CN111890127A - Cutting state edge intelligent monitoring method based on online incremental wear evolution model - Google Patents

Abstract

The technology discloses a cutting state edge intelligent monitoring method based on an online incremental wear evolution model, which utilizes a tool edge intelligent monitoring system to acquire cutting force signals, vibration signals and acoustic emission signals in the machining process of machining equipment in real time; establishing a wireless data transmission network, controlling a data transmission path by using a domain controller, and finally transmitting data from a sensing layer to an edge computing layer; preprocessing data by using a Field Programmable Gate Array (FPGA) hardware system of an edge calculation layer to obtain a training sample, determining an internal topological structure of the online sequential increment extreme learning machine by combining the feature number and the sample number of the training sample, and finally transmitting the data acquired by a sensor to the trained edge calculation layer under the same working condition; sending a real-time monitoring result to a cloud computing layer by utilizing an SDN technology layer in the tool edge intelligent monitoring system; and applying the tool wear monitoring result by using the application service in the cloud computing layer. The technology has high real-time performance, good accuracy and strong data processing capability.

Description

Cutting state edge intelligent monitoring method based on online incremental wear evolution model

Technical Field

The technology relates to the field of cutter abrasion monitoring of machining equipment, in particular to an intelligent cutting state edge monitoring method based on an online incremental abrasion evolution model.

Background

In the field of machining, the wear amount of a tool has an important meaning, and the wear amount directly affects the residual life of the tool and the size, surface roughness and the like of a machined workpiece, but the tool inevitably has a wear phenomenon during the cutting process. Until now, the method for monitoring the wear of the cutter is mainly divided into a direct method and an indirect method, wherein the direct method refers to a contact method and an optical image method, but due to the limitation of test conditions, the direct method needs to shut down processing equipment in different time stages, and the needed infrastructure is complex and is not easy to operate; the indirect method is to judge the wear state of the cutter by monitoring a cutting force signal, a cutter vibration signal, an acoustic emission signal, a noise signal and the like, preprocessing the signals and then extracting characteristic information in the cutting process of the cutter, wherein the accuracy of a monitoring result mainly depends on the intelligent learning model construction level and experience of a user.

Whether the intelligent learning tool selects reasonably has great influence on the monitoring efficiency and the diagnosis service level of the tool abrasion loss. The intelligent learning tools for monitoring the tool wear amount in the prior art all adopt a batch learning method and an off-line training mode, so that the requirements of on-line monitoring are met, previous learning achievements need to be abandoned, and the learning and training need to be repeated according to newly added data and all past data. Relearning all of the data results in a significant amount of time and space resources being consumed. As the scale of online monitoring data increases, the demand for time and space increases rapidly, and eventually, the learning speed cannot catch up with the data updating speed. In addition, the application of the intelligent learning tool for monitoring the tool wear amount in the prior art also assumes that the tool wear state during the cutting process is monitored in a mode known in advance. However, in practical applications, the training samples are usually not available all at once, various tool wear patterns are often obtained gradually over time, and the information reflected by the samples may change over time. Therefore, in order to enable the intelligent learning tool for monitoring the cutter abrasion loss to have strong learning capacity, a sufficient number of cutter abrasion patterns can be learned and can be communicated; the tool wear model learning system also has continuous self-improvement and updating capability, and can continuously relearn a new tool wear model in the cutting process, and gradually improve the knowledge and understanding of the tool wear state in the cutting process; meanwhile, the system also has the capability of detecting and analyzing the wear state of the cutter in the cutting process fast enough; the technology is based on the idea of edge intelligence, adopts an online incremental learning strategy, takes a single-layer feedforward neural network as a base, establishes a rapid online sequential neural network model, and uses the model as an intelligent learning tool to monitor the cutter abrasion in the cutting process in an online and real-time manner.

The invention discloses a method for monitoring the wear of a cutter of machining equipment (CN102091972B), which is used for judging the wear degree of the cutter by detecting the three-phase output current of a driving motor in cutting machining and utilizing a wavelet decomposition and polynomial fitting method. However, the method utilizes wavelet decomposition to extract features, the extracted result has high empirical components, the application effect mainly depends on the calculation level of application personnel, and when the accuracy requirement of polynomial fitting is high, a large amount of time is required to complete the fitting, so that the real-time performance and the accuracy are insufficient.

The invention discloses a tool wear online monitoring method based on wavelet packet analysis and RBF neural network (CN108356606A), which provides a method for monitoring signals by using the RBF neural network by detecting cutting force signals in the cutting process and the flank wear of a machined tool and analyzing and processing the signals through the wavelet packet. However, in the method, the implicit characteristics are extracted by utilizing wavelet packet analysis, the implicit relation between the cutting force signal and the cutter abrasion loss cannot be well reflected by the extracted result, and the cutter abrasion monitoring is carried out by utilizing the RBF neural network, so that the learning speed is low, the generalization capability is poor, the real-time performance of the monitoring result is insufficient, and the accuracy is low.

