CN108569607B - Elevator fault early warning method based on bidirectional gating cyclic neural network - Google Patents
Elevator fault early warning method based on bidirectional gating cyclic neural network Download PDFInfo
- Publication number
- CN108569607B CN108569607B CN201810647751.0A CN201810647751A CN108569607B CN 108569607 B CN108569607 B CN 108569607B CN 201810647751 A CN201810647751 A CN 201810647751A CN 108569607 B CN108569607 B CN 108569607B
- Authority
- CN
- China
- Prior art keywords
- neural network
- layer
- elevator
- bidirectional
- cyclic neural
- 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.)
- Active
Links
Classifications
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B66—HOISTING; LIFTING; HAULING
- B66B—ELEVATORS; ESCALATORS OR MOVING WALKWAYS
- B66B5/00—Applications of checking, fault-correcting, or safety devices in elevators
- B66B5/0006—Monitoring devices or performance analysers
- B66B5/0018—Devices monitoring the operating condition of the elevator system
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B66—HOISTING; LIFTING; HAULING
- B66B—ELEVATORS; ESCALATORS OR MOVING WALKWAYS
- B66B1/00—Control systems of elevators in general
- B66B1/02—Control systems without regulation, i.e. without retroactive action
- B66B1/06—Control systems without regulation, i.e. without retroactive action electric
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B66—HOISTING; LIFTING; HAULING
- B66B—ELEVATORS; ESCALATORS OR MOVING WALKWAYS
- B66B3/00—Applications of devices for indicating or signalling operating conditions of elevators
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B66—HOISTING; LIFTING; HAULING
- B66B—ELEVATORS; ESCALATORS OR MOVING WALKWAYS
- B66B5/00—Applications of checking, fault-correcting, or safety devices in elevators
- B66B5/02—Applications of checking, fault-correcting, or safety devices in elevators responsive to abnormal operating conditions
- B66B5/027—Applications of checking, fault-correcting, or safety devices in elevators responsive to abnormal operating conditions to permit passengers to leave an elevator car in case of failure, e.g. moving the car to a reference floor or unlocking the door
Abstract
The invention discloses an elevator fault prediction method based on a bidirectional gated cyclic neural network, which comprises the following steps: 1) sampling elevator vibration waveform data; 2) converting the elevator vibration waveform sample into a sequence form; 3) dividing the serialized samples into a training set and a testing set; 4) constructing a bidirectional gating cyclic neural network framework; 5) training a bidirectional gating cyclic neural network framework; 6) performing prediction test by using the test set to obtain a bidirectional gating cyclic neural network prediction model; 7) and (4) performing elevator fault prediction classification by using a bidirectional gated cyclic neural network prediction model to obtain a predicted elevator waveform result. The invention also discloses an elevator fault diagnosis method based on the bidirectional gated recurrent neural network, and the judgment result shows that the elevator is in a fault state and then an alarm is sent. The two methods of the invention have extremely high accuracy and real-time performance for diagnosing and predicting the elevator condition.
Description
Technical Field
The invention belongs to the technical field of elevator fault detection and early warning, relates to an elevator fault prediction method based on a bidirectional gated cyclic neural network, and further relates to an elevator fault diagnosis method based on the bidirectional gated cyclic neural network.
Background
In recent years, with the continuous development of social economy, the number of urban high-rise buildings is rapidly increased, the application of elevators is wider, the holding amount of the elevators is continuously increased, and the space of elevator updating markets and after-sale service markets is huge in the future. However, in recent years, elevator faults are frequent, and on one hand, the high accident rate and the high severity degree often cause serious casualties and large economic loss; on the other hand, at present, a plurality of maintenance and repair units have good and irregular storage of the running condition and the maintenance record of the elevator, so that a plurality of real data are difficult to be examined and a plurality of potential safety hazards are buried for the running of the elevator.
Therefore, how to effectively monitor the safe operation of the elevator and timely eliminate various elevator fault hidden dangers, the safety problem of the elevator is fundamentally solved, the safe use and operation of the elevator are guaranteed, and the elevator safety monitoring system has very important significance. At present, a relatively common elevator fault detection method is to detect in real time through an elevator fault monitoring system. The elevator fault monitoring system integrates data acquisition, communication, data analysis, fault diagnosis and computer control, collects the running state data of the elevator through the data acquisition equipment installed on site, then transmits the data to the monitoring center through the communication network, and realizes the monitoring of the running state of the elevator, the fault diagnosis, the real-time alarm and the like through analysis and processing, thereby providing guarantee for the safety of the elevator.
