 4G module 1034 and the WIFI module 1035 are both connected to the antenna 110 through signal lines for information transmission.
Therefore, the real-time performance of data acquisition and early warning information is further improved, a user can timely and accurately master the working state of each target device 400, different adjustments can be conveniently carried out on different target devices 400, and efficient and safe work of each target device 400 is guaranteed.
At present, the application of complex aluminum, magnesium and titanium light alloy castings is mainly concentrated in the fields of automobiles, rail transit, medical appliances and the like. The working environment of the casting shop is complex, and the prediction for monitoring the running state of the target equipment is a prediction with multiple fault factors and a complex nonlinear mapping relation. Therefore, the application also provides an early warning method for monitoring the running state of the complex equipment, and the early warning method for monitoring the running state of the target equipment provided by the application is explained in detail below by taking the surface drying furnace equipment as an example.
Fig. 6 is a schematic flow chart of an early warning method for monitoring an operating state of a target device according to some embodiments of the present application; as shown in fig. 6, in the embodiment of the present application, the early warning method includes the following steps:
step S10, using each monitoring index of the target device at the current time as input, calculating to obtain the influence factor of dynamic optimization BP neural network
Step S20, dynamically optimizing the weight of each neuron of the hidden layer of the BP neural network by adopting the influence factors to obtain an equipment running state early warning model of the BP neural network with dynamically optimized parameters;
step S30, based on the device running state early warning model, according to the acquired input data set of the target device 400, early warning is carried out on the running state of the target device at the next moment;
the input data set includes monitoring indexes of the current time and a plurality of previous times, and each monitoring index includes a monitoring index array value of the target device 400.
Specifically, the PLC106 acquires each monitoring index of the target device 400 at the current time and a plurality of previous times in advance to obtain a monitoring index array value of the target device 400, an operation state evaluation prediction value corresponding to the monitoring index array value at the current time is obtained by predicting a trained device operation state early warning model, and when predicting an operation state evaluation value of the target device 400 at the next time, the operation state evaluation prediction value corresponding to the current time is the operation state evaluation value at the current time.
The PLC106 uploads the data collected by different industrial data transmission protocols to the data processing and control unit 103 in an OPCUA protocol, the data processing and control unit 103 dynamically optimizes a device operating state early warning model of the BP neural network according to preset parameters, predicts the probability of a future failure of the target device 400, uploads the predicted probability data and device parameter monitoring data to the data visualization management platform 300 deployed on the server 200 in a wireless or wired network according to two different topics, namely, real-time data and early warning information, through a Message Queue Telemetry Transport (MQTT) standard data format, and notifies a relevant responsible person of the device failure intelligent early warning through a 4G module 1034 of E9510 in a short Message when the device failure prediction probability value exceeds a preset value (for example, 0.5).
Fig. 7 is a schematic flowchart of step S10 in the warning method for monitoring the operating state of the target device according to some embodiments of the present application; as shown in fig. 7, step S10 includes:
step S11: and taking each monitoring index of the target equipment at the current moment as an input variable of the equipment running state early warning model, carrying out normalization processing on the input variable to obtain an input vector, and calculating membership function values corresponding to each component in the input vector.
Specifically, the input variable is each monitoring index of the target device at time k, and the input variable is x (k) ═ x (k)1,x(k)2,……,x(k)n]N tableThe number of input variables is shown, and in the embodiment of the operation health degree of the dry-oven equipment, the number n of the input variables is selected to be 4, as shown in table 1.
TABLE 1 input variables for an example of the health of operation of a kiln plant
Normalizing the input variables, mapping all input variables into an interval of [ -1, 1], and normalizing the input variables according to a formula (1) to obtain normalized input vectors; equation (1) is as follows:
the input vector after normalization processing is:
where n is 4;
in formula (1), i is (1, 2, … n), ximaxIs the maximum value of the range of the ith monitoring index, ximinThe minimum value of the range of the monitoring index of the ith item.
Calculating the components of the input vector according to equation (2)
The corresponding membership function value; equation (2) is as follows:
in the formula (I), the compound is shown in the specification,
respectively representing the kth time, the kth-1 time and the ith input vector
And its corresponding jth membership function value; j is (1, 2, … m), j represents the j-th data subclass of m data subclasses divided according to different characteristics of data training samples, wherein the input data set is set as 4 data subclasses, m is 4, v, c and b are all parameters of a membership function model, before model training, the initial value of c is the value of m data subclass class centers after clustering, and the initial values of v and b are random values with the value of normal truncation distribution with the standard deviation of 0.1.