The invention patent "Tool monitor" (EP0334341a2) is characterized in that at least one acoustic emission sensor is installed near the processing equipment and is not in contact with the processing equipment, and the acoustic emission sensor is used for collecting the acoustic emission signal of the Tool of the processing equipment, the position of another acoustic emission sensor is set to detect the acoustic emission signal from the Tool of the processing equipment when the Tool is broken, the signal is preprocessed by combining a band-pass filter, and the state of the Tool is obtained by comparing the acoustic emission signal of the Tool of the processing equipment collected in real time with the acoustic emission signal when the Tool is broken. However, the band-pass filter cannot completely eliminate the influence of the interference factor, and the band-pass filter cannot acquire the implicit characteristics of the acoustic emission signal and the cutter state, so that the real-time state of the cutter cannot be completely reflected by the acoustic emission signal after filtering, and the real-time state of the cutter cannot be judged due to the fact that simple comparison judgment needs offline judgment, and the real-time performance of a monitoring result is insufficient.

The invention patent "Cutting tool wear monitor" (EP0165745a2) monitors wear of a rotating Cutting tool based on short circuit current, open circuit voltage and power generated on a workpiece during Cutting, the generated current, voltage and power increasing gradually as the tool wears until a sharp increase indicates damage to the tool due to excessive wear or breakage. However, the current and the voltage are easily affected by external factors, so that the monitoring result cannot accurately represent the real-time wear state of the cutter, and only whether the cutter is worn or not can be judged through the change of the current and the voltage, but the real-time wear amount of the cutter cannot be judged.

At present, the prior art of the indirect method adopts a batch learning method and an offline training mode, in order to meet the requirement of online monitoring, previous learning results need to be abandoned, and in addition, the newly added data and all the past data need to be relearned and trained together, so that all the data are relearned, and a large amount of time and space resources are consumed; with the continuous increase of the data scale, the requirements on time and space also increase rapidly, and finally, the learning speed cannot catch up with the data updating speed; in practical application, a training sample cannot be obtained all at once, various sensing data representing the wear state of the tool are obtained gradually along with time, and information reflected by the sample can be changed along with time; these problems severely impact the on-line requirements of the tool wear monitoring system during machining.

In addition, in the prior art of the indirect method, data processing in the processing process is completed in a server or a data center, and due to the fact that the amount of data generated in the processing process is large, the transmitted data is slow, the calculation speed of the data center is affected, the data processing time is prolonged, and the real-time requirement of a tool wear monitoring system in the processing process cannot be well met. Aiming at the problem, the technology plans to design an intelligent cutting state edge monitoring method based on an online incremental wear evolution model, and the system utilizes an online sequential incremental extreme learning machine model and an edge calculation method, so that the problems of low real-time performance and insufficient accuracy of the cutter wear monitoring system in the current machining process are effectively solved.

Disclosure of Invention

In order to solve the technical problems, the technology aims to provide an intelligent cutting state edge monitoring method based on an online incremental wear evolution model, which is high in real-time performance, good in accuracy and strong in data processing capacity.

In order to solve the technical problems, the technology adopts the following technical scheme:

the technology relates to an intelligent monitoring method for cutter abrasion based on edge calculation, which comprises the following steps:

s1: establishing an intelligent monitoring system for the edge of the cutter;

s2: the method comprises the steps that a three-way cutting force sensor of a sensing layer in an intelligent cutter edge monitoring system is used for collecting cutting force signals of an X axis, a Y axis and a Z axis in the machining process of machining equipment in real time, and meanwhile, a three-way vibration sensor and an acoustic emission sensor are used for respectively collecting vibration signals and acoustic emission signals in the machining process of the machining equipment;

s3: a ZigBee module, a Wi-Fi module and a Bluetooth module in the edge gateway are utilized to form a wireless data transmission network, a domain controller is utilized to control a data transmission path, and finally, data are transmitted from a sensing layer to an edge computing layer;

s4: preprocessing data by using an FPGA hardware system of an edge calculation layer to obtain a training sample, determining an internal topological structure of the on-line sequential increment extreme learning machine by combining the feature number and the sample number of the training sample, and finally transmitting the data acquired by the sensor to the edge calculation layer after training under the same working condition;

s5: sending a real-time monitoring result to a cloud computing layer by utilizing an SDN technology layer in the tool edge intelligent monitoring system;

s6: and applying the tool wear monitoring result by using the application service in the cloud computing layer.

Further, in S1, the intelligent tool edge monitoring system includes:

the sensing layer comprises a three-way cutting force sensor, a three-way vibration sensor and an acoustic emission sensor, the three-way cutting force sensor is used for collecting cutting force signals of an X axis, a Y axis and a Z axis in the machining process of the machining equipment in real time, the three-way vibration sensor is used for collecting vibration signals in the machining process of the machining equipment, and the acoustic emission sensor is used for collecting acoustic emission signals in the machining process of the machining equipment;

the data transmission layer comprises a plurality of edge gateways and a domain controller, wherein the edge gateways comprise a ZigBee module, a Wi-Fi module and a Bluetooth module, form a wireless data transmission network by utilizing the ZigBee module, the Wi-Fi module and the Bluetooth module, and control a data transmission path by utilizing the domain controller;

the edge calculation layer comprises a programmable logic gate array hardware system, the hardware function module comprises a self-encoder feature extraction module and an online sequential increment extreme learning machine module, the self-encoder feature extraction module is used for preprocessing data to obtain a training sample, the internal topological structure of the online sequential increment extreme learning machine module is determined by combining the feature number and the sample number of the training sample, and finally the data acquired by the sensor is transmitted to the edge calculation layer after training under the same working condition;

the software defined network technology layer comprises a plurality of SND controllers, the SND controllers are used for adjusting a transmission path between the edge computing layer and the cloud computing layer in real time, and real-time monitoring results are sent to the cloud computing layer in an optimal path;

and the cloud computing layer is used for finishing various specific applications including real-time early warning, visual display and decision assistance to users and industries.