However, the current elevator fault monitoring system has two fatal defects which cannot be avoided: firstly, the premonitory characteristics of the elevator before failure cannot be known in time, and the elevator cannot be judged and processed in time (in advance). The existing elevator fault monitoring system is more used for real-time alarm, various losses caused by the fault are reduced as far as possible after the fault occurs, the safety factor of the elevator in operation cannot be improved, the people trapping rate and the accident rate are reduced, and a plurality of adverse effects and negative emotions are inevitably generated; secondly, the elevator fault monitoring system can only give out whether an elevator is good or not and whether the elevator needs to be overhauled, but cannot give out the current damage degree of the elevator systematically. Such black box designs often fail to provide passengers with an intuitive elevator quality experience, and also often fail to alert maintenance personnel in advance.
Compared with the traditional elevator monitoring system, no matter whether the fault reason is quickly confirmed or the fault is found in time, the mode that the fault type is checked after rescue is avoided by first-out fault. The passive rescue mode of the existing elevator monitoring system wastes a large amount of manpower and material resources, and cannot reduce the existing accident rate, so that the traditional method can treat the symptoms and the root causes, and a new breakthrough is urgently needed. Therefore, there is an urgent need to develop a technology capable of predicting and urgently braking before a failure occurs, and simultaneously, feeding back a health condition in a normal operation.
Disclosure of Invention
The invention aims to provide an elevator fault prediction method based on a bidirectional gated cyclic neural network, which solves the problems that the occurrence of a fault cannot be predicted in advance, only passive rescue is needed, the fault is eliminated afterwards, and the health condition of an elevator cannot be reflected in real time in the prior art.
The invention also aims to provide an elevator fault diagnosis method based on the bidirectional gated recurrent neural network, which can judge the fault type of the elevator about to have faults and give an alarm in advance.
The technical scheme adopted by the invention is that an elevator fault prediction method based on a bidirectional gated recurrent neural network is implemented according to the following steps:
step 1: sampling elevator vibration waveform data, and converting the elevator vibration waveform data into a sample sequence which is sorted according to the collection time and takes the time sequence as a reference;
step 2: converting the elevator vibration waveform sample into a sequence form, wherein the sequence comprises two parts, the first part is an input signal sequence, and the second part is an output prediction sequence;
the signal sequence and the prediction sequence of the training sample are both taken from a sample sequence, wherein the difference between the signal sequence and the prediction sequence is p time units, the t-time data in the signal sequence is the t-p time data of the prediction sequence, the shapes of the signal sequence and the prediction sequence are the same, and the matrix shapes are [ the number of samples, the step length and the output dimension ];
and step 3: dividing the serialized samples into a training set and a testing set, wherein the training set data accounts for 70% of the total samples, and the testing set data accounts for 30% of the total samples;
and 4, step 4: constructing a bidirectional gate-controlled cyclic neural network framework,
the device comprises three parts, wherein the first part is an input part, the second part is an implicit layer part, the third part is an output layer part, and the input layer part only comprises one input layer; the hidden layer part comprises a plurality of hidden layers, and each hidden layer comprises a bidirectional gating recurrent neural network layer, a full connection layer and a discarding layer; the output layer only comprises one full connection layer; the prediction of the sequence is output through the last full-connection layer through the trained bidirectional gated cyclic neural network,
the other neural network layers except the first input layer are all linked with the previous neural network layer through an activation function, and output data of each layer are subjected to normalized processing to obtain a bidirectional gated cyclic neural network framework;
and 5: training the bidirectional gated cyclic neural network framework, traversing each training data in the training set every time, wherein each traversal is called a generation, so that the bidirectional gated cyclic neural network framework is subjected to multiple generation training, and a loss function is used for outputting a loss rate in the training process;
obtaining the optimal neural network framework parameters after training for a plurality of generations;
step 6: the test set is used for carrying out prediction test, if the difference between the predicted waveform of the test set and the waveform of the test set is too large, the parameter is incorrect, and if the phenomenon occurs, the hyper-parameter needs to be adjusted; then, training in the mode of the step 5 to obtain a bidirectional gating cyclic neural network prediction model;
and 7: elevator fault prediction is carried out by using a bidirectional gated cyclic neural network prediction model,
predicting a time unit each time, namely cycling the prediction, taking the predicted value and the previous 255 values each time to form an input, namely predicting a new value through 256 values each time, and cycling 256 times to predict a complete elevator waveform sequence;
and classifying the well-trained bidirectional gated recurrent neural network prediction model before calling the elevator waveform sequence so as to obtain a predicted elevator waveform result.