And step S12, calculating the similarity between the input vector and the m data subclasses according to the membership function value, and performing normalization processing to obtain an influence factor.
Specifically, according to formula (3):
obtaining an input vector
Similarity to each data subclass (1, 2, …, m data subclasses); according to formula (4):
will be provided with
Normalizing the similarity of each data subclass to obtain an influence factor
In formula (4), m is the number of data subclasses in the input data set of the device operation state early warning modelAmount, α
j(k) Representing the similarity of the input vector at time k to the ith data subclass of the m data subclasses,
representing the sum of the similarity of the input vector at the moment k and m data subclasses, i and j belonging to [1, m ∈]And m is a positive integer greater than 1. It should be noted that, in the following description,
and representing the influence factor of the j-th data subclass in the m data subclasses at the k time, wherein the influence factor represents the influence on the hidden layer neuron caused by the similarity between the input vector and the j-th data subclass.
In this embodiment of the present application, in step S20, the number of hidden layer neurons of the BP neural network is 3m, where in this embodiment, m is 4, the number of neurons in the hidden layer is 12, each of the m data subclasses has 3 hidden layer neurons belonging to this class, and the weights of these 3 neurons are subject to the hidden layer neurons belonging to this class
The degree of inhibition of these 3 neurons involved in the calculation process dynamically changes. The output result calculation process of the weight dynamic optimization feedforward neural network is as follows:
according to equation (5):
obtaining an input vector
Dynamically optimizing the calculation result input by each neuron of a hidden layer of the BP neural network in parameters; in the formula (5), w
jrRepresenting connection weights from the input (in) to an r-th neuron belonging to a j-th data subclass at the hidden layer (hide);
representing a bias value of an r-th neuron of a BP neural network hidden layer belonging to a j-th data subclass, wherein r represents the number of neurons belonging to each data subclass in the hidden layer, and r is (1, 2, 3);
for the influence factor, p represents the number of monitoring indexes of the target device at the time k.
Fig. 8 is a schematic flowchart illustrating a detailed process of step S30 in the method for warning about the operating condition of a device according to another embodiment of the present application, where, as shown in fig. 8, step S30 includes:
step S31, based on the device running state early warning model, obtaining a running state evaluation predicted value of the target device 400 at the next moment according to the obtained input data set of the target device; the operation state evaluation predicted value is used for representing the operation state of the target device 400 at the next moment;
and step S32, comparing the operation state evaluation predicted value with a preset operation state threshold value, responding to the fact that the operation state evaluation predicted value is larger than the preset operation state threshold value, and early warning the operation state of the target equipment at the next moment.
In step S31, the following equation (6) is used:
obtaining the output of a hidden layer of the parameter dynamic optimization BP neural network;
in formula (6), f (×) is a Sigmoid function.
According to formula (7) and formula (8):
yo(k)=f(yi(k)) (8)
and obtaining an operation state evaluation value of the target equipment at the next moment, wherein,yi (k) represents the input of the output layer of the BP neural network, b0Representing the BP neural network output layer bias value.
In the specific implementation of the present application, the device operating state table is shown in table 2.
TABLE 2 running state table of equipment
As can be seen from Table 2, when yo (k) is greater than 0.5, the intelligent sensing and control subsystem sends out early warning information that the equipment is about to fail to the data visualization management platform.
According to the method, various monitoring indexes of the target equipment at the current moment are used as input, influence factors are obtained through calculation, the weight of each neuron of a hidden layer of a BP neural network is dynamically optimized through the influence factors, an equipment running state early warning model with parameters dynamically optimized for the BP neural network is constructed, then the equipment running state early warning model of the BP neural network is dynamically optimized based on the constructed parameters, the running state of the target equipment at the next moment is predicted according to various monitoring indexes of the target equipment collected in advance, and the running state evaluation predicted value of the target equipment at the next moment is obtained; then, the running state evaluation predicted value is compared with a preset running state threshold value, so that early warning on the running state of the target equipment at the next moment is realized; and when the estimated predicted value of the running state is larger than the preset running state threshold value, early warning the running state of the target equipment at the next moment. According to the technical scheme, the active prediction of the equipment fault condition can be realized, the equipment fault is avoided in time, the operation and maintenance operation of the field equipment can be guided, and a scientific basis is provided for equipment state maintenance. According to the method, the influence factors are introduced into the hidden layer of the BP neural network, so that part of neurons are dynamically inhibited from participating in operation in the calculation process, network parameters are dynamically reduced in the calculation process, the network generalization is improved, the effect of accurately predicting the operation trend of large target equipment in advance is achieved in practical application, and the influence on production and the waste of manpower, material resources and financial resources caused by equipment faults are avoided.
