CNN-BLRNET thermal error prediction model and its transfer learning method and intelligent integration framework
By using the CNN-BLRNET thermal error prediction model and intelligent integrated framework in precision machine tools, the problems of thermal error prediction and control are solved, the processing accuracy and stability are improved, and adaptability to different working conditions and equipment is achieved.
Patent Information
- Application Number
- CN202211362282.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-02
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2042-11-02
AI Technical Summary
The prior art is difficult to effectively predict and control the thermal errors of precision machine tools, resulting in unstable machining accuracy, and insufficient generalization capabilities of traditional models, which are unable to adapt to changes in different working conditions and equipment.
A CNN-BLRNET thermal error prediction model is proposed. By embedding the BILSTM model into the RESNET network, combining the dropout layer and the maximum pooling layer, the spatial and temporal characteristics of the thermal error data are extracted, and the transfer learning method is adopted to improve the model's Generalization ability. At the same time, an intelligent integrated framework was designed to realize data acquisition, calculation and fault diagnosis, and provide thermal error control and status monitoring services.
It improves the accuracy and robustness of thermal error prediction, enhances the adaptability of the model, realizes adaptability to different working conditions and equipment, reduces the computing resource requirements for retraining, and improves processing accuracy and stability.
Smart Images

Figure CN115657600B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of mechanical error control, and specifically relates to a CNN-BLRNET thermal error prediction model and a transfer learning method and an intelligent integration framework thereof. Background Art
[0002] High-speed and high-precision machining has become the core of modern manufacturing, and the performance of precision machine tools is crucial. Error control and fault diagnosis are the key to ensuring machining performance. There is also a coupling between error control and fault diagnosis. Early faults are not easy to identify and may not affect machining efficiency, but when the fault deteriorates further, the accuracy stability will be affected. In previous studies, error control and fault diagnosis were studied separately, but in fact, error control and fault diagnosis are two important aspects of digital machining. Through fault diagnosis, early warnings of tool wear and damage and bearing damage can be reported, and then tools and bearings can be replaced in time. According to the requirements of digital machining, real-time thermal error control, condition monitoring and fault diagnosis during machining are essential.
[0003] Precision machine tools have thermal errors due to friction and electric heat. For precision machine tools, high-speed spindles are widely used. Built-in motors and high-speed bearings are placed in high-speed spindles, so that the motor can directly drive the spindle, reduce transmission errors and vibrations, and greatly improve transmission performance. However, the entire high-speed spindle is affected by a variety of heat sources, resulting in uneven temperature field distribution and different thermal deformation of each component, resulting in thermal errors. Research results show that 70% of machine tool errors are caused by thermal errors. Therefore, it is necessary to reduce thermal errors.
[0004] Unconventional working conditions such as high speed, high acceleration, large load, and large displacement have a significant impact on vibration, impact, and deformation, which in turn lead to various failures of mechanical components such as rolling guides, bearings, couplings, cams, gears, worm wheels, and worm shafts, and finally cause motion errors, component wear, and even unexpected downtime of machine tools. Correctly evaluating the current working status of machine tools and predicting the development trend of machine tool operating status are of great guiding significance for machine tool maintenance. Online monitoring and fault diagnosis allow for predictability of faults, and then maintenance does not need to wait until the fault actually occurs. Through online detection and collection and analysis of operating status, possible fault conditions are predicted based on normal state parameter standards and historical maintenance knowledge, providing a basis for formulating a reasonable maintenance plan. Summary of the invention
[0005] In view of this, the purpose of the present invention is to provide a CNN-BLRNET thermal error prediction model and its transfer learning method and intelligent integration framework. The CNN-BLRNET thermal error prediction model can extract the spatial and temporal characteristics of thermal error data, and can improve the prediction accuracy and robustness; the transfer learning method can improve the generalization ability of the model to adapt to different working conditions and different machine tool equipment; the intelligent integration framework realizes the internal and external thermal error data collection of the machine tool spindle, and can provide data services including thermal error control, state monitoring and fault diagnosis.
[0006] In order to achieve the above object, the present invention provides the following technical solutions:
[0007] The present invention firstly proposes a CNN-BLRNET thermal error prediction model, which includes sequentially setting a first input layer, a CNN layer, a BLRNET model, a dropout layer, a flatten layer, a dense layer and a first output layer;
[0008] The BLRNET model is formed by embedding a BILSTM model into a RESNET network, and the BLRNET model includes a second input layer, a ReLU layer and a second output layer, at least two RESNET units connected in series are arranged between the ReLU layer and the output layer, the RESNET unit includes a direct mapping part and a skip connection part, the direct mapping part includes a convolutional layer I, a maximum pooling layer I and a BILSTM layer arranged in sequence, the skip connection part includes a convolutional layer II and a maximum pooling layer II arranged in sequence, and an addition operator is arranged between the BILSTM layer and the maximum pooling layer II.
[0009] Furthermore, the BILSTM model includes an input layer, a forward LSTM layer, a backward LSTM layer and an output layer. The forward LSTM layer and the backward LSTM layer include several LSTM units. The principle of the LSTM unit is:
[0010] f t =σ(W f [h t-1 , x t ]+b f )
[0011] i t =σ(W f [h t-1 , x t ]+b i )
[0012] C t =tanh(W f [h t-1 , xt ]+b c )
[0013]
[0014] O t =σ(W o [h t-1 , x t ]+b o )
[0015] h t =O t *tanh(C t )
[0016] Where t and t-1 are the current and previous moments respectively; f t 、i t and O t They are forget gate, input gate and output gate respectively; x t , C t and h t are the input, unit state and output at the current moment; x t-1 , C t-1 and h t-1 are the input, unit state and output at the previous moment respectively; is the candidate cell state; W f , W i , W c and W o is the weight matrix; b f 、b i 、b c and b o is the bias term; * denotes element-wise multiplication; σ(·) and tanh(·) are activation functions.
[0017] Furthermore, the principle of BILSTM is:
[0018] h t =f(w1*x t +w2*h t-1 )
[0019] h′ t =f(w3*x t +w4*h′ t-1 )
[0020] o t =g(w5*h t +w6*h t )
[0021] Where f(·) represents the forward operation function; g(·) represents the backward operation function; ht represents the forward output; h′ t represents the backward output; w1, w2, w3, w4, w5 and w6 represent the weight matrices.
[0022] Furthermore, the principle of the RESNET unit is that the data is converted into nonlinearity through the ReLU layer:
[0023]
[0024] Where x represents the input data; relu(·) represents the ReLU activation function, x i Represents the i-th input data;
[0025] The data can be input as X1 to directly map the part, and the data can also be input as X2 to the remaining part. For convolution layer I:
[0026]
[0027] Among them, X1 represents the input of the direct mapping part; W1 represents the convolution weight matrix; Represents the input value; Represents the convolution kernel value; b1 represents the bias parameter in the convolution operation; represents the actual convolution operation bias; f1(·) represents the actual convolution operation; conv1d(·) represents the convolution operation;
[0028] For max pooling layer I:
[0029]
[0030] Where X′1 represents the output of the maximum pooling layer I; maxpooling(·) represents the pooling operation; Represents the pooled input value;
[0031] The BILSTM layer extracts temporal features:
[0032] h t =f(w1*X′1+w2*h t-1 )
[0033] h′ t =f(w3*X′1+w4*h′ t-1 )
[0034] The output of the direct mapping section is:
[0035] F(X,W)=h t +h′ t
[0036] Wherein, F(X, W) represents the output of the direct mapping part;
[0037] For convolutional layer II:
[0038]
[0039] Among them, X2 represents the input of the direct mapping part; W2 represents the convolution weight matrix; Represents the input value; represents the convolution kernel value; b2 represents the bias parameter in the convolution operation; represents the actual convolution operation bias; f1(·) represents the actual convolution operation; conv1d(·) represents the convolution operation;
[0040] The output of the jump connection part is:
[0041]
[0042] Among them, G(X) represents the output of the jump connection part; Represents the pooled input value;
[0043] The output of the BLRNET unit is:
[0044] H(X)=F(X,W)+G(X)
[0045] Where H(X) represents the output of the BLRNET unit.
[0046] The present invention also proposes a transfer learning method for the CNN-BLRNET thermal error prediction model as described above, retaining the parameters of the first input layer, CNN layer, BLRNET model dropout layer, flatten layer and first output layer in the original CNN-BLRNET thermal error prediction model unchanged, retraining and updating the dense layer, and obtaining a transfer learning model of the CNN-BLRNET thermal error prediction model.
[0047] The present invention also proposes an intelligent integrated framework, including a data acquisition system and a data calculation system;
[0048] The data acquisition system collects machine tool working status data and workpiece quality data; all collected data are transmitted to the edge node and the cloud center through the local area network and the industrial Internet respectively;
[0049] The computing system performs data computing on the collected data, performs self-learning with the support of machine learning, reinforcement learning, deep learning and transfer learning, establishes the fault diagnosis model, thermal error model, three-dimensional model, control strategy and knowledge of the machine tool as a digital twin model, and provides data services including condition monitoring, fault diagnosis and thermal error control;
[0050] The cloud center is embedded with the CNN-BLRNET thermal error prediction model as described in any one of claims 1 to 4, and the cloud center performs preliminary training on the CNN-BLRNET thermal error prediction model based on historical thermal error data; the CNN-BLRNET thermal error prediction model that has completed preliminary training is transmitted to the edge node, and the edge node performs local training on the CNN-BLRNET thermal error prediction model using real-time thermal error data, so that the CNN-BLRNET thermal error prediction model can adapt to the operating conditions of each edge node.