The invention adopts another technical scheme that an elevator fault diagnosis method based on a bidirectional gating cyclic neural network is implemented according to the following steps:
step 1: classifying and defining samples in an elevator vibration database by experts according to fault types contained in the signals to finish sample labeling;
step 2: converting the samples after the classification definition into a sample sequence,
the sample sequence comprises two parts, wherein the first part is a signal sequence, and the sampling interval is correspondingly determined according to the setting of sampling time; the second part is a label sequence, and the label sequence marks the classification of the corresponding signal sequence;
converting the signal sequence data into a matrix form, wherein the matrix shape of the signal sequence is [ the number of sequence samples, the step length and the input data dimension ]; meanwhile, converting the label sequence data corresponding to the signal sequence into a matrix form, wherein the matrix shape of the label sequence is [ the number of sequence samples, the dimension of the output label ];
and step 3: dividing the serialized samples into a training set and a testing set,
wherein the training set data accounts for 60% of the total sample, the test set data accounts for 30% of the total sample, and the cross validation set accounts for 10%;
and 4, step 4: constructing a bidirectional gate-controlled cyclic neural network framework,
the bidirectional gating cyclic neural network framework comprises three parts, wherein the first part is an input layer part, the second part is an implicit layer part, and the third part is an output layer part; the input layer part only comprises one input layer; the hidden layer part comprises a plurality of layers including a bidirectional gating cyclic neural network layer, a full connection layer and a discarding layer; the output layer part only comprises a Soft-Max layer,
the trained bidirectional gated cyclic neural network outputs a judgment result of a sequence through a last Soft-Max layer, wherein the rest bidirectional gated cyclic neural network layers except the input layer are all linked with the previous bidirectional gated cyclic neural network layer through an activation function;
the output data of each layer is subjected to normalization processing by using batch normalization;
and 5: training the constructed bidirectional gated cyclic neural network framework, traversing each training data in the training set each time, wherein each traversal is called a generation, and enabling the bidirectional gated cyclic neural network framework to perform a plurality of generations of training;
testing a plurality of generations by using the data in the test set to obtain the data accuracy rate, and outputting the loss rate by using a loss function; after training for a plurality of generations, obtaining an optimal bidirectional gating recurrent neural network diagnosis model;
step 6: and (4) judging the result of over-fitting,
performing overfitting test by using a cross validation set, wherein if the accuracy is greatly reduced, an overfitting phenomenon occurs, and if the overfitting phenomenon occurs, the hyper-parameters are adjusted;
if the number of layers of the bidirectional gated cyclic neural network is reduced, training needs to be carried out again through the step 5, and a bidirectional gated cyclic neural network diagnosis model is obtained;
and 7: performing elevator fault judgment on the elevator vibration waveform by using a bidirectional gated cyclic neural network diagnosis model, wherein the judgment standard is related to the expert marking in the step 1 and conforms to the national or industry related standard, and extracting a judgment result from the last Soft-Max layer;
and 8: and outputting a judgment result, wherein the judgment result indicates that the elevator is in a fault state and then sends an alarm.
The beneficial effects of the invention are that the invention comprises the following aspects:
1) the vibration waveform detection model based on the bidirectional gated cyclic neural network is based on a large amount of elevator fault waveform data, the health condition of an elevator can be monitored in real time through training, the waveform data is input in real time through a sensor, the existence of faults and fault types can be rapidly and accurately detected and judged, elevator faults possibly occurring in a short period can be predicted according to the current elevator waveform data, passengers can be evacuated by stopping nearby, preventive work such as maintenance is informed, the elevator fault rate is greatly reduced, and economic loss and adverse effects are reduced; meanwhile, the model can also provide possible fault types and health conditions of the elevator in real time, and an elevator black box is broken.
2) The vibration waveform detection model based on the bidirectional gated cyclic neural network has extremely strong generalization capability, is based on a large amount of elevator waveform data, has extremely high accuracy and real-time performance on diagnosis and prediction of elevator conditions once training is successful, and can be applied to elevators on various occasions.
3) Based on huge data information of a cloud, the bidirectional gating cyclic neural network training model can be updated and perfected in real time. Compared with the mode of rescue after the fact, the rescue time is shortened, and the rescue success rate is improved, the invention provides a new early warning mode, which can greatly reduce the accident rate and is very suitable for the elevator industry which is rapidly developed at present.
Detailed Description
The bidirectional gated neural network model can be used for sequence group memory identification processing, and the elevator waveform is used as data formed by a time sequence, and can be digitally processed to label the elevator waveform data, generate a dot matrix and other methods, so that the elevator waveform sequence can be recorded into the bidirectional gated neural network model in a digital form, and the functions of prediction and diagnosis are completed.
The method comprises an elevator fault prediction method and a diagnosis method based on a bidirectional gated cyclic neural network, wherein the core frameworks of a prediction model and a diagnosis model adopted by the two methods are similar, but the input, the output and the application of the two methods are completely different. Corresponding to the method process, the prediction model is based on the waveform after the waveform is predicted from the last time point, and a time sequence is output; and the diagnosis model is used for outputting whether the waveform has faults or not and the fault type based on the existing waveform. In application, the acquired waveform is always an already-generated waveform, only the acquired waveform is diagnosed, only the fault type of the generated fault can be judged, and the early warning effect cannot be achieved, and the prediction model can input the waveform to be generated into the diagnosis method as input quantity, so that the early warning effect is really achieved. Therefore, the method of the invention firstly outputs the prediction waveform through the prediction model and then outputs the prediction result through the diagnosis model, thereby realizing the early warning function.