In this embodiment of the present application, before performing dynamic optimization on the weight of each neuron of the hidden layer of the BP neural network by using the influence factor in step S20, the method further includes, according to a pre-obtained training set, training an equipment operation state early warning model of the BP neural network dynamically optimized by using parameters based on an adam optimization algorithm, to obtain the equipment operation state early warning model, which specifically is:
(1) and acquiring the clustering center of the training set based on a K-means method so as to divide the training set into a plurality of data subclasses.
Specifically, firstly, a batch of collected target equipment monitoring parameter historical records and equipment operation records are screened to form a training set,
X
lshowing various monitoring indexes of the target equipment at a certain moment,
is the running state that the target equipment is to enter, which is judged by equipment running records and relevant experts;
it is indicated that an abnormality has occurred in the device,
indicating that the device is working properly.
Then, randomly selecting detection index data of m target devices from the training set D as an initial clustering center of the training set D; the training set D comprises a plurality of time sequence sample data, each time sequence sample data comprises a time sequence sample number series value of the target equipment and a sample state evaluation value corresponding to the time sequence sample number series value, and the sample state evaluation value is obtained through manual labeling.
Then, according to the distance from each sample data in each sample data set in the training set D to m initial clustering centers, dividing the sample data set into m classes, and according to the formula (9)
Calculating the clustering center of each data subclass; in the formula, CtRepresents the clustering center of the t-th class, t ∈ [1, m]M is a positive integer greater than 1, X represents a set of input vectors in the time series sample data;
according to each time sequence sample data in the training set, obtaining a clustering center CtThe training set is subdivided into m classes;
performing loop iteration on the division of the training set until the clustering center CtNo change is made, resulting in m data subclasses.
(2) And optimizing parameters of the equipment running state early warning model according to the clustering center of the training set based on the adam optimization algorithm to obtain the equipment running state early warning model.
Specifically, based on an AdamaOptizer optimizer, according to a data training set, carrying out iterative optimization training on parameters of the equipment running state model according to a preset loss function to obtain an equipment running state model;
wherein the predetermined loss function is:
wherein l is the number of samples in the training set, i represents the ith sample in the training set,
for the operational state to be entered by the target device,
indicating that the target device is abnormal,
the target equipment works normally, P is the prediction probability of the sample class, and theta generally refers to the parameter needing to be trained.
Wherein, the center C is clusteredtAs membership function model:
μ(k)=exp[-(x+v×μ(k-1)-c)2/2b2]
initial value of c, initial value of other BP neural network
b
0The value is a random value of a truncated normal distribution with a standard deviation of 0.1, and the truncated normal distribution is limited on the basis of a standard normal distribution (gaussian distribution) so as to obtain that the generated data is within a certain range, for example, the range of the data generated by the standard normal distribution is from negative infinity to positive infinity, and the range of the data generated by the truncated normal distribution is (standard deviation of mean-2 times, standard deviation of mean +2 times). In general, w
jr,
b
0Is initially set at [ -0.2, 0.2 [)]With random numbers in a normal distribution.
Further, AdamaOptizer is selected as a parameter optimizer, and the training weight is 0.00005; the maximum iteration number is set to 1000, the purpose of minimizing J (theta) (approaching to 0) is achieved, and further the optimized values of the parameters v, b and c in the membership function model can be obtained and used in practical application.
Fig. 9 is a functional block diagram of an intelligent sensing and control subsystem in an early warning system for monitoring an operating state of a target device according to some embodiments of the present application, where as shown in fig. 9, the intelligent sensing and control subsystem includes a computing unit, a building unit, and an early warning unit. The calculation unit is configured to take various monitoring indexes of the target equipment at the current moment as input and calculate to obtain an influence factor; the building unit is configured to dynamically optimize the weight of each neuron of the hidden layer of the BP neural network by adopting the influence factors to obtain an equipment running state early warning model of the BP neural network with dynamically optimized parameters; the early warning unit is configured to perform early warning on the running state of the target equipment at the next moment according to the acquired input data set of the target equipment based on the equipment running state early warning model;
the input data set comprises various monitoring indexes of a current moment and a plurality of previous moments, and each monitoring index comprises a monitoring index array value of target equipment.
The early warning system for monitoring the operating state of the target device, provided by the embodiment of the application, can implement the processes and steps of the early warning method for monitoring the operating state of the target device in any embodiment, and achieve the same beneficial effects, which are not repeated herein.
The above description is only a preferred embodiment of the present application and is not intended to limit the present application, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the present application shall be included in the protection scope of the present application.