In S3, the domain controller is used to control the transmission path of the data, and the edge transmission and internal transmission methods are used to optimize the transmission path, so as to adjust the data transmission path between the sensing layer and the edge calculation layer.

Further, in S4, a self-encoder feature extraction module in the FPGA hardware system is used to perform feature extraction on the information acquired by the sensor in real time, and an online sequential incremental limit learning module in the FPGA hardware system is used to perform online sequential incremental learning on the information after feature extraction.

Further, in S5, the SND controller communicates with the OpenFlow-compatible switch using an OpenFlow protocol running on the transport layer security protocol, and forwards the received tool wear amount data to the cloud computing layer along a specified data forwarding path using rules defined by the SND application.

Further, in S6, the application service includes a visual display function, and the real-time tool wear amount is displayed as a waveform graph on the liquid crystal display screen by using the visual display function; the real-time early warning function is utilized to apply the cutter abrasion loss to a warning trigger for setting fixed parameters in advance, and when the abrasion loss is not in a normal range, warning triggering is carried out; and the auxiliary decision function is utilized, and the auxiliary adjustment is carried out on the machining parameters of the machining equipment by combining the tool abrasion loss, so that the adjustment of the machining parameters of the machining equipment is more reasonable.

Still further, the implementation method of the sensing layer in S2 is as follows: the perception layer comprises sensors of cutting force, vibration, acoustic emission, noise and the like, and a three-way vibration sensor, a three-way cutting force sensor and an acoustic emission sensor are used for respectively acquiring a vibration signal, a cutting force, an acoustic emission signal and the like in the processing process of the processing equipment in the processing process.

Still further, the sensor in S2 is connected to RAM through ZigBee₁When the data is received again, the data is shifted to the left by 8 bits and added with the data stored last timeAnd obtaining 16 bits of data representing original information of tool wear, wherein the first bit represents a sign bit, 2 bits to 6 bits represent integer bits, and the last 10 bits represent decimal bits.

Still further, the implementation method of the data transmission layer in S3 is as follows:

a ZigBee module: the ZigBee module and the domain controller form a star network topological structure, so that data conversion between ZigBee and the domain controller can be realized, and the domain controller can control a data transmission path of the ZigBee module in real time;

a Wi-Fi module: establishing a Wi-Fi wireless transmission network through a router, forming various transmission paths between a sensing layer and an edge computing layer by using different wireless transmission protocols, and adjusting the Wi-Fi transmission paths by controlling the transmission protocols through a domain controller;

a Bluetooth module: the transmission link between the sensing layer and the edge computing layer is established by utilizing a plurality of Bluetooth modules, different pairing information is distributed to different devices, and the transmission path is changed by adjusting the pairing devices through the domain controller.

Still further, the FPGA implementation method of the edge calculation layer in S4 is as follows: the facility includes:

the self-encoder feature extraction module comprises an 8 x 128 input double-port Random Access Memory (RAM) module, a data conversion module, a feature extraction module and a 16 x 128 input double-port RAM module, wherein the 8 x 128 input double-port RAM module is used for storing original 8-bit data information received from ZigBee, the data conversion module is used for combining two 8-bit data into 16-bit data, the feature extraction module is used for performing feature extraction on a series of 16-bit data containing cutter wear loss information, and finally storing the extracted data into the 16 x 128 input double-port RAM module;

an on-line sequential increment extreme learning machine module comprises 4 16 multiplied by 128 input double-port RAM modules, an activation function operation module, a multiplication and addition module, an inverse matrix operation module and a generalized inverse matrix operation module, the device comprises a plurality of 16X 128 input double-port RAM modules, an on-line sequential increment limit learning machine hidden layer and hidden layer connection weight, an on-line sequential increment limit learning machine input layer and hidden layer connection threshold, normalized flank wear loss, an on-line sequential increment limit learning machine hidden layer and output layer connection weight and on-line sequential increment limit learning machine real-time monitoring data output, wherein the 5X 128 input double-port RAM modules respectively store the on-line sequential increment limit learning machine hidden layer and hidden layer connection weight, the on-line sequential increment limit learning machine input layer and hidden layer connection threshold, the normalized flank wear loss, the on-line sequential increment limit learning machine hidden layer and output layer connection weight, an activation function operation module adopts a Sigmoid function, performs function operation on hidden layer data in a lookup table mode, and a generalized inverse matrix module performs generalized inverse.