The invention relates to an elevator fault prediction method based on a bidirectional gated cyclic neural network, which is implemented according to the following steps:
step 1: sampling the elevator vibration waveform data, and converting the elevator vibration waveform data into a sample sequence which is ordered according to acquisition time and takes the time sequence as a reference.
Step 2: converting the elevator vibration waveform sample into a sequence form, wherein the sequence comprises two parts, the first part is an input signal sequence, and the second part is an output prediction sequence;
the signal sequence and the prediction sequence of the training sample are both taken from a sample sequence, wherein the difference between the signal sequence and the prediction sequence is p time units, namely p is the time length required to be predicted, namely t-time data in the signal sequence, and is t-p time data of the prediction sequence, the shapes of the signal sequence and the prediction sequence are the same, and the matrix shapes are [ the number of samples, the step length (the same as the number of sampling points), and the output dimension ].
In an embodiment, p is 1, i.e. one time unit length of data is predicted at a time. Each elevator waveform contains 2000 sampling points, the shape of the signal sequence is the same as that of the prediction sequence, and the signal sequence and the prediction sequence are [ the number of samples, the step length (the number of the sampling points is the same), and 1], namely the dimension of data output each time is 1.
And step 3: and dividing the serialized samples into a training set and a testing set, wherein the training set data accounts for 70% of the total samples, and the testing set data accounts for 30% of the total samples.
And 4, step 4: constructing a bidirectional gate-controlled cyclic neural network framework,
the method for constructing the bidirectional gated cyclic neural network architecture comprises three parts, wherein the first part is an input part, the second part is an implicit layer part, and the third part is an output layer part. The input layer portion comprises only one input layer. The hidden layer part comprises a plurality of hidden layers, wherein the hidden layers comprise a bidirectional gated recurrent neural network layer, a full connection layer and a drop layer (dropout layer). The output layer comprises only one fully connected layer. And outputting the prediction of the sequence through the last full-connection layer by the trained bidirectional gated recurrent neural network. And the rest of the neural network layers except the first input layer are linked with the previous neural network layer through an activation function. For example: firstly, using an input layer, and then using eight hidden layers, wherein the hidden layers are a first bidirectional gated cyclic neural network layer, a first discarding layer, a second bidirectional gated cyclic neural network layer, a second discarding layer, a third bidirectional gated cyclic neural network layer, a third discarding layer, a fourth bidirectional gated cyclic neural network layer and a fourth discarding layer in sequence; and finally, an output layer part, namely a Soft-Max layer. The activation function uses the Relu activation function. And carrying out normalization processing on the output data of each layer by using batch normalization to obtain a bidirectional gated cyclic neural network framework.
And 5: and training the obtained bidirectional gated cyclic neural network framework, traversing each training data in the training set every time, wherein each traversal is called a generation, so that the bidirectional gated cyclic neural network framework is subjected to multiple generation training, and a loss function is used for outputting a loss rate in the training process. And finally obtaining the optimal neural network framework parameters after training for a plurality of generations.
In this embodiment, a global random initialization mode is used to initialize parameters, then an adam optimizer is used to train for 3000 generations, the learning rate is 0.0015, and the loss function uses a root mean square as the loss function.
Step 6: using the test set to carry out prediction test, if the waveform difference between the test set prediction waveform and the test set waveform is too large, indicating that the parameters are incorrect,
if the phenomenon occurs, the hyper-parameters need to be adjusted, such as changing the learning rate, changing the training generation, and adjusting the number of hidden layers, such as increasing or decreasing the bidirectional gating recurrent neural network layer;
and then training by using the mode of the step 5, thereby obtaining the bidirectional gating cyclic neural network prediction model with good generalization.
And 7: elevator fault prediction is carried out by using a bidirectional gated cyclic neural network prediction model,
predicting a time unit each time, namely cycling the prediction, taking the predicted value and the previous 255 values each time to form an input, namely predicting a new value through 256 values each time, and cycling 256 times to predict a complete elevator waveform sequence;
and classifying the well-trained bidirectional gated recurrent neural network prediction model before calling the elevator waveform sequence so as to obtain a predicted elevator waveform result.
The number of hidden layers of the neural network can be modified, two-way gated recurrent neural network layers can be increased or decreased, discarding layers can be fully connected, or the number of fully connected layer units can be modified.
The optimizer may also select a Momentum optimizer or an SGD optimizer.
The activation function selects a ReLU, Leaky Relu, Sigmoid, or tanh activation function.
Different loss functions may be used for training, and the cross-entropy loss function may use the mean square error or the mean difference as the loss function.
The neural network hyper-parameters can be adjusted, such as the number of training generations, the size of sequence n-steps, the learning rate, the length of input sequence, the dimension of input sequence, and the size of sequence window in the bidirectional gated cyclic neural network layer.