[0051] Furthermore, the edge node is provided with an intelligent gateway, and the intelligent network point includes a processor module, a ZigBee communication module, a WIFI communication module and a power management module. The ZigBee communication module is used to communicate with each data acquisition node of the data acquisition system, and the WIFI communication module is used to exchange information with the cloud center.
[0052] Furthermore, in condition monitoring and fault diagnosis, effective separation and feature extraction of harmonic and impulse components need to be achieved by:
[0053] Assume that the signal s is linearly composed of K components with different shapes, that is, The basic model of morphological component analysis is based on the following two assumptions:
[0054] (1) For a given signal s k For each component of Usually L k >>N, then the problem to be solved is:
[0055]
[0056] Among them, α k represents the sparse representation coefficient; represents a sparse solution;
[0057] (2) For a given signal s l Each component of , when k≠l:
[0058]
[0059] Among them, α l represents the non-sparse representation coefficient; represents a non-sparse solution;
[0060] The above two assumptions ensure that the dictionary T k It is possible to distinguish signals of different morphological components, and the sparse decomposition of the signal s is obtained by solving the following problem
[0061]
[0062] This problem is a non-convex NP-hard problem, which is transformed into the next convex optimization problem:
[0063]
[0064] The signal to be analyzed s is the harmonic component s A and the impact component s B The linear combination of , the model satisfies the following three conditions:
[0065] (1) For the harmonic component s A , constructing an overcomplete dictionary through harmonic wavelet transform
[0066]
[0067] in, α represents the sparse representation coefficient; represents a sparse solution;
[0068] Get the sparse solution and get the sparse representation coefficient α A ;
[0069] (2) For the shock wave component s B , we get a non-sparse solution, which means that the dictionary T A Two separate components can be distinguished:
[0070] (3) For the shock wave component s B , an overcomplete dictionary constructed by multiwavelet transform The shock wave component s can be sparsely represented B and harmonic components A , to solve the problem of multi-component sparse separation of signals:
[0071]
[0072] Convert it into a convex optimization problem:
[0073]
[0074] and The sparse matrix of L is the norm of ||α A ||1+||α B ||1 at s = T A α A +T B α B The optimal solution under the constraints; if the signal is noisy, the signal cannot be simply decomposed into harmonic components s A and the impact component s B , needs to be expanded to:
[0075]
[0076] The parameter ε characterizes the noise level in the signal s. In order to obtain the optimal solution, the Lagrange multiplier method is used to give a suitable Lagrange operator, and the constrained optimization problem is transformed into an unconstrained optimization problem:
[0077]
[0078] where λ = [λ A ,λ B ] T represents the Lagrangian operator, λ is a function of s, α and ε; the objective function of morphological component analysis is:
[0079]
[0080] Where J(·) represents the objective function of morphological component analysis.
[0081] Furthermore, thermal error control requires first predicting the thermal error. The method for predicting the thermal error using the CNN-BLRNET thermal error prediction model is as follows:
[0082] (1) Input the thermal error data into the input layer of the CNN-BLRNET thermal error prediction model.
[0083] (2) Use the CNN layer to process the data to extract important local features of the data; use the maximum pooling layer to further extract the local maximum features and obtain the initial feature information of the data;
[0084] (3) Extract the temporal characteristics of data using the BLRNET model;
[0085] (4) The data enters the dropout layer, flatten layer, and dense layer in turn, converting the two-dimensional vector into a one-dimensional vector for output to obtain the thermal error prediction value.
[0086] The beneficial effects of the present invention are:
[0087] The CNN-BLRNET thermal error prediction model of the present invention can extract spatial features in thermal error data by setting a CNN layer, and forms a BLRNET model by embedding a BILSTM model into a RESNET network. The BILSTM model extracts temporal features in thermal error data, and utilizes the jump connection part in the RESNET network to avoid the loss of some important information in the data transmission process, thereby improving the prediction accuracy and robustness. By adding a ReLU layer, the collinearity of the thermal error data can be reduced to avoid overfitting; by adding a maximum pooling layer after the convolution layer, the maximum information of the local features can be extracted and the size of the RESNET network can be reduced, thereby reducing the number of model parameters and improving the calculation speed; therefore, the CNN-BLRNET thermal error prediction model of the present invention can extract spatial features and temporal features of thermal error data, and can improve the prediction accuracy and robustness.
[0088] Traditional deep learning models are only built for one working condition and have insufficient generalization capabilities; traditional modeling methods rely on large-capacity data to improve model accuracy, but for thermal error prediction, due to the limitations of test instruments and equipment, the amount of measured data is limited. Therefore, how to use a small data set to train an excellent model is an important issue. The error model should be robust and have strong generalization capabilities to adapt to different working conditions. In addition, due to changes in the research object, retraining the error model requires a lot of computing time and resources. The transfer learning method of the CNN-BLRNET thermal error prediction model of the present invention, and by transferring the structure and parameters of the previously trained CNN-BLRNET thermal error prediction model, a new model is trained for different working conditions, and both training efficiency and robustness are improved.
[0089] The intelligent integrated framework of the present invention, by setting up a data acquisition system and a data computing system, uses the data acquisition system to collect machine tool working status data and workpiece quality data, and the collected data is transmitted to the edge node and the cloud center for storage; at the same time, the computing system is used, and self-learning is performed with the support of machine learning, reinforcement learning, deep learning and transfer learning to establish the fault diagnosis model, thermal error model, three-dimensional model, control strategy and knowledge of the machine tool as a digital twin model, and then provide data services, including thermal error control, state monitoring and fault diagnosis. For state monitoring, feature extraction is performed on the collected signal; for fault diagnosis, appropriate measures are taken to ensure the processing process; for error control, the servo system is adjusted in real time; that is, the intelligent integrated framework of the present invention integrates error control and error diagnosis services, and can achieve high-quality processing. BRIEF DESCRIPTION OF THE DRAWINGS
[0090] In order to make the purpose, technical solution and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration:
[0091] Figure 1 A system diagram of an embodiment of the intelligent integration framework of the present invention;
[0092] Figure 2 This is the principle block diagram of the data acquisition system;
[0093] Figure 3 It is a wireless tree structure network diagram;
[0094] Figure 4 This is the structure diagram of the intelligent gateway node;
[0095] Figure 5 is the cluster head node graph;
[0096] Figure 6 Provides a diagram of the model training system architecture;
[0097] Figure 7 is a schematic diagram of the thermal deformation of the main shaft;
[0098] Figure 8 It is a structural schematic diagram of an embodiment of the CNN-BLRNET thermal error prediction model of the present invention;
[0099] Fig. 9 It is the structural diagram of LSTM unit;
[0100] Fig.10 This is the schematic diagram of BILSTM;
[0101] Fig.11 This is the RESNET network structure diagram;
[0102] Fig.12 This is the structure diagram of the BLRNET model;
[0103] Fig.13 This is the principle diagram of transfer learning;
[0104] Fig.14 The transfer learning model TLM of CNN-BLRNET structure;
[0105] Fig.15 Workflow diagram for condition monitoring and fault diagnosis services;
[0106] Fig.16 Thermal error control working process;
[0107] Fig.17 The thermal error of the main axis; (a) thermal error data of working condition #1; (b) thermal error data of working condition #2;
[0108] Fig.18CNN-BLRNET prediction performance; (a) fitting error data for condition #1; (b) prediction error data for condition #2;
[0109] Fig.19 is the fitting error; Fig. 20 is the prediction error;
[0110] Fig.21 is the loss function;
[0111] Fig. 22 TLM of CHD25A; (a) Model comparison #Ⅰ; (b) Model comparison #II; (c) Model comparison III; (d) Model comparison IV;
[0112] Fig.23 TLM of the TGK46100 machine tool spindle; (a) Model comparison #Ⅰ; (b) Model comparison #II; (c) Model comparison III; (d) Model comparison IV;
[0113] Fig.24 The actual picture of the processing experiment; (a) processing technology; (b) processing workpiece;
[0114] Fig.25 are vibration signals of different damages; (a)-(h) are vibration signals of states 1-8 respectively;
[0115] Fig.26 The frequency domain spectra in different states; (a)-(h) are the frequency domain spectra of states 1-8 respectively;
[0116] Fig. 27 The impact part of the arc surface cam in the damaged state; (a)-(h) are the vibration impact parts in states 1-8 respectively;
[0117] Fig.28 The corresponding relationship between the impact component in normal operation and motion cycle;
[0118] Fig.29 It is the correspondence between abnormal operation and impact components in the motion cycle. DETAILED DESCRIPTION
[0119] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.
[0120] 1. Intelligent integration framework
[0121] like Figure 1As shown, the intelligent integration framework of this embodiment includes a data acquisition system and a data computing system. The data acquisition system collects machine tool working status data and workpiece quality data; all collected data are transmitted to edge nodes and cloud centers through local area networks and industrial Internet respectively; the computing system performs data calculation on the collected data, performs self-learning with the support of machine learning, reinforcement learning, deep learning and transfer learning, establishes the fault diagnosis model, thermal error model, three-dimensional model, control strategy and knowledge of the machine tool as a digital twin model, and provides data services including state monitoring, fault diagnosis and thermal error control. Specifically, the working status data of the machine tool includes internal data and external data, and the collection of internal and external data is realized through self-perception. Internal and external data include processing process information, location information and processing parameters. With the help of local area networks and industrial Internet, the data is transmitted to edge nodes and cloud centers, and through data calculation, and self-learning with the support of machine learning, reinforcement learning, deep learning and transfer learning, the fault diagnosis model, thermal error model, three-dimensional model, control strategy and knowledge of the machine tool are established as a digital twin model, and then data services can be provided, including thermal error control, state monitoring and fault diagnosis. For condition monitoring, feature extraction is performed on the collected signals; for fault diagnosis, appropriate measures can be taken to ensure the processing process; for error control, it is used to make real-time adjustments to the servo system.