Still further, the calculation method of the self-encoder in S4 is as follows:

given a set of unlabeled force sensor acquired data samples { x^m}^M ₁The coding network passing a coding function f_θEach training sample x^mTransformed into a coded vector h^m：

h^m＝f_θ(x^m)＝s_f(Wx^m+b) (1)

In the formula, s_fAn activation function for the coding network; θ is a set of parameters of the coding network, and θ ═ W, b }; then encodes vector h^mBy decoding functions

Inverse transformation to x^mA reconstructed representation of

In the formula, s_gAn activation function for a decoding network; θ ' is a parameter set of the decoding network, and θ ' ═ W ', d }; AE by minimizing x^mAnd

reconstruction error of

And finishing the training of the whole network:

if the vector h is coded^mCan reconstruct x well^mThen, the training sample data is considered to extract most of characteristic information which can represent the sample data and is contained in the training sample data; but only retains x^mThe information of the self-encoder is not enough to enable the self-encoder to obtain a useful feature representation, because the processing environment of the processing equipment is complex, sample data is easily interfered by various factors, and in addition, the processing conditions and working conditions are continuously changed due to complex tasks, so that the parameters and properties of the sample under the same processing condition are fluctuated, and therefore certain condition constraint needs to be given to the self-encoder to enable the self-encoder to learn a feature representation with good robustness; the denoising autoencoder solves the problem by reconstructing sample data containing noise; the core idea is as follows: the coding network adds noise with certain statistical characteristics into sample data, then codes the sample after the noise is added, and the decoding network estimates the original form of the interfered sample from the data which is not interfered according to the statistical characteristics of the noise, so that the denoising autoencoder learns more robust characteristics from the noise-containing sample, and the sensitivity of the denoising autoencoder to tiny random disturbance is reduced;

first, for sample x^mAccording to q_DRandom noise is added in distribution to become noisy samples

Namely, it is

In the formula, q_DRandomly concealing noise for binomial;

then the training of the de-noising self-encoder is completed by optimizing the following objective function

Still further, the FPGA implementation method of the self-encoder feature extraction module in S4 is as follows:

determining the internal implementation structure of the self-encoder feature extraction module by the formula (1), finishing operation in a pipeline mode, wherein the operation is divided into three steps of addition, multiplication and comparison, the three steps are circularly realized, the operation is finished, and when the rising edge of a clock comes, data is converted from the data conversion module and the RAM_WThe memory module is simultaneously input into the multiplication module, the operation result is temporarily stored in the memory, and when the next clock rising edge comes, the data in the memory and the RAM_bThe data in the memory and the RAM are simultaneously transmitted to an addition module, and the result is stored in the next memory, when the rising edge of the next clock comes_jThe data in the memory and RAM are simultaneously transmitted to the multiplication module, the result is stored in the next memory, when the next clock rising edge comes_dThe data in (1) are simultaneously transmitted to the adding module, and the result is stored in

Finally will be

The data in the data conversion module and the data in the data conversion module are simultaneously input into the comparison module, W and B are adjusted, and the adjusted data are respectively stored in the RAM_wAnd RAM_bAnd (5) waiting for next training.

Still further, the calculation method of the online sequential incremental limit learning machine in S4 is as follows:

the learning process of the output weight of the single hidden layer feed-forward neural network by the online sequential increment extreme learning algorithm is mainly divided into two parts, wherein the first part is an initial stage, namely the output weight beta of the single hidden layer feed-forward neural network is obtained through a small number of samples, the second part is an online learning part, namely the output weight beta of the single hidden layer feed-forward neural network learned in the initial stage is updated by using a single sample or a sample data block, and the initial stage is provided with the output weight beta of the single hidden layer feed-forward neural network learned in the initial stageSegment, assume there is N₀An arbitrary training sample (X)_i,t_i) Wherein X is_i＝[x_i1,x_i2,…,x_in]^T∈Rⁿ，t_i＝[t_i1,t_i2,…,t_im]^T∈R^m(ii) a Hopefully, the idea of the basic extreme learning algorithm is utilized to satisfy | | H₀β-T₀Minimum beta | |₀Which comprises the following steps:

in this case, β can be calculated by the generalized inverse calculation method₀：

In the formula

When a new sample enters the model, assume that there is N₁The samples are entered into the model, where the basic concept of extreme learning algorithm is used to hopefully find the equation beta that satisfies⁽¹⁾：

According to the calculation method of the generalized inverse, β⁽¹⁾The values of (A) are:

in the formula

For online learning, we wish to assign β⁽¹⁾Is expressed as beta⁽⁰⁾，K₁，H₁And T₁Function of, K₁Can be expressed as:

while

From this, β is⁽¹⁾Namely:

therefore, we can get a recurrence formula of online learning:

while

Can be obtained from the Woodbury formula:

order to

Equation (15) can be expressed as:

if the online mode is to enter the system as a piece of data, then the update formula (17) has the following simple form:

still further, the FPGA implementation method in the training process of the online sequential incremental limit learning machine in S4 is as follows:

the FPGA implementation method of the on-line sequential increment extreme learning machine module training part can be obtained by the formula (18), is implemented in a pipeline mode and comprises four steps of multiplication and addition, subtraction, activation function calculation, matrix transposition calculation and generalized inverse matrix solution; data in RAM (OSELM) and RAM when a rising clock edge is imminent_wThe data in the buffer memory are simultaneously transmitted to the multiply-add module and temporarily stored in a 16-bit memory, and when the next clock rising edge comes, the data in the memory and the RAM_bThe data in the formula (6) are simultaneously transmitted to an addition module, the calculation result is directly input to an activation function solving module to obtain an H matrix in the formula (6), and the matrix is stored in an RAM_HIn the memory, when the rising edge of the clock comes, the data in the memory is input into the transposition calculation module, and the result is stored in the RAM_HTIn a memory, and RAM_HData RAM in memory_HTThe data in the memory is simultaneously transmitted to the multiplication module to obtain the training part parameter K, and the result is stored in the RAM_kA memory; when the rising edge of the clock comes, the RAM is started_kData in (1), RAM_HTData and RAM in_TThe data in the RAM are simultaneously transmitted to the multiplication module, and the result is stored in the RAM_βIn the memory, when the next clock rising edge comes, the RAM is started_βData and RAM in_TThe data in the buffer memory are simultaneously transmitted to a subtraction module, the result is stored in a memory, and when the next clock rising edge comes, the RAM is used_kData of (5) and data after data is input into inversion module, and RAM_HTThe data in (1) and the data in the memory are simultaneously input into the multiplication module, and the result and the RAM are simultaneously input into the multiplication module_βThe data in the memory are simultaneously input into the addition module to obtain the calculated beta⁽²⁾And circulating the steps to finally obtain the output weight beta after training.

Still further, the FPGA implementation method of the online sequential incremental limit learning machine monitoring process in S4 is as follows:

the FPGA implementation method of the on-line sequential increment extreme learning machine module monitoring part can be obtained by the formula (6), is implemented by adopting a pipeline mode and comprises 5 steps of multiplication and addition, summation and activation function calculation; when the rising edge of the clock comes, RAM₁The data in the buffer memory is transmitted to a multiplication and addition module and temporarily stored in a 16-bit memory, and when the data amount reaches the number of input layer neurons, the data amount is stored in a RAM (random access memory) from a value memory_wReading corresponding weight values, calculating multiple groups of data simultaneously, temporarily storing the calculation results in a memory, transmitting the data in the memory to a summation module when the next clock rising edge comes, and simultaneously, storing the data in the memory in a threshold value memory RAM_bReading corresponding threshold value, making summation calculation, directly inputting calculation result into activating function solving module, storing solved result in memory, when clock rising edge comes, making result in memory and stored in output weight value memory RAM_βAnd simultaneously outputting the corresponding output weight beta to the multiplication and addition module, and summing all the results when the rising edge of the next clock comes to obtain the online abrasion loss of the cutter.

Still further, the FPGA implementation method of the online sequential incremental limit learning machine activation function calculation in S4 is as follows:

the solution of the activation function is realized by adopting a table look-up method, and the activation function is y-1/(1 + e)^-x) When x is-5, y is 0.0067, when x is 5, y is 0.9933, because only two decimal values are needed to be reserved for the final result when calculating the activation function, it can be set that when x is greater than 5, y is 1, when x is less than-5, y is 0, and [ -5,5 [ -5]The interval is divided equally into 100 equal parts at intervals of 0.1, the function value corresponding to each point is calculated using MATLAB, and the function value of + -0.05 for that point is regarded as the function value for that point, for example, [ -4.95, -4.85]The function value of (c) is defined as a function value of-4.9, which is [ -4.85, -4.75 [ ]]The function value of (2) is determined as the function value of (x) -4.8, all the sections and the corresponding function values are stored in the FPGA by using a lookup table, and when the input value is in a fixed range, the corresponding function value can be output.

Preferably, the calculation method of the generalized inverse matrix solution in S4 is as follows:

setting An n-dimensional matrix An × n, and performing LU decomposition on An × n to obtain:

from formula (20):

u_1j＝a_1j,j＝1,2,…,n，l_i1＝a_i1/u₁₁,i＝2,3,…,n (21)

further obtained by multiplication of matrices

The ith row element of U can be obtained

By

If u_ijNot equal to 0, one can obtain:

solving for L and U matrices by recursive solutions, e.g.

Finally obtaining all elements of the L and U matrixes;

according to A^-1＝(LU)^-1＝U^-1L^-1Then, it is required to obtain U^-1、L^-1Namely, the method can be used for preparing the anti-cancer medicine,

let the inverse matrix of U be V, then

It is possible to obtain,

all elements of the V matrix are calculated by using a recurrence method, and the L matrix is only transposed when the inverse matrix of the L is calculated, and then the calculation is carried out according to the method.

Further, the FPGA implementing method of the generalized inverse matrix solving module in S4 is as follows:

from equation (21), first, the data stored in RAM_aCorresponds to a_1jElements of j-1, 2, …, n are directly transferred to RAM_uCorresponds to u_1jJ is stored in RAM at 1,2, …, n address_aCorresponds to a_i12,3, …, n and the element stored in RAM_uU in₁₁Is divided to obtain l_i1And store it in RAM_lThen, using the RAM after the update_uAnd RAM_lMultiplying known elements in the RAM by the product of the known elements in the RAM_aThe corresponding elements in the sequence are subtracted to obtain the corresponding u_iiStoring the result in RAM_uIn the corresponding address; finally, the steps are repeated by using different elements in the memory until the RAM is obtained_uAnd RAM_lAll corresponding elements in the list;

the method can be obtained by the formula (22), and the step of solving the inverse matrix of the upper triangular matrix by using the FPGA comprises the following steps: will be stored in RAM_uThe upper triangular matrix U in the V matrix is subjected to reciprocal transformation by using an inversion module to obtain V in the V matrix_iiElements, and storing them in RAM_vIn RAM_uIn corresponds to u_ikAnd elements stored in RAM_vIn (1) corresponds to v_kjThe elements of (a) are subjected to multiplication and addition operation and finally are summed with v_iiPerforming multiplication to obtain V matrix, and storing in RAM_lThe corresponding lower triangular matrix L in the middle is firstly transposed, and then the steps are repeated.