The invention relates to an elevator fault diagnosis method based on a bidirectional gating cyclic neural network, which is implemented according to the following steps:
step 1: and classifying and defining the samples in the elevator vibration database by experts according to fault types contained in the signals to finish sample labeling (for short, expert labeling).
Step 2: converting the samples after the classification definition into a sample sequence,
the sample sequence comprises two parts, wherein the first part is a signal sequence, and the sampling interval is correspondingly determined according to the setting of sampling time; the second part is a label sequence, and the label sequence marks the classification of the corresponding signal sequence;
converting the signal sequence data into a matrix form, wherein the matrix shape of the signal sequence is [ the number of sequence samples, the step length (namely the number of sampling points) and the dimension of input data ]; and meanwhile, converting the label sequence data corresponding to the signal sequence into a matrix form, wherein the matrix shape of the label sequence is [ the number of sequence samples, and the dimension of the output label ].
And step 3: dividing the serialized samples into a training set and a testing set,
wherein the training set data accounts for 60% of the total sample, the test set data accounts for 30% of the total sample, and the cross validation set accounts for 10%.
And 4, step 4: constructing a bidirectional gate-controlled cyclic neural network framework,
the bidirectional gating cyclic neural network framework comprises three parts, wherein the first part is an input layer part, the second part is an implicit layer part, and the third part is an output layer part; the input layer part only comprises one input layer; the hidden layer part comprises a plurality of layers including a bidirectional gated recurrent neural network layer, a full connection layer and a discarding layer (namely a dropout layer); the output layer part only comprises a Soft-Max layer.
And outputting the judgment result of the sequence by the trained bidirectional gated cyclic neural network through the last Soft-Max layer. Wherein the other two-way gated recurrent neural network layers except the (first layer) input layer are all linked with the previous two-way gated recurrent neural network layer through activation functions, such as: the method comprises the following steps that an input layer is used, eight hidden layers are used, and the hidden layers are a first bidirectional gated recurrent neural network layer, a first discarding layer, a second bidirectional gated recurrent neural network layer, a second discarding layer, a third bidirectional gated recurrent neural network layer, a third discarding layer, a fourth bidirectional gated recurrent neural network layer and a fourth discarding layer in sequence; and finally, an output layer part only comprises a Soft-Max layer (one layer). The activation function uses the Relu activation function.
In the embodiment, the input data format of the input layer is [ sample number, sequence sample number, step length (i.e. sampling point number), input data dimension ]; in the hidden layers, the bidirectional long-gated recurrent neural network layer is set to be 512 hidden units, four layers of bidirectional gated neural network layers are used, a discarding layer (namely a dropout layer) is connected behind each bidirectional gated neural network layer, the step length of each gated recurrent neural network layer is related to a sampling point and is set to be 256, and the last dropout layer is connected with a full-connection layer; the output layer is a Soft-max layer, and the Soft-max layer is connected with the full connection layer.
The output data of each layer is normalized using batch normalization. In the embodiment, the sequence comprises 256 elements, that is, 256 sampling points, which are adjusted by parameters in training, and when the hidden layer is four layers and the number of neurons in each layer of the bidirectional gated neural network is 512, a better result is obtained.
And 5: training the constructed bidirectional gated cyclic neural network framework, traversing each training data in the training set each time, wherein each traversal is called a generation, and enabling the bidirectional gated cyclic neural network framework to perform a plurality of generations of training;
testing a plurality of generations by using the data in the test set to obtain the data accuracy rate, and outputting the loss rate by using a loss function; after training for a plurality of generations, an optimal bidirectional gating recurrent neural network diagnosis model is obtained.
In this embodiment, a global random initialization mode is used to initialize parameters; training by using an adam optimizer, training 2000 generations, and the learning rate is 0.00015; the loss function uses cross entropy as the loss function.
Step 6: and (4) judging the result of over-fitting,
performing an overfitting test by using a cross validation set, wherein if the accuracy is greatly reduced, an overfitting phenomenon occurs;
if the overfitting phenomenon occurs, adjusting the hyperparameters, such as modifying the discarding rate of discarded layers, modifying the number of fully connected layers, modifying the learning rate, modifying the training generation or adjusting the number of hidden layers;
if the number of layers of the bidirectional gated recurrent neural network is reduced, the training needs to be carried out again through the step 5, so that the bidirectional gated recurrent neural network diagnostic model with good generalization is obtained.
And 7: and (3) carrying out elevator fault judgment on the elevator vibration waveform needing to be diagnosed by using the trained bidirectional gated cyclic neural network diagnosis model, wherein the judgment standard is related to the expert marking in the step 1 and conforms to the national or industry related standard, and extracting a judgment result from the last Soft-Max layer.
And 8: and outputting a judgment result, wherein the judgment result indicates that the elevator is in a fault state and then sends an alarm.