[0122] 1.1 Self-perception process
[0123] 1.1.1 Data Collection
[0124] The working status information of the machine tool and the quality information of the workpiece are the basis for the monitoring and error control of the entire digital machining process. In order to obtain comprehensive operating information of the machining process, a combination of multiple sensors is used to collect machine tool status information. The reasonable selection and optimal configuration of various sensors are important conditions for obtaining the required information. Incorrect selection and configuration may affect the integrity and reliability of the collected information and may reduce the reliability of state identification and error control results. Therefore, how to optimize the configuration of sensors and improve the reliability and fault tolerance of sensor networks has become a problem of concern. Generally speaking, the physical quantities reflecting the operating status of machine tools include internal data and external data. External data mainly refers to vibration, cutting force, acoustic emission, temperature, torque and speed, etc. External data is the main data source; internal data is millisecond-level high-frequency data with a large amount of data. Internal data mainly includes motor current, motor voltage, encoder displacement and grating signal. Specifically, machine tool external data includes the following aspects:
[0125] (1) Motion axis data: current, position, speed, temperature, etc.;
[0126] (2) Spindle data: power, torque, speed, temperature, etc.
[0127] (3) Operation status data: temperature, vibration, PLC, I / O, etc.;
[0128] (4) Operation data: start, shut down, power off, emergency stop, etc.;
[0129] (5) Program data: program line number, tool, processing time, etc.
[0130] The working status information of the machine tool is easy to measure on the machine tool, such as encoder signals, servo motor currents, vibration signals, acoustic emission signals, and vibration signals of the transmission system. Generally speaking, the above signals are selected as working condition monitoring signals to form characteristic quantities of the machine state. In this embodiment, the dimensions and microscopic morphology of the machined workpiece are measured using a three-coordinate measuring machine and a surface roughness tester, respectively. There is a correlation between the processing quality, accuracy, and efficiency and the collected data. Therefore, the processing status of the machine tool can be indirectly obtained by sensing internal and external data. Figure 2 A typical multi-sensor network system for collecting machine status information is shown. Hall sensors are used to detect the current of the servo motors of the spindle and feed drive system, encoders and grating rulers are used to obtain position information, and acceleration sensors are used to pick up vibration information of the spindle box and worktable. In addition, all data are sent to the data sub-station via wireless and wired transmission, and an intelligent gateway is set as an edge node to provide error prediction services and fault diagnosis services. Data is uploaded to the cloud center via industrial Ethernet to facilitate data analysis and use. Data storage is performed by Hadoop, and data calculation is based on Spark.
[0131] 1.1.2 Sensor network configuration
[0132] According to the different functions of different nodes in the wireless network, the devices are divided into coordinators, routing nodes and terminal nodes. The coordinator is the center of the ZigBee sensor network, playing the role of establishing the network, maintaining the server and managing other nodes in the network. The routing node is equivalent to the role of a router, forwarding data from the lower layer nodes to the upper layer, and passing instructions from the upper layer to the lower layer, and can be used to expand the signal range. The terminal node is the bottom layer, realizing data collection and uploading. Both the coordinator and the router are full-function nodes.
[0133] In the sensor network, the terminal node will be equipped with sensors and collect the physical parameters of the machine tool and send the collected information to the upper node. The routing node plays the role of expanding the network range, wirelessly searching for nearby nodes that can be added, and can transmit data and request data transmission. The role of the coordination node is reflected in establishing a wireless sensor network, transmitting instructions from the upper computer to the lower node, managing each node, saving node information and assigning addresses. The terminal node is the most basic part of the wireless network, which determines whether the collected data is accurate and stable. It mainly consists of four parts: controller module, power module, sensor module and serial communication module.
[0134] Each machine tool has multiple wireless sensor nodes. The nodes on each machine tool use the ZigBee protocol to independently form a wireless network. This embodiment adopts a cluster network structure, such as Figure 3 As shown. On the one hand, the wireless sensor network forms a network for the wireless sensors and intelligent gateway nodes currently waiting to join the network; on the other hand, the number of acquisition nodes under each ZigBee communication channel is set, and then the intelligent gateway node waits for the cloud center to send the networking command of the acquisition node, so as to realize the self-organizing network of the acquisition node under each ZigBee communication channel, thereby forming a tree network topology. According to the acquisition parameter configuration, the acquisition under each ZigBee channel is performed, the acquisition parameters of the node are configured, and the signal acquisition starts until the command to start the acquisition is sent. Finally, the signals in each ZigBee network topology are aggregated into each intelligent gateway node for data sorting, analysis and processing, and local storage, and the corresponding data is transmitted back to the cloud center as needed.
[0135] 1.1.3 Edge computing intelligent gateway node configuration
[0136] (1) Intelligent gateway node structure
[0137] The structure of the intelligent gateway node is as follows Figure 4As shown. The intelligent gateway node for edge computing of wireless sensor networks consists of a processor module, a ZigBee communication module, a WIFI communication module, a power management module, etc. The processor module uses the CPU I.MX6ULL of the ARMCortex-A7 low-power processor architecture as the main control core, and expands 8GB 8MMC and 512MBDDR3L. The ZigBee communication module is used to communicate with the cluster acquisition nodes in the entire cluster network. The WIFI communication module is used to exchange information with the data center. The processor module is the core of the edge computing intelligent gateway node. It is responsible for the data transmission and reception scheduling between the ZigBee communication module and the WIFI communication module, and then aggregates, converts and stores the data in the cluster, and completes the intelligent processing tasks according to the deployed computing model. The power management module is responsible for converting the rated voltage provided by the battery into the voltage required by each functional module to maintain the power supply of the entire node.
[0138] The ZigBee communication module is mainly responsible for communicating with the acquisition nodes, including self-organizing networks, beacon transmission, data transmission, etc. CC2530F256 is used as the control core of the ZigBee wireless communication module, and is equipped with CC2592 as a radio frequency (RF) power amplifier chip. CC2530F256 is a ZigBee RF chip developed by TI based on the IEEE802.15.4 protocol. It operates in the 2.4GHz wireless frequency band, provides 16 different communication channels, and has a transmission rate of 250kbps. CC2592 is a RF front-end chip developed by TI. Its function is to amplify the signal in the ZigBee wireless sensor network by increasing the transmission power and improving the receiving sensitivity. The ZigBee protocol stack based on TIMAC runs on the ZigBee communication module.
[0139] The WIFI communication module is mainly used to exchange information with the host, including network parameter settings, acquisition parameter settings, data analysis and processing result return, etc. The wireless communication module in the edge computing intelligent gateway node of the wireless sensor network uses the USB WIFI-RTL8188eus produced by Realtek, and the edge computing intelligent gateway node provides a USB HOST interface to support the module.
[0140] (2) Gateway topology construction and management
[0141] The traditional star network topology has the problems of heavy communication load and insufficient coverage, and its communication quality is easily affected by the environment. The emergence of routing protocols and multi-hop networks has solved this problem to a certain extent. However, each node in the network needs to reserve channels and transmit data based on the carrier sense multiple access / collision avoidance mechanism of the IEEE802.15.4 wireless communication protocol. The traditional multi-hop network causes the original narrow-edge wireless network bandwidth to face huge transmission pressure, and the communication load is still very high.
[0142] Multiple gateway nodes are used as cluster head nodes to manage each cluster in the cluster network, such as Figure 5 As shown in the figure. The gateway node communicates wirelessly with the data center based on the IEEE 802.11 protocol. Compared with the traditional gateway, it has the characteristics of passive wireless and independent layout. The edge computing smart gateway exists as the closest hop to the device in the multi-hop network. It uses its own data processing function to process and forward the data collected by the collection node. Routing nodes can be added between the edge computing smart gateway and the data center according to the communication situation. Due to the small amount of data to be forwarded, the overall communication pressure of the network and the storage and computing pressure of the data center are reduced.
[0143] For ZigBee wireless networks, it is unreliable to distinguish different communication channels by the physical direction differences of the edge computing smart gateway, and adjacent clusters may interfere with each other due to network fluctuations. In the designed topology, different communication channels are used for networking and data interaction to avoid interference between adjacent clusters. In addition, the data transmission work of the large nodes at the outermost layer of the network is handled in a distributed manner, thereby dispersing the bandwidth pressure of the entire network.
[0144] The data exchange between the edge computing intelligent gateway and the workstation is based on IEEE 802.11. The data center allocates a separate ZigBee communication channel for the edge computing intelligent portal for cluster network management. The ZigBee protocol provides 16 different channels. In order to avoid possible interference between them, spaced communication channels are selected, and a total of 8 channels are available. In the cluster network, each intelligent gateway of the edge computing performs self-organizing network and data transmission under the specified communication channel. The edge computing intelligent gateway receives the control command of the workstation from top to bottom, and through command parsing, encapsulates the command packet using the control command protocol of the acquisition node, and then sends the command packet to the acquisition node. The acquisition node aggregates the data to its corresponding cluster head node from bottom to top, and the edge computing intelligent gateway acts as the cluster head node; the edge computing intelligent gateway node decapsulates, reconstructs, stores, and processes the data packet, and then returns the result to the workstation.
[0145] 1.2 Data Calculation
[0146] 1.2.1 Data Calculation Principles
[0147] Based on the cloud-edge collaboration technology, the collected data is preprocessed. Then the historical data is transmitted to the cloud center, the error model is preliminarily trained, and the preliminarily trained error model is deployed to the edge node. In addition, the real-time data is used to locally train the preliminarily trained model to ensure that the model adapts to the latest operating conditions and obtains a well-trained model. The data calculation process is as follows Figure 6 As shown, historical data is used as input to perform preliminary training in the cloud center. In addition, real-time data is used as input to perform fine training on the edge node.