Still further, the implementation method of the SDN technology layer in S5 is as follows: the facility comprises an SND controller and an OpenFlow switch, wherein the SND controller comprises OpenFlow (a network communication protocol) software, the SND controller is used for communicating with the OpenFlow-compatible switch by using an OpenFlow protocol running on a transport layer security protocol, and forwarding the received tool wear data to a cloud computing layer along a specified data forwarding path by using rules defined by an SND application.

Still further, the implementation method of the cloud computing layer in S6 is as follows: the layer mainly utilizes a cloud service system to store monitoring data in a cloud space and establishes a cloud service standard access protocol, wherein an application program comprises 3 application functions of visual display, real-time early warning and decision assistance, wherein the visual display means that the real-time cutter abrasion loss is displayed on a liquid crystal display screen as a waveform graph and the like; the real-time early warning means that the tool wear loss is used for setting a warning trigger of fixed parameters in advance, and when the wear loss is not in a normal range, the warning trigger is triggered; the decision assistance means that the tool abrasion loss is used for performing auxiliary adjustment on the processing parameters of the processing equipment, so that the adjustment of the processing parameters of the processing equipment is more reasonable.

Compared with the prior art, the beneficial technical effects of the technology are as follows:

1. the real-time monitoring data are processed on the edge nodes of the edge computing layer near the processing equipment, and the real-time monitoring data are not processed on the cloud computing layer, so that the method has the following advantages: (1) the edge node is closer to the position of processing equipment, so that the data processing speed and the data transmission speed are obviously improved, and the time delay is greatly reduced; (2) the data acquired by the sensor of the sensing layer processed by the edge calculation is small data, and the cost advantage is achieved in the aspects of data calculation and storage; (3) the data volume needing to be uploaded to a cloud computing layer is greatly reduced through the edge analysis capability of the edge computing nodes, so that the bandwidth pressure uploaded to the cloud computing layer is reduced; (4) the network of the intelligent cutter edge monitoring system is protected from being attacked through the edge nodes, and the network safety is improved.

2. The online sequential increment extreme learning machine model adopted by the technology has the following advantages: the intelligent learning tool for monitoring the cutter abrasion loss has strong learning capability, and can learn enough cutter abrasion patterns to be communicated; (2) the intelligent monitoring system for the edge of the cutter has continuous self-improvement and updating capability, and can continuously relearn the newly appeared abrasion mode of the cutter in the cutting process and gradually improve the knowledge and understanding of the abrasion state of the cutter in the cutting process; (3) the intelligent monitoring system for the edge of the cutter has the capability of detecting and analyzing the wear state of the cutter in the cutting process fast enough.

Based on the idea of edge intelligence, the technology processes the monitoring data at the edge side of the processing equipment, reduces the fixed delay of data transmission and reduces the data processing time of the main server, and the system adopts the idea of minimum delay, the idea of pipeline programming and the idea of parallel computing in the implementation process, so that the hardware and network resources can be effectively saved and the utilization rate of the resources can be improved on the premise of keeping higher throughput rate and transmission rate. The multi-channel data parallel computation is realized by adopting a Field Programmable Gate Array (FPGA) in the edge computation layer, the efficiency of the system can be effectively improved, and the self-encoder can obtain high-level data characteristics by continuously learning the characteristics of the original data to represent the original data, so that the most core characteristics of the original data can be well extracted by removing noise in the data, and the dimensionality of the data can be reduced; the online sequential incremental extreme learning machine network has the advantages of simple structure, excellent generalization performance, high training speed and high accuracy of prediction results.

Drawings

The present technology is further described in the following description with reference to the drawings.

FIG. 1 is a block diagram of a system for intelligently monitoring the edge of a tool based on an online incremental wear evolution model;

FIG. 2 is a connection diagram of internal modules for implementing an edge computation layer based on an FPGA according to the present technology;

FIG. 3 is a schematic diagram of an FPGA implementation of the self-encoder feature extraction module of the present technology;

FIG. 4 is a schematic diagram of an FPGA implementation of the on-line sequential incremental limit learning machine training portion of the present technique;

FIG. 5 is a schematic diagram of an FPGA implementation of the monitoring portion of the on-line sequential incremental limit learning machine of the present technology;

FIG. 6 is a schematic diagram of an FPGA implementation of the generalized inverse matrix calculation of the online sequential incremental extreme learning machine according to the present technology.

Detailed Description

In order to make the technical problems, technical solutions and advantageous effects to be solved by the present technology clearer and more obvious, the present technology is further described in detail below with reference to the accompanying drawings.