Examples
1. Parameter setting and application of the bidirectional gating recurrent neural network prediction model.
The elevator waveform sequence is used as a data set, the waveform sequence of each period of which contains 256 sampling points, and here, one hundred thousand sequences are used as an overall sample, 70% of which are training samples and 30% of which are test samples. 7 neural network layers are used, a Relu function is used between each layer of neurons as an activation function, the input layer is a full-connection layer, the hidden layer is a four-layer bidirectional gated recurrent neural network layer and a full-connection layer, the output layer is a softmax layer, and the softmax layer extracts the neural network layer and classifies the neural network layer. The shape of the input sequence is [70000,256,1], namely 70000 samples, 256 steps correspond to 256 sampling points, and the dimension of the input data is 1. The output sequence samples are [70000,256,1], i.e. the output data is identical to the input data. Training with adam optimizer, 5000 generations of training, learning rate of 0.01. The root mean square error function is used as the loss function. And (3) predicting the running waveform of the elevator system by using the trained bidirectional gated cyclic neural network model, wherein the prediction error reaches 5.61%.
2. Parameter setting and application of a bidirectional gating recurrent neural network diagnosis model.
The elevator waveform sequence is used as a data set, the waveform sequence of each period of which contains 256 sampling points, and twelve thousand sequences are used as an overall sample, wherein 60% is used as a training sample, 30% is used as a test sample, and 10% is used as a cross-validation sample. The model uses 7 layers of neural network layers, a Relu function is used between each layer of neurons as an activation function, the input layer of the model is a full-connection layer, the hidden layer is a four-layer bidirectional gated cyclic neural network layer and a full-connection layer, the output layer is a softmax layer, and the softmax layer extracts the final classification. The input sequence shape is [72000,256,1], i.e. 72000 samples, 256 steps correspond to 256 sample points, and the input data dimension is 1. The output sequence sample is [72000,256,1], i.e. the output data is identical to the input data. The training is carried out by using an adam optimizer, the learning rate is 0.01, the learning rate of each 10 generations is reduced by 10 percent on the current basis, and the generation training is carried out by 1000 generations. The overall recognition accuracy of the neural network diagnosis model obtained after complete training on various elevator faults is 96.43%.
Common elevator fault positions are mainly concentrated on a car and a traction machine. The relevant vibration waveform and motion waveform can be measured by a vibration sensor and a motion sensor aiming at the car; the temperature waveform, the voltage current waveform and the vibration waveform can be measured by an infrared array sensor, a voltage current sensor and a vibration sensor for the traction machine. Based on the elevator fault detection and early warning technology of the bidirectional gated cyclic neural network, the accuracy of identifying the voltage and current waveform sequence fault reaches 98.86 percent, the accuracy of identifying the elevator car motion waveform fault reaches 96.77 percent, the accuracy of identifying the tractor temperature fault reaches 94.21 percent, and the accuracy of identifying the tractor vibration fault reaches 96.97 percent; the accuracy rate of the method for identifying various faults can reach 96.43 percent by integrating the method, and the method can embody the advantages of the correctness of the model and the extremely high accuracy rate.
The prior art can not predict various waveform sequences generated by an elevator system, and the elevator fault detection and early warning technology based on the bidirectional gated cyclic neural network can predict waveforms in a short time in the future, and accurately predict the change condition of the waveforms in the short time of the elevator with high precision by using a root mean square error function as a loss function. The elevator state real-time analysis and monitoring can be predicted and diagnosed in real time, and early warning can be performed in time once the elevator has abnormal symptoms, so that the hysteresis of the conventional elevator monitoring system is improved; as an intelligent detection technology, the invention can quickly judge the possible faults of the elevator in a short period just by the elevator oscillogram before the safety device, thereby greatly reducing the fault accident rate of the elevator; meanwhile, compared with the feedback speed of the existing monitoring system, the system has more reaction time, greatly reduces casualties and economic loss, and fundamentally reduces the accident rate of the elevator.