[0148] 1.2.2 Online monitoring and fault diagnosis
[0149] In condition monitoring and fault diagnosis, it is necessary to achieve effective separation and feature extraction of harmonic and impact components. In order to achieve effective separation and feature extraction of harmonic and impact components, the morphological component analysis theory is introduced into the multi-component separation of mechanical faults. This embodiment proposes a multi-component sparse separation method based on Q-switched wavelet and multi-wavelet transform. The method uses Q-switched wavelet and multi-wavelet transform to construct a sparse representation dictionary. Based on the basic idea of morphological component analysis, a multi-component sparse separation objective function is established. Finally, the effective separation of harmonic and impact components is achieved.
[0150] Assume that the signal s is linearly composed of K components with different shapes, that is, The basic model of morphological component analysis is based on the following two assumptions:
[0151] (1) For a given signal s k For each component of Usually L k >>N, then the problem to be solved is:
[0152]
[0153] Among them, α k represents the sparse representation coefficient; represents a sparse solution;
[0154] The definition is essentially that the signal s k Complete breakdown of the portions.
[0155] (2) For a given signal s l Each component of , when k≠l:
[0156]
[0157] Obtain a non-sparse solution. Among them, α l represents the non-sparse representation coefficient; represents a non-sparse solution;
[0158] The above two assumptions ensure that the dictionary T k It is possible to distinguish signals of different morphological components, and the sparse decomposition of the signal s is obtained by solving the following problem
[0159]
[0160] This problem is a non-convex NP-hard problem, which is transformed into the next convex optimization problem:
[0161]
[0162] In order to achieve effective separation and feature extraction of harmonic components and shock components, two dictionaries that can sparsely represent harmonic components and shock components are constructed based on the idea of morphological component analysis. Then, the objective function of the multi-component sparse separation method is constructed using the morphological component analysis method. The signal to be analyzed s is the harmonic component s A and the impact component s B The linear combination of , the model satisfies the following three conditions:
[0163] (1) For the harmonic component s A , constructing an overcomplete dictionary through harmonic wavelet transform
[0164]
[0165] in, α represents the sparse representation coefficient; represents a sparse solution;
[0166] Get the sparse solution and get the sparse representation coefficient α A ;
[0167] (2) For the shock wave component s B , and we get a non-sparse solution (i.e. larger), which means that the dictionary T A Two separate components can be distinguished:
[0168] (3) For the shock wave component s B , an overcomplete dictionary constructed by multiwavelet transform The shock wave component s can be sparsely represented B and harmonic components A , to solve the problem of multi-component sparse separation of signals:
[0169]
[0170] This is a non-convex NP-hard problem, which is transformed into a convex optimization problem:
[0171]
[0172] That is to say, and The sparse matrix of L is the norm of ||α A ||1+||α B ||1 at s = T A α A +T B α B The optimal solution under the constraints; if the signal is noisy, the signal cannot be simply decomposed into harmonic components s A and the impact component s B , needs to be expanded to:
[0173]
[0174] The parameter ε characterizes the noise level in the signal s. In order to obtain the optimal solution, the Lagrange multiplier method is used to give a suitable Lagrange operator, and the constrained optimization problem is transformed into an unconstrained optimization problem:
[0175]
[0176] where λ = [λ A ,λ B ] T represents the Lagrangian operator, λ is a function of s, α and ε; the objective function of morphological component analysis is:
[0177]
[0178] Where J(·) represents the objective function of morphological component analysis.
[0179] 1.2.3 Thermal Error Prediction
[0180] In order to simplify the temperature response of the spindle, the spindle is regarded as a one-dimensional slender axis, such as Figure 7 As shown. Assume that the spindle is a uniform and thin rod, its sides are insulated, and there is no convective heat transfer with the external environment. It is also assumed that the temperature distribution inside the spindle is the same on the same cross section, and the temperature function T is only related to the coordinate x and time T. The bearing is installed at the left end of the shaft and acts as a stable heat source, resulting in axial error. The heat conduction equation is:
[0181]
[0182] Where L, k, T(x, t), ρ, t, x and c represent length, scale factor, temperature, density, operating time, axial position and specific heat capacity, respectively.
[0183] When t is constant and the shaft temperature is equal to the external temperature, it is defined as the initial condition. That is: T0 represents the external temperature.
[0184] The Fourier transform is used to solve the heat conduction equation, and the final solution is:
[0185]
[0186] Among them, ξ is the intermediate conversion variable.
[0187] According to the above equation, we can get The temperature response is calculated as:
[0188]
[0189] The thermal error is:
[0190]
[0191] The thermal error at time 0 is:
[0192] ΔL(0)=0
[0193] The thermal error at time Δt is:
[0194]
[0195] The thermal error at 2Δt is:
[0196]
[0197] The thermal error at 3Δt is:
[0198]
[0199] The thermal error at time nΔt is:
[0200]
[0201] The thermal error is closely related to the historical thermal error at previous moments, indicating that the thermal error has memory and nonlinear characteristics. It can be applied to time series prediction networks.
[0202] 1.2.3.1CNN-BLRNET Thermal Error Prediction Model
[0203] like Figure 8As shown, the CNN-BLRNET thermal error prediction model of this embodiment includes a first input layer, a CNN layer, a BLRNET model, a dropout layer, a flatten layer, a dense layer and a first output layer arranged in sequence. The BLRNET model is formed by embedding a BILSTM model into a RESNET network, and the BLRNET model includes a second input layer, a ReLU layer and a second output layer, at least two RESNET units connected in series are arranged between the ReLU layer and the output layer, and the RESNET unit includes a direct mapping part and a jump connection part, the direct mapping part includes a convolution layer I, a maximum pooling layer I and a BILSTM layer arranged in sequence, the jump connection part includes a convolution layer II and a maximum pooling layer II arranged in sequence, and an addition operator is arranged between the BILSTM layer and the maximum pooling layer II.
[0204] (1) BILSTM model
[0205] The BILSTM model includes an input layer, a forward LSTM layer, a backward LSTM layer, and an output layer. The forward LSTM layer and the backward LSTM layer include several LSTM units, such as Fig. 9 As shown, the principle of the LSTM unit is:
[0206] f t =σ(W f [h t-1 , x t ]+b f )
[0207] i t =σ(W f [h t-1 , x t ]+b i )
[0208] C t =tanh(W f [h t-1 , x t ]+b c )
[0209]
[0210] O t =σ(W o [h t-1 , x t ]+b O )
[0211] h t =O t *tanh(C t )
[0212] Where t and t-1 are the current and previous moments respectively; f t 、i t and O t They are forget gate, input gate and output gate respectively; x t , C t and h t are the input, unit state and output at the current moment; x t-1 , C t-1 and h t-1 are the input, unit state and output at the previous moment respectively; is the candidate cell state; W f , W i , W c and W o is the weight matrix; b f , b i , b c and b o is the bias term; * denotes element-wise multiplication; σ(·) and tanh(·) are activation functions.
[0213] LSTM cannot fully utilize the temporal characteristics of future data. Thermal error has memory behavior, and the output at the next moment is related to historical and future data. This embodiment uses BILSTM to perform bidirectional prediction. The working principle of BILSTM is as follows: Fig.10 As shown in the figure. The principle of BILSTM is:
[0214] h t =f(w1*x t +w2*h t-1 )
[0215] h′ t =f(w3*x t +w4*h′ t-1 )
[0216] o t =g(w5*h t +w6*h t )
[0217] Where f(·) represents the forward operation function; g(·) represents the backward operation function; h t represents the forward output; h′ t represents the backward output; w1, w2, w3, w4, w5 and w6 represent the weight matrices.
[0218] (2) BLRNET Network
[0219] like Fig.11As shown in Figure 1, RESNET is mainly composed of convolutional layers and activation layers. The convolutional layer extracts local features. The rectified linear unit (ReLU) reduces collinearity, enhances nonlinear modeling capabilities, and prevents model overfitting. The existing RESNET is divided into two parts, namely the direct mapping part and the skip connection part. The direct mapping part is the main body of RESNET, which is used to extract the main feature information from the initial data; the skip connection part retains the initial information to prevent the model from losing feature information when extracting the main features; thereby ensuring the integrity of the data and improving the prediction accuracy.
[0220] The original error data is X. For the direct mapping part, the convolution layer is:
[0221]
[0222] Among them, X1 represents the input of the direct mapping part; W1 represents the convolution weight matrix; Represents the input value; Represents the convolution kernel value; b1 represents the bias parameter in the convolution operation; represents the actual convolution operation bias; f1(·) represents the actual convolution operation; conv1d(·) represents the convolution operation.
[0223] In the activation function layer:
[0224]
[0225] In the above example, X1 represents the input data; relu(·) represents the ReLU activation function, and x i Represents the i-th input data.
[0226] For the direct mapped part:
[0227]
[0228] Among them, F(X,W) represents the output of the directly connected part; X′1 represents the input of the directly mapped part; W2 represents the convolution weight matrix; Represents the input value; represents the convolution kernel value; b2 represents the bias parameter in the convolution operation; represents the actual convolution operation bias; f1(·) represents the actual convolution operation; conv1d(·) represents the convolution operation.
[0229] For the crossover part, the mathematical expression is:
[0230]
[0231] G(X) represents the output of the crossover part; W3 represents the convolution weight matrix; b3 represents the bias parameter in the convolution operation; Represents the input value; Represents the convolution kernel value; Indicates the actual convolution operation bias.