As shown in fig. 1, the method for intelligently monitoring the cutting state edge based on the online incremental wear evolution model in the method of the present technology includes the following steps:

s1: constructing an intelligent monitoring system for the edge of the cutter, as shown in FIG. 1;

specifically, in S1, a tool edge intelligent monitoring system is constructed, which is divided into five logical entities, namely a sensing layer, a data transmission layer, an edge computing layer, an SDN technology layer, and a cloud computing layer. The sensing layer consists of sensors such as cutting force, vibration, acoustic emission and noise, a three-way cutting force sensor is used for collecting cutting force signals of an X axis, a Y axis and a Z axis in the machining process of the machining equipment in real time, a three-way vibration sensor is used for collecting vibration signals in the machining process of the machining equipment, an emission sensor is used for collecting acoustic emission signals in the machining process of the machining equipment, and a noise sensor is used for collecting noise in the machining process of the machining equipment; the data transmission layer consists of a plurality of edge gateways and a domain controller, wherein the edge gateways comprise a ZigBee module, a Wi-Fi module and a Bluetooth module, a wireless data transmission network is formed by utilizing the ZigBee module, the Wi-Fi module and the Bluetooth module, and a data transmission path is controlled by utilizing the domain controller; the edge calculation layer is composed of an FPGA hardware system, the hardware function module comprises a self-encoder feature extraction module and an online sequential increment extreme learning machine module, the self-encoder feature extraction module is used for preprocessing data to obtain a training sample, an internal topological structure of the online sequential increment extreme learning machine module is determined by combining the feature number and the sample number of the training sample, and finally the data acquired by the sensor is transmitted to the edge calculation layer after training under the same working condition; the SND technical layer consists of a plurality of SND controllers, the transmission path between the edge computing layer and the cloud computing layer is adjusted in real time by using the SND controllers, and the real-time monitoring result is sent to the cloud computing layer in an optimal path; the cloud computing layer finishes various specific applications to users and industries, and the application in the technology comprises real-time early warning, visual display and decision assistance, wherein the visual display means that the real-time cutter abrasion loss is displayed on a liquid crystal display screen as a waveform graph and the like; the real-time early warning means that the tool wear loss is used for setting a warning trigger of fixed parameters in advance, and when the wear loss is not in a normal range, the warning trigger is triggered; the decision assistance means that the tool abrasion loss is used for performing auxiliary adjustment on the processing parameters of the processing equipment, so that the adjustment of the processing parameters of the processing equipment is more reasonable.

S2: the sensing layer comprises sensors of cutting force, vibration, acoustic emission, noise and the like, the three-way cutting force sensor of the sensing layer is used for collecting cutting force signals of an X axis, a Y axis and a Z axis in the machining process of the machining equipment in real time, and meanwhile, the three-way vibration sensor and the acoustic emission sensor are used for respectively collecting vibration signals and acoustic emission signals in the machining process of the machining equipment.

S3: utilize zigBee module, Wi-Fi module, bluetooth module in the marginal gateway to constitute wireless data transmission network to utilize domain controller control data transmission's route, include the following:

s3.1: a star network topology structure is formed by the ZigBee module and the domain controller, so that data conversion between the ZigBee and the domain controller can be realized, and the domain controller can control a data transmission path of the ZigBee module in real time;

s3.2: establishing a Wi-Fi wireless transmission network through a router, forming various transmission paths between a sensing layer and an edge computing layer by using different wireless transmission protocols, and adjusting the Wi-Fi transmission paths by controlling the transmission protocols through a domain controller;

s3.3: the transmission link between the sensing layer and the edge computing layer is established by utilizing a plurality of Bluetooth modules, different pairing information is distributed to different devices, and the transmission path is changed by adjusting the pairing devices through the domain controller.

S4: the method comprises the steps of preprocessing data by utilizing an FPGA hardware system of an edge calculation layer to obtain a training sample, determining an internal topological structure of the online sequential increment extreme learning machine by combining the feature number and the sample number of the training sample, and finally transmitting data acquired by a sensor to the trained edge calculation layer under the same working condition.

Specifically, the FPGA-implemented internal module connection diagram of the edge computing layer in S4 is shown in fig. 2, and includes:

the self-encoder feature extraction module comprises an 8 x 128 input double-port RAM module, a data conversion module, a feature extraction module and a 16 x 128 input double-port RAM module, wherein the 8 x 128 input double-port RAM module is used for storing original 8-bit data information received from ZigBee, the data conversion module is used for combining two 8-bit data into 16-bit data, the feature extraction module is used for performing feature extraction on a series of 16-bit data containing cutter wear loss information, and finally storing the extracted data into the 16 x 128 input double-port RAM module;

an on-line sequential increment extreme learning machine module comprises 4 16 multiplied by 128 input double-port RAM modules, an activation function operation module, a multiplication and addition module, an inverse matrix operation module and a generalized inverse matrix operation module, the device comprises a plurality of 16X 128 input double-port RAM modules, an on-line sequential increment limit learning machine hidden layer and hidden layer connection weight, an on-line sequential increment limit learning machine input layer and hidden layer connection threshold, normalized flank wear loss, an on-line sequential increment limit learning machine hidden layer and output layer connection weight and on-line sequential increment limit learning machine real-time monitoring data output, wherein the 5X 128 input double-port RAM modules respectively store the on-line sequential increment limit learning machine hidden layer and hidden layer connection weight, the on-line sequential increment limit learning machine input layer and hidden layer connection threshold, the normalized flank wear loss, the on-line sequential increment limit learning machine hidden layer and output layer connection weight, an activation function operation module adopts a Sigmoid function, performs function operation on hidden layer data in a lookup table mode, and a generalized inverse matrix module performs generalized inverse.

