Crane operation state monitoring method and monitoring system

Abstract

According to the crane running state monitoring method and system, state parameter data during running of a crane are collected through a sensor and transmitted into a monitor, the running state of the crane is judged through a monitoring model, the input part of the monitoring model comprises an encoder and a decoder, the encoder adds Gaussian noise to the input data, and the decoder outputs the Gaussian noise to the input data; and the decoder reconstructs the noise-containing data added with the noise, so that the influence of the data noise on the detection result is effectively eliminated, and the monitoring precision is improved. A main body part of the monitoring model comprises a multi-layer convolution layer, a pooling layer, a full connection layer and a softmax classifier, the weight of each operation state category is calculated according to the number of samples in the training process, all the samples are subjected to weighted average to calculate a loss function so as to update model parameters, monitoring result deviation caused by sample imbalance can be solved, and the monitoring accuracy is improved. The intelligent monitoring level of the operation state of the crane is effectively improved, and the operation safety and reliability of the crane are guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to crane intelligent detection technology, in particular to a crane operation state monitoring method and monitoring system. BACKGROUND

[0002] Large crane equipment, due to its own structure and working characteristics, determines that it is a construction machinery with a large hidden danger. Therefore, in order to ensure the stability and safety of the crane construction operation, the running state of the crane needs to be monitored and fault warning. The traditional crane monitoring system mainly relies on threshold alarm and manual inspection, which is difficult to identify the running state of the crane in real time and accurately predict potential failures. At present, some researches use intelligent algorithms to improve the monitoring ability of the crane, but there are problems such as data noise and the influence of unbalanced classification data collected on site on the monitoring ability. The traditional monitoring method relies on fixed rules or simple models, which is difficult to adapt to the state monitoring and identification requirements under complex working conditions; some algorithms are complicated and difficult to meet the real-time requirements of online monitoring. SUMMARY

[0003] The technical problem to be solved by the present application is to overcome the defects of data noise and unbalanced classification data affecting the monitoring accuracy, and to provide a crane operation state monitoring method and monitoring system.

[0004] The technical scheme adopted by the present application to solve the above technical problems is: a crane operation state monitoring method, state parameter data of the crane during operation is collected through a sensor, the collected data is transmitted into a monitor, and a monitoring model deployed in the monitor is used to judge the running state of the crane according to the collected data, the monitoring model is a convolutional neural network model, including an input part, a main part and an output part, the input part is used to receive state parameter data and transmit the processed data to the main part; the main part extracts features in the data and outputs the running state type; the input part includes an encoder and a decoder, the encoder adds Gaussian noise to the input data, and the decoder reconstructs the noisy data with noise; the parameters of the encoder and the decoder are trained by minimizing the loss function to make the reconstructed data close to the data without adding noise; the main part includes multiple convolutional layers, pooling layers, fully connected layers and softmax classifiers, the output of the decoder is subjected to multiple convolution and pooling operations, and then transmitted to the fully connected layer and the softmax classifier for running state classification; the main part calculates the weight of each running state category according to the number of samples during the training process, and all samples are weighted and averaged to calculate the loss function to update the model parameters; the output part outputs the diagnosis result according to the running state classification of the main part.

[0005] The state parameter data includes the lifting weight, lifting torque, lifting height, lifting amplitude, wind speed and motor winding temperature of the crane.

[0006] The collected state parameter data is spliced into input data of the monitoring model after standardization processing.

[0007] The encoder of the input part adds Gaussian noise to the state parameter data and performs normalization processing, and the formula is: wherein, is the input data, is the noisy data, obeys the standard normal distribution, is a noise factor for adjusting the noise intensity, is used to limit the result in the interval [0, 1].

[0008] The encoder is provided with two hidden layers, and a batch normalization layer is added after the first hidden layer.

[0009] The decoder is provided with one hidden layer, the number of nodes of the decoder hidden layer is 32, and the number of nodes of the two hidden layers of the encoder is 64 and 32 respectively.

[0010] In the main part, the convolution layer performs convolution operation on the input data and the convolution kernel to generate a feature map, after the convolution operation, the ReLU activation function is used to perform nonlinear transformation on the output of the convolution layer, and then the maximum pooling is used to capture the signal characteristics; after the multiple convolution layers and the pooling layers, the generated feature map is first flattened, then the full connection layer is connected after the flattening, and finally the softmax classifier is used to classify the running state.

[0011] The main part includes a plurality of stacked convolution modules, each convolution module is provided with a convolution layer, a BN layer and a pooling layer, and a Dropout layer is arranged after the convolution module to prevent overfitting.