[0232] The output of the BLRNET unit is the linear sum of the outputs of the direct mapping part and the skip connection part:
[0233] H(X)=F(X,W)+G(X)
[0234] Among them, X is the input of the BLRNET unit; H(X) is the output of the BLRNET unit; F(X, W) is the output of the direct mapping part; G(X) is the output of the cross-connection part; W is the weight matrix, representing the convolutional layer in this embodiment.
[0235] Thermal error is spatiotemporal data. CNN can extract spatial features, and BILSTM can extract temporal features. In order to fully extract spatiotemporal features, RESNET is combined with BILSTM to construct BLRNET. Specifically, BILSTM is embedded into RESNET and placed after the convolutional layer of RESNET, such as Fig.12 As shown. The introduction of skip connections avoids the loss of some important information during data transmission, thereby improving prediction accuracy and robustness. The data was normalized in the previous data processing section, the batch normalization layer was removed, and the ReLU layer was added to reduce collinearity and avoid overfitting. A maximum pooling layer was added after the convolution layer to extract the maximum information of local features and reduce the size of BLRNET, thereby reducing the number of model parameters and improving calculation speed. The data was converted to nonlinearity through the activation function ReLU, and:
[0236]
[0237] Where X represents the input data; relu(·) represents the ReLU activation function, x i Represents the i-th input data;
[0238] The data can be input as X1 directly into the mapped part, and the data can also be input as X2 into the remaining part. For convolutional layer I:
[0239]
[0240] Among them, X1 represents the input of the direct mapping part; W1 represents the convolution weight matrix; Represents the input value; Represents the convolution kernel value; b1 represents the bias parameter in the convolution operation; represents the actual convolution operation bias; f1(·) represents the actual convolution operation; conv1d(·) represents the convolution operation.
[0241] For the max pooling layer I:
[0242]
[0243] Where X′1 represents the output of the maximum pooling layer I; maxpooling(·) represents the pooling operation; Represents the input value;
[0244] The BILSTM layer extracts temporal features:
[0245] h t =f(w1*X′1+w2*h t-1 )
[0246] h′ t =f(w3*X′1+w4*h′ t-1 )
[0247] The output of the direct mapping section is:
[0248] F(X,W)=h t +h′ t
[0249] Among them, F(X,W) represents the output of the direct mapping part;
[0250] For convolutional layer II:
[0251]
[0252] Among them, X2 represents the input of the direct mapping part; W2 represents the convolution weight matrix; Represents the input value; represents the convolution kernel value; b2 represents the bias parameter in the convolution operation; b′ j l represents the actual convolution operation bias; f1(·) represents the actual convolution operation; conv1d(·) represents the convolution operation;
[0253] The output of the jump connection part is:
[0254]
[0255] Among them, G(X) represents the output of the jump connection part; Represents the input value;
[0256] The output of the BLRNET unit is: H(X) = F(X,W) + G(X)
[0257] Where H(X) represents the output of the BLRNET unit.
[0258] The core principle of BLRNET is the same as that of RESNET. The jump connection part is used to save the complete spatiotemporal information, thereby ensuring the network accuracy. BLRNET also introduces a maximum pooling layer to extract the local maximum feature. The most important feature of BLRNET is the introduction of BILSTM into the traditional RESNET, which is used to obtain temporal features. Finally, the extraction of spatiotemporal information is realized. The introduction of RESNET avoids the excessive loss of important features during data transmission, thereby improving accuracy.
[0259] (3) CNN-BLRNET thermal error prediction model
[0260] To evaluate the CNN-BLRNET thermal error prediction model, the following prediction accuracy metrics are used:
[0261]
[0262]
[0263]
[0264]
[0265] Where n is the length; is the predicted data; y i It is real data; is the average value.
[0266] The number of BLRNET units needs to be optimized. By using different numbers of layers for error prediction, as shown in Table 1, when the number of BLRNET units is 4, the model accuracy and efficiency are optimal.
[0267] Table 1. The influence of network depth on prediction accuracy
[0268]
[0269] The method for predicting thermal error using the CNN-BLRNET thermal error prediction model is:
[0270] (1) Input the thermal error data into the input layer of the CNN-BLRNET thermal error prediction model.
[0271] (2) Use the CNN layer to process the data to extract important local features; use the maximum pooling layer to further extract the local maximum features and obtain the initial feature information;
[0272] (3) Extract the temporal characteristics of data using the BLRNET model;
[0273] (4) The data enters the dropout layer, flatten layer, and dense layer in turn to reduce the dimension of the output data of the output model, convert the two-dimensional vector into a one-dimensional vector for output, and obtain the thermal error prediction value.
[0274] (4) Transfer learning method of CNN-BLRNET
[0275] Traditional deep learning models are only built for one working condition and have insufficient generalization capabilities; traditional modeling methods rely on large-capacity data to improve model accuracy, but for error prediction, the amount of data is limited. Therefore, how to use very few data sets to train an excellent model is an important issue. The error model should be robust and have strong generalization capabilities to adapt to different working conditions. In addition, due to changes in the research object, retraining the error model requires a lot of computing time and resources. The transfer learning method of the CNN-BLRNET thermal error prediction model in this embodiment, and by transferring the structure and parameters of the previously trained CNN-BLRNET thermal error prediction model, a new model is trained for different working conditions, and the training efficiency and robustness are improved. Fig.13 As shown in Figure 1, a new model is trained for machine tool #II by transferring the structure and parameters of the previously trained CNN-BLRNET model for machine tool #I. Although the thermal errors of machine tools #I and #II are different, the thermal error mechanism is the same. The CNN-BLRNET model is pre-trained on machine tool #I. Then, a transfer learning model (TLM) is established by retraining the dense layers of CNN-BLRNET and freezing the other layers, as shown in Figure 1. Fig.14 shown.
[0276] 1.3 Data Services
[0277] 1.3.1 Online monitoring and fault diagnosis services
[0278] Processing state identification and fault diagnosis are the decision-making links of the entire monitoring system. It analyzes and infers based on the symptoms of the monitored and diagnosed objects, and finally determines whether the processing process and the machine are in a normal state. If an abnormality occurs, the location and cause of the fault must be found, and then an effective intervention decision is made. At present, the most common and reliable diagnostic decision-making process is still carried out by domain experts based on empirical knowledge. With the development and improvement of artificial intelligence, expert systems based on symbolic reasoning and intelligent diagnostic systems based on numerical calculations have been developed. Fig.15The working process of the condition monitoring and fault diagnosis service is shown. First, a ZigBee sensor network is constructed to obtain the working condition information of the machine tool. The working condition information includes vibration, cutting force, temperature, acoustic emission and other signals. The quality information of the workpiece includes the geometric error, waviness, roughness, etc. of the machined surface, and these signals provide data support for feature extraction. The second is feature extraction. Due to the complexity of the machining process, the collected information usually contains a lot of noise, and the collected data should be reprocessed, that is, the feature extraction of the signal is necessary. According to the characteristics of the signal, appropriate signal processing methods are used to extract signal features, including wavelet analysis, S transform and stochastic resonance. Next, the extracted signal features and workpiece quality information are used to identify the current machining process and operating status. Then the intelligent integration framework makes diagnosis and decision. If the machining process is normal, the current state will be maintained and the machining will continue. If an abnormal state is found, the fault must be diagnosed and the error source must be tracked, and then corresponding measures must be taken. For example, the machine is shut down for maintenance, or the machining parameters are adjusted.
[0279] 1.3.2 Thermal Error Control Service
[0280] Fig.16 The working process of the thermal error control service is shown. After obtaining the thermal error of the machine tool under the current working condition, the intelligent gateway is used to reprocess the collected data and use the real-time data to finely train the error model. The historical data is transmitted for preliminary training of the error model, and preliminary training is performed in the cloud center. The cloud center has powerful computing and storage capabilities, so the computing efficiency is high. The edge node uses real-time data as input to finely train the error model. In this way, frequent comprehensive training is avoided. The calculation module is constructed based on the edge node, and the predicted error is obtained through the error prediction model; the compensation value is the inverse of the predicted error, and the compensation axis is determined according to the direction of the predicted error. The compensation axis is then transmitted to the PLC of the CNC system for thermal error control, and the CNC system corrects the position of the feed drive system. Through the design of the intelligent integration framework, real-time and historical data are reasonably used. The update of the transfer learning model of the CNN-BLRNET model is divided into fine training and preliminary training. The historical data is used for preliminary training in the cloud center and fine training is performed in the edge node. The transfer learning model of the CNN-BLRNET model provides reasonable efficiency for the entire intelligent integration framework.
[0281] 2. Case Study
[0282] 2.1 Thermal Error Control
[0283] 2.2.1 Error model training and prediction
[0284] The milling and turning machining center is used as the research object, and closed-loop control is performed on the X, Y and Z axes. The X, Y and Z axes are driven by motors through synchronous belts, and the transmission ratio of the synchronous belts is 20:40. The screw leads of the X, Y and Z axes are 16mm, 20mm and 20mm respectively. This embodiment uses a total of five displacement sensors to collect data by collecting thermal errors. Fig.17 (a) shows the filtered interpolation error data for operating condition #1. Fig.17 (b) shows the error data set for operating condition #2. The training data for this experiment used the thermal error data from sensor S3.
[0285] The network model is built based on Python's Keras, and the Tensorflow module is used to build CNN-BLRNET. There are many hyperparameters, such as learning rate, batch size, time step, optimizer, and epoch size. These hyperparameters are determined, the epoch size is 50, the optimizer is Adam, the learning rate is 0.0001, and the effects of the time step on the prediction accuracy and convergence are listed in Table 2. When the time step is 10, the prediction accuracy and convergence time are 95.26% and 32.7404s, respectively. In addition, when the time step is 10, the prediction accuracy is much higher than when the time step is 5 and 15, respectively. Therefore, in this embodiment, the time step is set to 10.