Claims (1)
1. An elevator fault diagnosis method based on a bidirectional gating cyclic neural network is characterized by comprising the following steps:
step 1: classifying and defining samples in an elevator vibration database by experts according to fault types contained in the signals to finish sample labeling;
step 2: converting the samples after the classification definition into a sample sequence,
the sample sequence comprises two parts, wherein the first part is a signal sequence, and the sampling interval is correspondingly determined according to the setting of sampling time; the second part is a label sequence, and the label sequence marks the classification of the corresponding signal sequence;
converting the signal sequence data into a matrix form, wherein the matrix shape of the signal sequence is [ the number of sequence samples, the step length and the input data dimension ]; meanwhile, converting the label sequence data corresponding to the signal sequence into a matrix form, wherein the matrix shape of the label sequence is [ the number of sequence samples, the dimension of the output label ];
and step 3: dividing the serialized samples into a training set and a testing set,
wherein the training set data accounts for 60% of the total sample, the test set data accounts for 30% of the total sample, and the cross validation set accounts for 10%;
and 4, step 4: constructing a bidirectional gate-controlled cyclic neural network framework,
the bidirectional gating cyclic neural network framework comprises three parts, wherein the first part is an input layer part, the second part is an implicit layer part, and the third part is an output layer part; the input layer part only comprises one input layer; the hidden layer part comprises a plurality of layers including a bidirectional gating cyclic neural network layer, a full connection layer and a discarding layer; the output layer part only comprises a Soft-Max layer,
the trained bidirectional gated cyclic neural network outputs a judgment result of a sequence through a last Soft-Max layer, wherein the rest bidirectional gated cyclic neural network layers except the input layer are all linked with the previous bidirectional gated cyclic neural network layer through an activation function;
the output data of each layer is subjected to normalization processing by using batch normalization;
and 5: training the constructed bidirectional gated cyclic neural network framework, traversing each training data in the training set each time, wherein each traversal is called a generation, and enabling the bidirectional gated cyclic neural network framework to perform a plurality of generations of training;
testing a plurality of generations by using the data in the test set to obtain the data accuracy rate, and outputting the loss rate by using a loss function; after training for a plurality of generations, obtaining an optimal bidirectional gating recurrent neural network diagnosis model;
step 6: and (4) judging the result of over-fitting,
performing overfitting test by using a cross validation set, if the accuracy is greatly reduced, generating overfitting phenomenon, and adjusting hyper-parameters if the overfitting phenomenon occurs, wherein the discarding rate of a discarded layer is modified, the number of full-connection layers is modified, the learning rate is modified, the training generation is modified, or the number of hidden layers is adjusted;
if the number of layers of the bidirectional gated cyclic neural network is reduced, training needs to be carried out again through the step 5, and a bidirectional gated cyclic neural network diagnosis model is obtained;
and 7: performing elevator fault judgment on the elevator vibration waveform by using a bidirectional gated cyclic neural network diagnosis model, wherein the judgment standard is related to the expert marking in the step 1 and conforms to the national or industry related standard, and extracting a judgment result from the last Soft-Max layer;
and 8: and outputting a judgment result, wherein the judgment result indicates that the elevator is in a fault state and then sends an alarm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810647751.0A CN108569607B (en) | 2018-06-22 | 2018-06-22 | Elevator fault early warning method based on bidirectional gating cyclic neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810647751.0A CN108569607B (en) | 2018-06-22 | 2018-06-22 | Elevator fault early warning method based on bidirectional gating cyclic neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108569607A CN108569607A (en) | 2018-09-25 |
| CN108569607B true CN108569607B (en) | 2020-10-27 |
Family
ID=63573492
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810647751.0A Active CN108569607B (en) | 2018-06-22 | 2018-06-22 | Elevator fault early warning method based on bidirectional gating cyclic neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108569607B (en) |
Families Citing this family (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109607344B (en) * | 2018-12-20 | 2020-04-14 | 华北水利水电大学 | Neural network based vertical elevator fault prediction system and method |
| CN109893732B (en) * | 2019-02-28 | 2021-04-06 | 杭州智瑞思科技有限公司 | Mechanical ventilation man-machine asynchrony detection method based on recurrent neural network |
| CN110135241B (en) * | 2019-03-27 | 2021-08-31 | 浙江新再灵科技股份有限公司 | Statistical analysis system of elevator stroke |
| CN110009427B (en) * | 2019-04-10 | 2021-07-16 | 国网浙江省电力有限公司 | Intelligent electric power sale amount prediction method based on deep circulation neural network |
| CN110287583B (en) * | 2019-06-21 | 2021-01-29 | 上海交通大学 | Industrial equipment residual life prediction method based on cyclic neural network |