[0012] The main part calculates the weight of each class according to the total number of training samples, the total number of classes and the number of samples in each class, weights the cross-entropy loss function of the samples in the class by using the calculated weight, averages the loss functions of all weighted samples, and uses the average loss function to optimize the model parameters.

[0013] The application discloses a monitoring system for a crane operation state monitoring method, which comprises sensors arranged on the crane and used for monitoring state parameters in real time when the crane is running, and a monitor used for judging whether the crane operation state is abnormal according to the state parameters, wherein the monitor comprises a data processing module arranged on the crane and a master control module arranged on a cloud, and the data processing module is in wireless communication connection with the master control module; the data processing module is deployed with a monitoring model input part comprising an encoder and a decoder, and the master control module is deployed with a monitoring model main part and an output part used for extracting data features and giving a crane operation state classification result.

[0014] The application has the beneficial effects that the encoder and the decoder trained by the input part can effectively eliminate the influence of data noise on the detection result by adding Gaussian noise for data reconstruction, thereby improving the monitoring precision. The main part of the model adopts the trained convolutional neural network, combines the data processing of the input part, reduces the model complexity while ensuring the monitoring precision, and is more conducive to realizing the real-time state monitoring. The sample loss function is weighted by calculating the weight of the operation state category, the monitoring result deviation caused by the sample imbalance can be solved, the intelligent level of the crane operation state monitoring is effectively improved, the operation safety and reliability of the crane are ensured, and the application is suitable for intelligent safety monitoring and fault early warning of large hoisting equipment such as tower cranes and bridge cranes. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 Fig. 1 is a network structure schematic diagram of the monitoring model of the application.

[0016] Figure 2 Fig. 2 is a network structure schematic diagram of the input part of the application. DETAILED DESCRIPTION

[0017] The technical solutions of the application are described in detail below in combination with the drawings and specific embodiments. The specific contents listed in the following embodiments are not limited to the technical features necessary for the technical problems to be solved by the technical solutions recited in the claims. At the same time, the enumeration is only a part of the application, not all embodiments.

[0018] The application adopts an intelligent monitoring system to monitor the operating state of the crane, the monitoring system comprises sensors and a monitor, the sensors are arranged on the crane and are used to monitor various state parameters in real time when the crane is operating, such as the lifting weight, lifting moment, lifting height, lifting amplitude, wind speed and motor winding temperature of the crane. According to the type of the monitored state parameters, multiple corresponding sensors are arranged. The monitor is arranged with a monitoring model, which is used to process the state parameter data monitored by the sensors, and judge whether the operating state of the crane is abnormal according to the extracted data features. The monitoring system collects the state parameter data of the crane when operating through the sensors, transmits the collected data into the monitor, and judges the operating state of the crane according to the collected data by using the monitoring model arranged in the monitor.

[0019] The monitoring model is a convolutional neural network model, comprising an input part, a main part and an output part. The input part is used to receive the state parameter data and transmit the processed data to the main part. The main part is used to extract the features in the data and output the operating state type. The output part outputs the diagnosis result according to the operating state type output by the main part.

[0020] Figure 1 The embodiment shown is an embodiment schematic diagram of the monitoring model. The input layer in the figure is the input part, the main part comprises a convolutional layer, a pooling layer, a fully connected layer and other neural network structures, and the output layer is the output part.

[0021] In an embodiment of the application, the input part comprises an encoder and a decoder, the encoder is used to encode the input data, and the decoder reconstructs the encoded data into the original data. In order to reduce the influence of data noise on the monitoring accuracy, the input data is added with Gaussian noise by the encoder, and then the original signal features are reconstructed from the noisy data by the decoder, so as to improve the robustness of the crane operating state monitoring model. The parameters of the encoder and the decoder are trained by minimizing the loss function, so that the reconstructed data is close to the data without adding noise. The data output by the decoder is used as the input of the main part.

[0022] Figure 2 The embodiment shown is an embodiment schematic diagram of the input part. In the figure - The data without adding noise is obtained by adding noise - , and the hidden layer data is obtained by processing the data by the encoder - , and then the data is reconstructed by the decoder to obtain the reconstructed data - In the present application, Gaussian noise is added to the data, and the data is limited to remain between 0 and 1 after the Gaussian noise is added. The encoder of the input part adds Gaussian noise to the state parameter data and performs normalization processing, and the formula is: wherein, is the input data, is the noisy data, is the Gaussian noise, which is subject to standard normal distribution, is a noise factor for adjusting the noise intensity, is used to limit the result to the interval [0, 1].