[0286] Table 2 Effect of time step
[0287]
[0288] The prediction accuracy and convergence are affected by the batch size, as shown in Table 3. The smaller the batch size, the better the prediction performance. However, if the batch size is too small, the time consumption will increase significantly. When the batch size is 32, the prediction accuracy is 95.26% and the convergence time is 21.724s. When the batch size is 32, the prediction accuracy is much higher than that when other batch sizes are used, and the time consumption is reasonable. Therefore, the batch size is set to 32.
[0289] Table 3 Effect of batch size
[0290]
[0291] CNN and BLRNET contain convolutional layers. For convolutional layers, the most important parameter is the kernel size, which is the range of convolutions for each convolution operation. By increasing the number of kernels, more features can be extracted. However, it should also be kept within a valid range to prevent overfitting. Another important parameter is the step size of the kernel sliding, also known as the sliding step size. The sliding step size controls the number of potential features between two adjacent data collected. The parameters of the convolution kernels in CNN and BLRNET are the same, and the number of kernels and the sliding step size should be consistent. The input dimension is equal to the output dimension. The parameters of the proposed CNN-BLRNET model are optimized through multiple experiments, and then the convolutional layer is finally determined, as shown in Table 4.
[0292] Table 4 Structure and parameters of convolutional layer
[0293]
[0294] It is necessary to select a suitable number of neurons for BILSTM. The number of neurons should not be too large, and multiple comparative experiments are required to predict accuracy and time consumption. Finally, the number of neurons is determined to be 16. The main parameters of the entire CNN-BLRNET model are shown in Table 5.
[0295] Table 5 Parameters of CNN-BLRNET
[0296]
[0297] To verify the CNN-BLRNET model, the CNN-RESNET-BILSTM model is compared with it. The CNN-RESNET-BILSTM model is designed by connecting CNN, traditional RESNET and BILSTM in series. In addition, BILSTM is placed behind traditional RESNET, the CNN layer first extracts some local features, RESNET further extracts local features, and finally BILSTM extracts temporal features.
[0298] In the CNN-RESNET-BILSTM model, the extraction of temporal features is separated from the extraction of local features. Considering that RESNET and BILSTM are linearly connected, the information transmission of the network model may be biased, and the model operation time of CNN-RESNET-BILSTM is longer than that of CNN-BLRNET. For CNN-RESNET-BILSTM, temporal and spatial feature extraction is performed separately, and the interrelated features between the two features cannot be fully extracted, which reduces the prediction accuracy to a certain extent. Therefore, this example embeds BILSTM into RESNET to construct the BLRNET model, extracts temporal and spatial features synchronously, and adds the maximum pooling layer after the CNN layer to extract the maximum features of local data, thereby improving the accuracy and robustness of the model. Compared with the fitting results of CNN-BLRNET, CNN-RESNET-BILSTM and CNN-BLRNET without the MaxPooling layer, as shown in Fig.18 (a). To illustrate the generalization performance, the prediction results are shown in Fig.18 (b) as shown.
[0299] The evaluation indicators of CNN-BLRNET and CNN-RESNET-BILSTM without MaxPooling layer were calculated, as shown in Table 6. The fitting abilities of CNN-BLRNET with MaxPooling layer and CNN-BLRNA, CNN-RESNET-BILST without MaxPooling layer are 95.63%, 95.17% and 95.16% respectively. Table 6 shows that CNN-BLRNET outperforms CNN-RESNET-BILSTM in terms of prediction performance. BILSTM embedded in RESNET is better than the serial linear connection of BILSTM and RESNET. The fitting performance of CNN-BLRNET is better than that of CNN-BLRNET without Maxpooling layer. The results show that it is effective to use Maxpooling layer to extract local features to the maximum extent. In terms of prediction performance, CNN-BLRNET has the best prediction performance, indicating that the CNN-BLRNET model proposed in this embodiment has good robustness and generalization performance.
[0300] Table 6 Model evaluation indicators
[0301]
[0302] In order to further test the prediction performance of CNN-BLRNET, the whole process is divided into fitting and prediction processes. In the early stage of time series prediction, gradient vanishing is easy to occur and a local optimal solution is obtained. Then the problem of inaccurate long-term prediction arises. Through multiple iterations, the network model is continuously optimized to build a local optimal network model. Then the prediction is carried out to verify the generalization performance and robustness of CNN-BLRNET. Finally, the control variable method is used to construct an error model based on CNN-RESNET-BILSTM, CNN-RESNET, CNN-BILSTM, BILSTM, GRU and BP. To ensure reliability, 10 predictions were performed under the same conditions, and the average value of the 10 prediction experiments was taken as the fitting result. The comparison of the fitting results is shown in the figure. Fig.19 As shown. Further verification of the extraction of spatiotemporal features is crucial for accurate model prediction. Fig. 20 The prediction results were compared in , and in order to make the results more convincing, the error data under condition #2 was used and a similar trend was found.
[0303] The evaluation indicators of the pre-trained CNN-BLRNET model are calculated, as shown in Table 7. The running times of CNN-BLRNET, CNN-RESNET-BILSTM, CNN-BILSTM, CNN-RESNET, BILSTM, and BP are 23.603s, 35.654s, 10.526s, 9.446s, 10.372s, 10.920s, 9.652s, and 7.366s, respectively. The RMSE, MSE, and MAE of CNN-BLRNET are the smallest among the above models, indicating that the CNN-BLRNET model has the highest prediction accuracy. The RMSE, MSE, and MAE of CNN-BLRNET and CNN-RESNET-BILSTM models are smaller than those of CNN-BILSTM and CNN-RESNET models. The RMSE, MSE, and MAE of CNN-BILSTM and CNN-RESNET are smaller than those of BILSTM and LSTM with a single time feature. The RMSE, MSE, and MAE of BP network are all larger than those of other models, and its prediction accuracy is lower than that of other models. Finally, the fitting abilities of CNN-BLRNET, CNN-RESNET-BILSTM, CNN-BILSTM, CNN-RESNET, BILSTM, and BP are 95.63%, 95.17%, 93.38%, 93.07%, 91.74%, 91.39%, 90.85%, and 89.22%, respectively.
[0304] The fitting accuracy and efficiency of CNN-BLRNET are better than those of CNN-RESNET-BILSTM. In addition, the fitting accuracy of CNN-BLRNET and CNN-RESNET-BILSTM is significantly better than that of CNN-BILSTM and CNN-RESNET. CNN-BLNET and CNN-RESNET-BILSTM realize the extraction of spatiotemporal features, while CNN-BILSTM only extracts temporal features and CNN-RESNET only extracts spatial features. Then the necessity of extracting spatiotemporal features of error data is demonstrated. More importantly, it is an effective and feasible solution to extract spatiotemporal features using the proposed CNN-BLRNET model. Compared with CNN-BILSTM, BILSTM and LSTM lack CNN. Compared with CNN-RESNET, BILSTM and LSTM lack convolutional layers. Convolution operations can extract local features, which is beneficial to improve model accuracy. Similarly, the prediction accuracy and efficiency of CNN-BLRNET are better than CNN-RESNET-BILSTM, CNN-RESNET, CNN-BILSTM, BILSTM and GRU and BP models. Embedding BILSTM into RESNET can indeed identify more connected features and improve model accuracy. In addition, placing BILSTM to extract temporal features behind CNN to extract spatial features can also improve model accuracy. In addition, spatiotemporal feature extraction is indeed more effective than single feature extraction.
[0305] Table 7 Pre-training model evaluation indicators
[0306]
[0307] Finally, we get the loss function, such as Fig.21 As shown in the figure. In a total of 50 iterations, CNN-BLRNET, CNN-RESNET-BILSTM, CNN-RESNET, and CNN-BILSTM converge to stability. The stability of BILSTM and GRU is slightly lower than that of CNN-BLRNET, CNN-RESNET-BILSTM, CNN-RESNET, and CNN-BILSTM. Among the above models, LSTM and BP have the worst stability. More importantly, the convergence of CNN-BLRNET and CNN-RESNET-BILSTM is significantly stronger than that of other models.
[0308] 2.1.2 Transfer Learning Model
[0309] The TLM of CNN-BLRNET is established by retraining the parameters of the dense layer. TLM is a transfer learning model, and its main layer structure and parameters are shown in Table 8.
[0310] Table 8 TLM parameter settings of CNN-BLRNET
[0311]
[0312] Error prediction is performed with TLM of CNN-BLRNET. Error data of CHD25A spindle is used as input. Fig. 22 (a) shows the comparison between the TLM of CNN-BLRNET and the pre-trained CNN-BLRNET model as shown. Fig. 22 (b) shows the prediction results obtained by TLM of CNN-RESNET-BILSTM and pre-trained CNN-RESNET-BILSTM model. TLM of BILSTM is compared with the pre-trained BILSTM model, as shown in Fig. 22 (c) as shown. Fig. 22 (d) shows the prediction results obtained by the TLM of LSTM and the pre-trained LSTM model. The TLM of CNN-BLRNET, CNN-RESNET-BILSTM, BILSTM, and LSTM are compared with the pre-trained CNN-BLRNET, CNN-RESNET-BBILSTM, BILSDM, and LSTM models. The average of multiple prediction results is taken. The prediction accuracy of the TLM of CNN-BLRNET is better than that of the pre-trained CNN-BLRNET model. The prediction accuracy of the TLM of CNN-RESNET-BILSTM is much higher than that of the pre-trained CNN-RESNET-BILSTM.