| CN110287924A (en) * | 2019-06-28 | 2019-09-27 | 电子科技大学 | A kind of soil parameters classification method based on GRU-RNN model |
| CN110386520A (en) * | 2019-08-05 | 2019-10-29 | 浙江一网通信息科技有限公司 | A kind of Elevator operation control system and method |
| CN110766070A (en) * | 2019-10-22 | 2020-02-07 | 北京威信通信息技术股份有限公司 | Sparse signal identification method and device based on cyclic self-encoder |
| CN111060815B (en) * | 2019-12-17 | 2021-09-14 | 西安工程大学 | GA-Bi-RNN-based high-voltage circuit breaker fault diagnosis method |
| CN111263326B (en) * | 2020-01-09 | 2022-11-04 | 中国人民解放军国防科技大学 | Vehicle position prediction method based on multiple fusion convolution GRU |
| CN112052894B (en) * | 2020-09-02 | 2023-02-07 | 北京理工大学 | Real-time monitoring and evaluating method for power cabin state |
| CN112648923A (en) * | 2020-12-12 | 2021-04-13 | 山东省农业机械科学研究院 | Virtual calculation method for detecting size of product part |
| CN112631434B (en) * | 2021-01-11 | 2022-04-12 | 福州大学 | Deep learning-based vibrotactile coding and decoding method |
| CN113033174B (en) * | 2021-03-23 | 2022-06-10 | 哈尔滨工业大学 | Case classification method and device based on output type similar door and storage medium |
| CN113086798B (en) * | 2021-04-06 | 2022-02-08 | 中南大学 | Elevator fault detection method based on gated cyclic network and typical correlation analysis |
| CN114955771B (en) * | 2022-05-13 | 2023-06-09 | 江苏省特种设备安全监督检验研究院 | Elevator control system fault monitoring method based on finite state machine |
| CN115086006B (en) * | 2022-06-13 | 2024-02-02 | 安徽工业大学 | Distributed application program encryption traffic classification method based on bidirectional gating logic unit |
| CN116304781B (en) * | 2022-09-09 | 2024-02-02 | 武汉中数医疗科技有限公司 | Thyroid sampling data identification method based on cyclic neural network |
| CN116383697B (en) * | 2023-03-29 | 2024-02-06 | 四川省交通建设集团有限责任公司 | Bituminous pavement compaction degree prediction method based on double-gate circulation network |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN202414914U (en) * | 2011-12-27 | 2012-09-05 | 阳西县电梯配件有限公司 | Elevator safety detection device based on characteristic signals |
| CN102491140A (en) * | 2011-12-27 | 2012-06-13 | 阳西县电梯配件有限公司 | Characteristic-signal-based elevator safety checking device and elevator safety checking method |
| CN104239905A (en) * | 2013-06-17 | 2014-12-24 | 上海盖普电梯有限公司 | Moving target recognition method and intelligent elevator billing system having moving target recognition function |
| CN103678952A (en) * | 2013-11-14 | 2014-03-26 | 昆明理工大学 | Elevator risk evaluation method |
| CN107194053B (en) * | 2017-05-16 | 2020-10-20 | 歌拉瑞电梯股份有限公司 | Intelligent elevator control system operation fault prediction method |
| CN108516432A (en) * | 2018-04-10 | 2018-09-11 | 百世杉工业技术有限公司 | A kind of elevator reparing maintenance system based on Internet of Things |
-
2018
- 2018-06-22 CN CN201810647751.0A patent/CN108569607B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN108569607A (en) | 2018-09-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108569607B (en) | Elevator fault early warning method based on bidirectional gating cyclic neural network | |
| CN110647133B (en) | Rail transit equipment state detection maintenance method and system | |
| CN108764601B (en) | Structural health monitoring abnormal data diagnosis method based on computer vision and deep learning technology | |
| CN111351664B (en) | Bearing temperature prediction and alarm diagnosis method based on LSTM model | |
| CN113469060A (en) | Multi-sensor fusion convolution neural network aeroengine bearing fault diagnosis method | |
| CN111274737A (en) | Method and system for predicting remaining service life of mechanical equipment | |
| CN105095963A (en) | Method for accurately diagnosing and predicting fault of wind tunnel equipment | |
| CN111639467A (en) | Aero-engine service life prediction method based on long-term and short-term memory network | |
| CN108709744B (en) | Motor bearings method for diagnosing faults under a kind of varying load operating condition | |
| CN111523081A (en) | Aircraft engine fault diagnosis method based on enhanced gated cyclic neural network | |
| CN107844067B (en) | A kind of gate of hydropower station on-line condition monitoring control method and monitoring system | |
| CN110375983A (en) | Failsafe valve real-time diagnosis system and diagnostic method based on time series analysis | |
| Son et al. | Deep learning-based anomaly detection to classify inaccurate data and damaged condition of a cable-stayed bridge | |
| CN103674538A (en) | Multi-fault mode identification method and device of swashplate of helicopter | |
| CN111678699B (en) | Early fault monitoring and diagnosing method and system for rolling bearing | |
| CN116610998A (en) | Switch cabinet fault diagnosis method and system based on multi-mode data fusion | |
| CN112036087A (en) | Multi-strategy fused nuclear power key equipment fault diagnosis method and system | |
| CN113962308A (en) | Aviation equipment fault prediction method | |
| CN111260823B (en) | Fault diagnosis method based on I/O (input/output) measuring point fault dependency matrix | |
| CN113987905A (en) | Escalator braking force intelligent diagnosis system based on deep belief network | |
| CN110320802B (en) | Complex system signal time sequence identification method based on data visualization | |
| CN115600695A (en) | Fault diagnosis method of metering equipment | |
| CN115062867A (en) | Bearing fault prediction method based on Teager energy spectrum and improved CNN model | |
| Lv et al. | A weight recognition method for movable objects in sealed cavity based on supervised learning | |
| Rombach et al. | Contrastive feature learning for fault detection and diagnostics in railway applications |
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 |