[0023] Forward propagation of the encoder-decoder model: the encoder can be set to a multi-layer structure, for example, two hidden layers are set, and through the multi-layer denoising autoencoder structure, the deep feature representation in the crane detection signal is extracted layer by layer. A batch normalization (BN) layer can be introduced in the encoder, which can effectively alleviate the problems of gradient disappearance in deep network and gradual shift of data distribution to the limit interval of nonlinear function with the increase of the number of layers by normalizing the distribution of each layer input, thereby accelerating the training convergence and improving the stability of the features. For example, when two hidden layers are set, a batch normalization layer is added after the first hidden layer. The decoder maps the noisy data to the reconstructed data through the encoder-decoder, thereby improving the stability.

[0024] wherein, is the noisy data, , are the weight and bias parameters of the encoder, respectively, is the encoder hidden layer data, , are the weight and bias parameters of the decoder, respectively, is the encoder output, is the mean of the input data, is the variance of the input data, is a very small constant value.

[0025] In the training process of the encoder-decoder, mean squared error (MSE) is used as the loss function to measure the difference between the reconstructed data and the original data.

[0026] wherein, is the batch size.

[0027] Back propagation and parameter update: constantly update parameters to minimize mean square error, constantly iterate to get optimal parameters until the training is completed after reaching the set conditions.

[0028] wherein, is the learning rate, is the partial derivative of .

[0029] The data monitored by the related sensors is processed by the encoder and decoder of the input part, and the encoded feature data obtained after processing is taken as the input of the main part, and the crane operation state type code is taken as the output of the classification result of the main part.

[0030] The main part includes multiple convolutional layers, pooling layers, fully connected layers and softmax classifiers. The output of the decoder is subjected to multiple convolution and pooling operations. The convolutional layer performs convolution operation on the input data and the convolution kernel to generate a feature map. After the convolution operation, the output of the convolutional layer is subjected to nonlinear transformation by using the ReLU activation function, and then maximum pooling is used to capture signal features. After multiple convolutional layers and pooling layers, the generated feature map is first subjected to flattening operation, and then connected to the fully connected layer, and finally the softmax classifier is used to classify the operation state.

[0031] The convolutional layer uses one-dimensional convolution operation to perform convolution operation on the input data and the convolution kernel to generate a feature map.

[0032] wherein, is the output obtained by inputting the input data through the th convolution kernel, is the edgeless padding convolution operation, is the th convolution region in the th layer, is the convolution kernel width, is the th weight in the th convolution kernel, is the bias vector of the th feature map of the convolutional layer in the th layer.

[0033] Activation operation: ReLU is used as the activation function of the convolutional neural network. After the convolution operation, the activation function performs nonlinear transformation on the output of the convolutional layer.

[0034] Pooling operation: The pooling operation follows the convolution layer, which uses maximum pooling to capture the most valuable signal features.

[0035] wherein, is the output value obtained by average pooling, is the width of the pooling operation region, represents the activation value of the layer at the th feature map at position n.

[0036] After multiple convolution layers and pooling layers, the generated feature map is first flattened, and then connected to the fully connected layer.

[0037] wherein, is the output value of the th neuron of the th layer.

[0038] Loss function calculation: In order to solve the problem of unbalanced classification samples, a weight distribution strategy based on class frequency is adopted to improve the cross-entropy loss function, so that the model treats all classes equally during training.

[0039] wherein, is the weighted loss function, m is the number of samples in the current batch; k is the sample index (k=1,2,…,m); is the weight of class ; is the predicted probability distribution of the kth sample; is the encoded true label vector of the kth sample; is the standard cross-entropy loss; represents the probability of predicting class c; is the total number of samples in the training set; is the total number of classes (i.e., C); represents the number of samples of class in the training set.

[0040] The main part calculates the weight of each classification category in advance according to the total number of training samples, the total number of categories, and the number of samples in each category, and the weight of each category = total number of training samples ÷ total number of categories × number of samples of class c. For each sample k, calculate its standard cross-entropy loss according to the real category of the sample , find the corresponding weight , the cross-entropy loss function of the samples in the category is weighted by the calculated weight, the loss function of all samples is averaged, and the total loss L of the batch is obtained to optimize the model parameters.

[0041] Back propagation and parameter update: constantly update the parameters to minimize the mean square error, and after constantly iterating to obtain the optimal weight and threshold, the training is completed.

[0042] , wherein, is the optimal parameter of the convolutional neural network.

[0043] The data processed by the input part is transmitted to the trained main part for prediction, and the diagnostic result is output by the output part.