[0313] Table 9 shows that the TLM calculation time of CNN-BLRNET is 18.713s, and its fitting ability is 96.66%. The calculation time of CNN-BLRNET is 23.603s, and its fitting ability is 95.63%. The TLM calculation time of CNN-BLRNET is 18.713s, and its fitting ability is 96.66%. The calculation time of CNN-BLRNET is 23.603s, and its fitting ability is 95.63%. The results show that TLM has better prediction accuracy and efficiency compared with the pre-trained model.
[0314] Table 9 TLM evaluation indexes of CHD25A machine tool spindle
[0315]
[0316] In order to make the conclusion more convincing, the error data of the TGK46100 spindle is also used. The thermal error of the TGK46100 spindle is measured. The robustness and generalization ability are verified using two sets of thermal error data of the TGK46100 spindle. TLMs of CNN-BLRNET, CNN-RESNET-BILSTM, BILSTM and LSTM are established. The thermal error of the TGK46100 spindle is used as input. Fig.23 The comparison between TLM and pre-trained models is shown. The prediction accuracy of TLM of CNN-BLRNET is better than that of CNN-BLRNET. The performance of TLM of CNN-BLRNET is also better than that of other TLMs.
[0317] The evaluation indicators are shown in Table 10. The results show that the prediction accuracy and efficiency of TLM of CNN-BLRNET are 19.055s and 95.50% respectively. The prediction accuracy and efficiency of CNN-BLRNET are 20.222s and 94.83% respectively. Compared with the pre-trained CNN-BLRNET model, the TLM of CNN-BLRNET has better prediction performance. The prediction accuracy and efficiency of TLM of CNN-RESNET-BILSTM, BILSTM and LSTM are also better than those of pre-trained CNN-RESNET-BILSTM, BILSDM and LSTM models respectively. The results show that TLM has good prediction performance. At the same time, the prediction accuracy and efficiency of CNN-BLRNET are better than those of CNN-RESNET-BILSTM, BILSTM and LSTM models. This further proves the superiority of the proposed BLRNET. In addition, the TLM of CNN-BLRNET can adapt well to the error prediction of other working conditions. Then the robustness of TLM of CNN-BLRNET is effectively improved by using transfer learning theory.
[0318] Table 10 TLM evaluation indexes of TGK46100 machine tool spindle
[0319]
[0320] 2.1.3 Error Control
[0321] This embodiment performs error control and then conducts actual processing experiments. Fig.24 (a) The feed rate is 500 mm / min, the spindle speed is 3500 r / min, and the tool used is a carbide milling cutter. Fig.24(b) shows the machined workpiece. The three-dimensional coordinate measuring machine is used to measure the geometric errors of the machined workpiece, and the measurement results are shown in Table 11. With the implementation of the framework, the measured contour error is reduced from 15.8μm to 9.2μm. The contour error is reduced by 41.2%. By implementing error control, the roundness error is reduced from 26.5μm to 8.6μm. The roundness error is reduced by 67.5%. By implementing error control, the verticality error is reduced from 38.7μm / m to -9.6μm / m. The effectiveness of the designed system is then verified.
[0322] Table 11 Comparison of results
[0323]
[0324] 2.2. Online monitoring and fault diagnosis services
[0325] 2.2.1 Feature Extraction
[0326] This embodiment collects vibration signals of different damage degrees of the automatic tool changer in one working cycle, and the signal sampling frequency is 6400 Hz. Fig.25 The time domain signal components of the vibration signal are complex, including impact signals and harmonic signals. It is difficult to observe a regular change trend from the impact components.
[0327] The collected vibration signals of the automatic tool changer with different damage degrees in one working cycle are analyzed by spectrum analysis to obtain the spectrum diagram, such as Fig.26 It is impossible to understand the abnormal information of each component of the automatic tool changer from the frequency domain.
[0328] The internal structure of the automatic tool changer is complex, and the measured signals include the vibration signals of the arc cam, helical gear and rolling bearing. The vibration signal of the arc cam is mainly the impact component, while the vibration signal of the helical gear is mainly the harmonic component. The morphological components of the shock wave and the harmonic component are different. Therefore, the multi-component sparse separation method is used to separate the harmonic component and the impact component in the vibration signal. Finally, the vibration signal is decomposed.
[0329] The redundant dictionary is constructed by Hermite spline multiwavelet to match the harmonic and impulse components. H and λ P The values of are 0.3 and 0.25 respectively. The penalty parameter μ of the split enhanced Lagrangian shrinkage method is 10. Then the impact component is obtained, as Fig. 27 As shown in the figure. The vibration signals in different states are separated by multi-component sparse separation method, and then the impact components in the original signal are extracted. The impact components obtained in different states are similar and consist of a series of impacts. The multi-component sparse separation method is used to effectively filter out the interference of harmonic components and noise in the original vibration signal.
[0330] 2.2.2 Processing state identification and fault diagnosis
[0331] The extracted impact component is Fig.28 As shown. Harmonic components and high-frequency noise are filtered out. However, it is still difficult to determine the possible fault location of the automatic tool changer. Due to the structure and working characteristics of the gear, shocks will be generated during the working processes of starting, holding the tool, pulling the tool, changing the tool, inserting the tool, and resetting. The shapes of these shocks are similar to the shapes generated by the fault. Therefore, the separation and sparse method cannot perform fault diagnosis on the automatic tool changer. According to the structure and working principle of the gear, it is found that the cam working cycle diagram corresponds to the shocks generated when the gear works normally. According to the designed working cycle diagram, the combined cam works regularly in a certain order. Impact component #1 corresponds to the starting action, and impact components #2 and #3 correspond to the grabbing action. Impact components #4 and #5 correspond to the stretching action, and impact components #6 and #7 correspond to the cutting action. Shock absorber components #8 and #9 correspond to the insertion action, and shock absorber component #10 corresponds to the reset action. These shocks also reflect the working process of the automatic tool changer. Therefore, fault diagnosis can be performed with the help of the motion cycle diagram and the extracted shock components of the vibration signal.
[0332] Fig.29 Shows Fig. 27 The impact component under the early and middle fault conditions is analyzed in combination with the working cycle diagram. Fig.29 and Fig.28 , it was found that there was a series of large-amplitude shocks between shocks #6 and #7, and the existence of these shocks was inconsistent with the working cycle process of the automatic tool changer. In addition, the shocks at other working times are also shown in the figure. 29 and 28 are similar. Therefore, it can be preliminarily judged that the working profile of the 180° indexing tool change arc cam is damaged, which is consistent with the actual damage setting position in this study. The shock components obtained under other conditions were analyzed and similar analysis results were obtained.
[0333] 3. Conclusion
[0334] An intelligent integration framework for high-precision machining is designed. Self-perception is achieved by configuring a ZigBee wireless sensor network and an intelligent gateway. The model training process is divided into preliminary training and fine training. The preliminary training is performed in the cloud center with historical data as input, and the fine training is performed on the edge node. Finally, data services such as online monitoring and fault diagnosis services and thermal error control services are provided. The main conclusions are as follows:
[0335] (1) An intelligent integration framework for high-precision machining is designed. Data perception and transmission are carried out through ZigBee wireless sensor networks and the Industrial Internet. In addition, an intelligent gateway is designed for edge computing. Data management and storage are carried out with the help of Hadoop. Then Spark is used for data calculation. For fault diagnosis, a multi-component sparse separation method is proposed. For error prediction, a CNN-BLRNET error model is proposed. Finally, online monitoring services, fault diagnosis services, and thermal error control services are provided. With the implementation of the intelligent integration framework, the measured contour error is reduced by 41.2%, and the roundness error of the machined workpiece is reduced by 67.5%. In addition, accurate diagnosis of tool change mechanism faults is achieved.
[0336] (2) On the one hand, the heat conduction equation is solved, and the error mechanism modeling can effectively reflect the memory and nonlinear characteristics, proving that the thermal error data has dual characteristics of time and space. On the other hand, BILSTM is embedded in the new RESNET, and BLRNET is proposed to realize the extraction of time and space features. BLRNET can extract local features with CNN layer and time features with BILSTM. In addition, MaxPooling layer is added to extract important local feature information. The proposed CNN-BLRNET is much superior to the serial network structure.
[0337] (3) The fitting abilities of CNN-BLRNET, CNN-RESNET-BILSTM, CNN-RESNET, CNN-BILSTM, BILSTM, and BP are 95.63%, 95.17%, 93.38%, 93.07%, 91.74%, 91.39%, 90.85%, and 89.22%, respectively. The results show that CNN-BLRNET has the best prediction performance. Among the above models, CNN-BLNET has the highest prediction accuracy and the best prediction performance. Practice has proved that CNN-BLRNET has good robustness and generalization ability. The TLM of CNN-BLRNET has good prediction performance. The TLM structure of CNN-BLRNET is the same as that of the pre-trained CNN-BLRNET. The dense layer is retrained and all other layers of the pre-trained CNN-BLRNET model are frozen. Therefore, the prediction performance and efficiency of the TLM of CNN-BLRNET are better than those of the pre-trained CNN-BLRNET model.
[0338] (4) A multi-component sparse component method based on Q-switched wavelet and multi-wavelet transform is proposed. On the basis of morphological component analysis, a redundant dictionary construction and sparse representation method based on Q-switched wavelet and multi-wavelet transform is proposed, and the separation method of different components of vibration signals is studied. The establishment of the objective function of the sparse separation method and the optimization solution method of the objective function based on the split augmented Lagrangian shrinkage algorithm are studied. Finally, the impact component in the original signal is extracted, and the motion period diagram of the automatic tool changer is used to locate the damage of the camber cam.
[0339] The above-described embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or changes made by those skilled in the art based on the present invention are within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.
Claims
1. A method for predicting thermal error using the CNN-BLRNET thermal error prediction model, Features: include: (1) Input the thermal error data into the input layer of the CNN-BLRNET thermal error prediction model; (2) Use the CNN layer to process the data to extract important local features of the data; use the maximum pooling layer to further extract the local maximum features and obtain the initial feature information of the data; (3) Extract the temporal characteristics of data using the BLRNET model; (4) The data enters the dropout layer, flatten layer, and dense layer in turn, converting the two-dimensional vector into a one-dimensional vector for output to obtain the thermal error prediction value; The CNN-BLRNET thermal error prediction model includes sequentially setting a first input layer, a CNN layer, a BLRNET model, a dropout layer, a flatten layer, a dense layer and a first output layer; The BLRNET model is formed by embedding a BILSTM network into a RESNET network, and the BLRNET model includes a second input layer, a ReLU layer, and a second output layer, at least two RESNET units connected in series are arranged between the ReLU layer and the second output layer, and the RESNET unit includes a direct mapping part and a jump connection part, the direct mapping part includes a convolution layer I, a maximum pooling layer I, and a BILSTM layer arranged in sequence, the jump connection part includes a convolution layer II and a maximum pooling layer II arranged in sequence, and an addition operator is arranged between the BILSTM layer and the maximum pooling layer II.
2. The method for predicting thermal error using the CNN-BLRNET thermal error prediction model according to claim 1, Features: The BILSTM network includes an input layer, a forward LSTM layer, a backward LSTM layer and an output layer. The forward LSTM layer and the backward LSTM layer include several LSTM units. The principle of the LSTM unit is: f t =σ(W f [h t-1 ,x t ]+b f ) i t =σ(W i [h t-1 ,x t ]+b i ) C t =tanh(W c [h t-1 ,x t ]+b c ) The t =σ(W O [h t-1 ,x t ]+b O ) h t =O t *tanh(C t ) Where t and t-1 are the current and previous moments respectively; f t 、i t and O t They are forget gate, input gate and output gate respectively; x t , C t and h t are the input, unit state and output at the current moment; x t-1 , C t-1 and h t-1 are the input, unit state and output at the previous moment respectively; is the candidate cell state; W f , W i , W c and W o is the weight matrix; b f 、b i 、b c and b o is the bias term; * denotes element-wise multiplication; σ(·) and tanh(·) are activation functions.
3. The method for predicting thermal error using the CNN-BLRNET thermal error prediction model according to claim 2, Features: The principle of BILSTM is: h t =f(w 1 *x t +w 2 *h t-1 ) h′ t =f(w 3 *x t +w 4 *h′ t-1 ) o t =g(w 5 *h t +w 6 *h t ) Where f(·) represents the forward operation function; g(·) represents the backward operation function; h t represents the forward output; h′ t Indicates backward output; w 1 、w 2 、w 3 、w 4 、w 5 and w 6 represents the weight matrix.
4. The method for predicting thermal error using the CNN-BLRNET thermal error prediction model according to claim 3, Features: The principle of the RESNET unit is that the data is converted into nonlinearity through the ReLU layer: Where X represents the input data; relu(·) represents the ReLU activation function, x i Represents the i-th input data; Data as X 1 Input direct mapping part, data is also as X 2 Input the remaining part, for convolutional layer I: Among them, X 1 Represents the input of the direct mapping part; W 1 represents the convolution weight matrix; Represents the input value; represents the convolution kernel value; b 1 Represents the bias parameter in the convolution operation; Indicates the actual convolution operation bias; f 1 (·) represents the actual convolution operation; conv1d(·) represents the convolution operation; For the max pooling layer I: Among them, X′ 1 represents the output of the maximum pooling layer I; maxpooling(·) represents the pooling operation; Represents the input value of the pooling operation; The BILSTM layer extracts temporal features: h t =f(w 1 *X′ 1 +w 2 *h t-1 ) h′ t =f(w 3 *X′ 1 +w 4 *h′ t-1 ) The output of the direct mapping section is: F(X,W)=h t +h′ t Among them, F(X,W) represents the output of the direct mapping part; For convolutional layer II: Among them, X 2 Indicates the input of the jump connection part; W 2 represents the convolution weight matrix; Represents the input value; represents the convolution kernel value; b 2 Represents the bias parameter in the convolution operation; Indicates the actual convolution operation bias; f 2 (·) represents the actual convolution operation; conv1d(·) represents the convolution operation; The output of the jump connection part is: Among them, G(X) represents the output of the jump connection part; x i ' l Represents the input value of the pooling operation; The output of the BLRNET unit is: H(X)=F(X,W)+G(X) Where H(X) represents the output of the BLRNET unit.
5. A transfer learning method for predicting thermal error using a CNN-BLRNET thermal error prediction model as claimed in any one of claims 1 to 4, Features: The parameters of the first input layer, CNN layer, BLRNET model dropout layer, flatten layer and the first output layer in the original CNN-BLRNET thermal error prediction model are kept unchanged, and the dense layer is retrained and updated to obtain the transfer learning model of the CNN-BLRNET thermal error prediction model.
6. An intelligent integrated framework system, Features: Including data acquisition system and data calculation system; The data acquisition system collects machine tool working status data and workpiece quality data; all collected data are transmitted to the edge node and the cloud center through the local area network and the industrial Internet respectively; The data computing system performs data computing on the collected data, performs self-learning with the support of machine learning, reinforcement learning, deep learning and transfer learning, establishes the fault diagnosis model, thermal error model, three-dimensional model, control strategy and knowledge of the machine tool as a digital twin model, and provides data services including condition monitoring, fault diagnosis and thermal error control; The cloud center is embedded with the CNN-BLRNET thermal error prediction model in the method for predicting thermal error using the CNN-BLRNET thermal error prediction model as described in any one of claims 1 to 4, and the cloud center performs preliminary training on the CNN-BLRNET thermal error prediction model based on historical thermal error data; the CNN-BLRNET thermal error prediction model that has completed preliminary training is deployed to the edge node, and the edge node performs local training on the CNN-BLRNET thermal error prediction model using real-time thermal error data, so that the CNN-BLRNET thermal error prediction model can adapt to the operating conditions of each edge node.
7. The intelligent integrated framework system according to claim 6, Features: The edge node is provided with an intelligent gateway node, which includes a processor module, a ZigBee communication module, a WIFI communication module and a power management module. The ZigBee communication module is used to communicate with each data acquisition node of the data acquisition system, and the WIFI communication module is used to exchange information with the cloud center.
8. The intelligent integrated framework system according to claim 6, Features: In condition monitoring and fault diagnosis, effective separation and feature extraction of harmonic and impulse components are required. The method is as follows: Assume that the signal s is linearly composed of K components with different shapes, that is, The basic model of morphological component analysis is based on the following two assumptions: (1) For a given signal s k For each component of Usually L k >>N, then the problem to be solved is: Among them, α k represents the sparse representation coefficient; represents a sparse solution; (2) For a given signal s l Each component of , when k≠l: Among them, α l represents the non-sparse representation coefficient; represents a non-sparse solution; The above two assumptions ensure that the dictionary T k It is possible to distinguish signals of different morphological components, and the sparse decomposition of the signal s is obtained by solving the following problem This problem is a non-convex NP-hard problem, which is transformed into the next convex optimization problem: The signal to be analyzed s is the harmonic component s A and the impact component s B The linear combination of , the model satisfies the following three conditions: (1) For the harmonic component s A , constructing an overcomplete dictionary through harmonic wavelet transform in, α represents the sparse representation coefficient; represents a sparse solution; Get the sparse solution and get the sparse representation coefficient α A ; (2) For the shock wave component s B , we get a non-sparse solution, which means that the dictionary T A Two separate components can be distinguished: (3) For the shock wave component s B , an overcomplete dictionary constructed by multiwavelet transform Can sparsely represent the shock wave component s B and harmonic components A , to solve the problem of multi-component sparse separation of signals: Convert it into a convex optimization problem: and The sparse matrix of L is the norm of ||α A || 1 +||α B || 1 When s = T A α A +T B α B The optimal solution under the constraints; if the signal is noisy, the signal cannot be simply decomposed into harmonic components s A and the impact component s B , needs to be expanded to: The parameter ε characterizes the noise level in the signal s. In order to obtain the optimal solution, the Lagrange multiplier method is used to give a suitable Lagrange operator, and the constrained optimization problem is transformed into an unconstrained optimization problem: where λ = [λ A ,λ B ] T represents the Lagrangian operator, λ is a function of s, α and ε; the objective function of morphological component analysis is: Where J(·) represents the objective function of morphological component analysis.
9. The intelligent integrated framework system according to claim 6, Features: Thermal error control requires first predicting the thermal error. The method for predicting the thermal error using the CNN-BLRNET thermal error prediction model is as follows: (1) Input the thermal error data into the input layer of the CNN-BLRNET thermal error prediction model; (2) Use the CNN layer to process the data to extract important local features of the data; use the maximum pooling layer to further extract the local maximum features and obtain the initial feature information of the data; (3) Extract the temporal characteristics of data using the BLRNET model; (4) The data enters the dropout layer, flatten layer, and dense layer in turn, converting the two-dimensional vector into a one-dimensional vector for output to obtain the thermal error prediction value.
Citation Information
Cited By
Power transmission line real-time regulation and control method and system based on CNN (Convolutional Neural Network) and Transform cooperation
CN120708010A
Real-time transmission line regulation method and system based on cooperation of CNN and transformer
CN120708010B