[0044] In order to better eliminate the noise influence of sensor error and extract the correlation characteristics of multiple observation variables, the application provides an optimized embodiment. Different types of state parameter data collected by multiple sensors are respectively standardized. The lifting weight and wind speed are standardized by Z-score to retain the distribution characteristics; the lifting height and motor winding temperature are normalized by MinMax to ensure that the data is in the [0, 1] interval; the lifting moment is standardized by logarithmic transformation to compress the large value range and highlight the small value change; the lifting amplitude is standardized by interval mapping to avoid invalid values in the sensor blind area. Different normalization methods are used for different parameter physical characteristics to avoid model bias caused by dimensional differences. After standardization, each observation variable is added with noise, and the noise factor can be independently set. The processed 6 observation variables form a 6-dimensional vector, the input layer and the output layer of the input part both contain 6 neuron nodes, the encoder sets two hidden layers with node numbers of 64 and 32 respectively. The decoder sets one hidden layer with a node number of 32. The encoder processes the 6-dimensional input vector, and multiple dimensions of features share a weight matrix to enhance the interaction between observation variables and suppress the influence of single variable noise by using the correlation between variables. The encoder compresses the dimensions, and the decoder restores the 6-dimensional vector output after processing. The main part receives the output of the decoder, recombines the 6-dimensional vector into a three-dimensional tensor with a time step of 6, and each time step corresponds to the characteristics of an observation variable. The convolution layer uses one-dimensional convolution, and the convolution kernel size is set to 3 to cover 3 consecutive time steps to capture the local correlation between different variables.

[0045] A plurality of stacked convolution modules can be arranged in the main body part, and a convolution layer, a BN layer and a pooling layer are arranged in each convolution module, and the number of channels of the plurality of convolution modules arranged in sequence gradually increases, for example, the number of channels of the first convolution module is 64, the single parameter time domain change such as the fluctuation of the lifting weight is focused on, the number of channels of the second convolution module is 128, the double parameter association such as the coupling of the lifting weight and the torque is captured, and the number of channels of the third convolution module is 256, and the multi-parameter joint anomaly is identified. Residual connection is used between different convolution modules to solve the gradient vanishing problem of the deep neural network. A Dropout layer is arranged after the convolution module to prevent overfitting.

[0046] The input part and the main body part of the monitoring model of the application are trained separately and different data processing is performed. The model of the input part can be deployed on the device site to process real-time data in time. The model of the main body part is deployed in the cloud to obtain more computing resources. The monitor of the application is divided into a data processing module arranged on the crane and a main control module arranged in the cloud, and the data processing module is in wireless communication connection with the main control module. The monitoring model input part including an encoder and a decoder is deployed in the data processing module, and the monitoring model main body part and the output part for extracting data features and giving the classification result of the crane running state are deployed in the main control module.

[0047] The application selects six variables of the tower crane running state as observation variables, which are: lifting weight, lifting torque, lifting height, lifting amplitude, wind speed and motor winding temperature. The six observation variables are all variables that can be collected by sensors during the operation of the tower crane. The running state of the crane is divided into three categories, in order: relatively safe state, safe to dangerous transition state, and relatively dangerous state. The codes of the three states are 100, 010 and 001, respectively. In the historical data corresponding to the tower crane running state information observation variables and the running state classification code, 24 training samples and 9 classification verification samples are selected, and the verification samples are not included in the training samples. The 24 training samples are shown in Table 1.

[0048] A denoising autoencoder (DAE) model framework including an encoder and a decoder is constructed and network parameters are configured, and the input and output of the model contain six neuron nodes. The encoder is composed of two hidden layers (the node numbers are 64 and 32, respectively), and the decoder is provided with a single hidden layer (32 nodes). The training sample with added noise and normalized is used as the network input, and the model is trained with the un-noisy data as the target output, so as to generate a tower crane feature processing model based on DAE.

[0049] The main body part convolutional neural network prediction model is constructed, the parameters of the network are set, 128 convolution kernels with a size of 3 are used in the convolution layer, a Relu activation function is adopted, feature compression is performed through a maximum pooling layer with a step of 2, the flattened input is input into a fully connected layer containing 64 neurons, and finally the classification result is output through a softmax function. The training process adopts an Adam optimizer with a learning rate of 0.001, uses sparse classification cross-entropy as the loss function, sets an early stopping mechanism (monitors the validation loss value, terminates training and restores the optimal weight if there is no improvement for 15 rounds, and adopts a learning rate dynamic decay strategy (the learning rate is reduced to 20% of the original value after the loss stagnates for 5 rounds, and the minimum is not less than 0.0001), and performs 200 rounds of iteration optimization through batch gradient descent (batch_size=16), thereby generating a crane operating state monitoring classification model based on a convolutional neural network.

[0050] Nine groups of samples are selected from the historical data of the corresponding relationship between the tower crane operating state information observation variables and the operating state classification codes as monitoring classification verification samples. After normalizing the state variables of the monitoring verification samples, the encoder output feature vector is obtained by inputting the monitoring verification samples into the trained DAE network. The classification result obtained by inputting the output result into the trained convolutional neural network is shown in Table 2.

[0051] The above description of the specific embodiments is only used to help understand the technical concept and core idea of the present application, although specific preferred embodiments are used herein to describe and explain the technical solutions, but it should not be understood as a limitation of the present application itself. Those skilled in the art can make various changes in form and details without departing from the technical concept of the present application. These easily thought changes or replacements should be covered within the protection scope of the present application.

Claims

1. A crane operation state monitoring method, state parameter data of a crane operation is collected by a sensor, the collected data is transmitted into a monitor, and the operation state of the crane is judged according to the collected data by using a monitoring model deployed in the monitor, characterized in that: The monitoring model is a convolutional neural network model, comprising an input part, a main part and an output part, the input part is used for receiving state parameter data and transmitting to the main part after processing; the main part extracts features in the data and outputs the running state type; the input part comprises an encoder and a decoder, the encoder adds Gaussian noise to the input data, the decoder reconstructs the noisy data with noise, and the parameters of the encoder and the decoder are trained by minimizing the loss function to make the reconstructed data close to the data without adding noise; the main part comprises a plurality of convolutional layers, pooling layers, fully connected layers and a softmax classifier, the output of the decoder is subjected to a plurality of convolution and pooling operations, and then transmitted to the fully connected layer and the softmax classifier for running state classification; the main part calculates the weight of each running state category according to the number of samples during the training process, and calculates the loss function by weighted average of all samples to update the model parameters; the output part outputs the diagnosis result according to the running state classification of the main part.

2. A method of monitoring the operating state of a crane as claimed in claim 1, characterized in that The state parameter data comprises the lifting weight, lifting moment, lifting height, lifting amplitude, wind speed and motor winding temperature of the crane.

3. A method of monitoring the operating state of a crane as claimed in claim 2, characterized in that: The collected state parameter data is spliced into the input data of the monitoring model after standardization processing.

4. A method of monitoring the operating state of a crane as claimed in claim 1, characterized in that: The encoder of the input part adds Gaussian noise to the state parameter data and performs normalization processing, and the formula is: wherein, is the input data, is the noisy data, obeys a standard normal distribution, is a noise factor that adjusts the noise intensity, is used to limit the result to the interval [0, 1].

5. A method of monitoring the operating state of a crane as claimed in claim 1, characterized in that: The encoder is provided with two hidden layers, and a batch normalization layer is added after the first hidden layer.

6. A method of monitoring the operating state of a crane as claimed in claim 5, characterised in that: The decoder is provided with one hidden layer, the number of nodes of the decoder hidden layer is 32, and the number of nodes of the two hidden layers of the encoder is 64 and 32 respectively.

7. A crane operating condition monitoring method as claimed in claim 1, characterised by: In the main part, the convolution layer performs convolution operation on the input data and the convolution kernel to generate a feature map, after the convolution operation, the ReLU activation function is used for nonlinear transformation of the output of the convolution layer, and then the maximum pooling is used to capture the signal features; After the plurality of convolution layers and the pooling layers, the generated feature map is first flattened, then connected to the fully connected layer, and finally the softmax classifier is used to classify the running state.

8. A crane operating condition monitoring method as claimed in claim 7, characterised by: The main part comprises a plurality of stacked convolution modules, each convolution module is provided with a convolution layer, a BN layer and a pooling layer, and a Dropout layer is arranged after the convolution module to prevent overfitting.

9. A method of monitoring the operating state of a crane as claimed in claim 1, characterized in that: The main part calculates the weight of each category according to the total number of training samples, the total number of categories and the number of samples in each category, weights the cross-entropy loss function of the samples in the category by using the calculated weight, averages the loss functions of all weighted samples, and uses the average loss function to optimize the model parameters.

10. A monitoring system for use in a crane operating condition monitoring method as claimed in claim 1, characterized in that The crane comprises sensors arranged on the crane for monitoring state parameters of the crane in real time when the crane is running, and a monitor for judging whether the running state of the crane is abnormal according to the state parameters, the monitor comprising a data processing module arranged on the crane and a master control module arranged in the cloud, and the data processing module is in wireless communication connection with the master control module; the data processing module is provided with a monitoring model input part comprising an encoder and a decoder, and the master control module is provided with a monitoring model main part and an output part for extracting data features and giving a classification result of the running state of the crane.

Images