The technology will explain how to utilize FPGA to realize the self-encoder feature extraction module and the online sequential increment extreme learning machine module. The technology is written by adopting VHDL hardware language, the used development platform is Quartus II13.0, and the model of the used FPGA chip is EP4SGX 230.

Specifically, the implementation of the self-encoder module in S4: given a set of unlabeled force sensor acquired data samples { x^m}^M ₁The coding network passing a coding function f_θEach training sample x^mTransformed into a coded vector h^m：

h^m＝f_θ(x^m)＝s_f(Wx^m+b) (1)

In the formula, s_fAn activation function for the coding network; θ is a set of parameters of the coding network, and θ ═ W, b }; then encodes vector h^mBy decoding functions

Inverse transformation to x^mA reconstructed representation of

In the formula, s_gAn activation function for a decoding network; θ ' is a set of parameters of the decoding network, and θ ' ═ W ', d }, AE by minimizing x^mAnd

reconstruction error of

And finishing the training of the whole network:

if the vector h is coded^mCan reconstruct x well^mThen it is considered that most of the feature information which can characterize the sample data and is contained in the training sample data is extracted, but only x is retained^mThe information of the self-encoder is not enough to enable the self-encoder to obtain a useful feature representation, because the processing environment of the processing equipment is complex, sample data is easily interfered by various factors, and in addition, the processing conditions and working conditions are continuously changed due to complex tasks, so that the parameters and properties of the sample under the same processing condition are fluctuated, and therefore certain condition constraint needs to be given to the self-encoder to enable the self-encoder to learn a feature representation with good robustness; the denoising autoencoder solves the problem by reconstructing sample data containing noise; the core idea is as follows: the coding network adds noise with certain statistical characteristics into sample data, then codes the sample after the noise is added, and the decoding network estimates the original form of the interfered sample from the data which is not interfered according to the statistical characteristics of the noise, so that the denoising autoencoder learns more robust characteristics from the noise-containing sample, and the sensitivity of the denoising autoencoder to tiny random disturbance is reduced;

first, for sample x^mAccording to q_DRandom noise is added in distribution to become noisy samples

Namely, it is

In the formula q_DRandomly concealing noise for binomial;

training of the DAE is then accomplished by optimizing the following objective function

Specifically, in S4, the FPGA of the self-encoder feature extraction module is implemented as follows, as shown in fig. 3, the internal implementation structure of the self-encoder feature extraction module is determined by formula (1), the operation is completed in a pipeline manner, the operation is divided into three steps of addition, multiplication, and comparison, and the three steps are implemented circularly to complete the operation, including the following steps:

s4.1: when the rising edge of the clock comes, the data slave data conversion module and the RAM_wThe storage module is simultaneously input into the multiplication module, and the operation result is temporarily stored in the memory;

s4.2: when next clock rising edge comes, data in the memory and RAM_bThe data in the data storage module are simultaneously transmitted to an addition module, and the result is stored into a next memory;

s4.3: when next clock rising edge comes, data in the memory and RAM_jThe data in the data storage module are simultaneously transmitted to a multiplication module, and the result is stored in a next memory;

s4.4: when next clock rising edge comes, data in the memory and RAM_dThe data in (1) are simultaneously transmitted to the adding module, and the result is stored in

S4.5: will most probably

The data in the data conversion module and the data in the data conversion module are simultaneously input into the comparison module, w and b are adjusted, and the adjusted data are respectively stored in the RAM_wAnd RAM_bAnd (5) waiting for next training.

Specifically, the calculation method of the online sequential extreme learning machine in S4 is as follows:

the learning process of the online sequential increment extreme learning machine algorithm for the output weight of the single hidden layer neural network is mainly divided into two parts, wherein the first part is an initial stage, namely, the learning process is realized by passingA small number of samples obtain the output weight beta of the single hidden layer feedforward neural network, the second part is an online learning part, namely, the output weight beta of the single hidden layer feedforward neural network learned in the initial stage is updated by using a single sample or a sample data block, and in the initial stage, N is assumed to exist₀An arbitrary training sample (X)_i,t_i) Wherein X is_i＝[x_i1,x_i2,…,x_in]^T∈Rⁿ，t_i＝[t_i1,t_i2,…,t_im]^T∈R^m(ii) a Hopefully, the idea of the basic extreme learning algorithm is utilized to satisfy | | H₀β-T₀Minimum beta | |₀Wherein:

in this case, β can be calculated by the generalized inverse calculation method₀：

Wherein

When a new sample enters the model, assume that there is N₁The samples are entered into the model, where the basic concept of extreme learning algorithm is used to hopefully find the equation beta that satisfies⁽¹⁾：

According to the calculation method of the generalized inverse, β⁽¹⁾The values of (A) are:

wherein the content of the first and second substances,

for online learning, we wish to assign β⁽¹⁾Is expressed as beta⁽⁰⁾，K₁，H₁And T₁Function of, K₁Can be expressed as:

while

From this, β is⁽¹⁾Namely:

therefore, we can get a recurrence formula of online learning:

while

Can be obtained from the Woodbury formula:

order to

Equation (15) can be expressed as:

if the online mode is to enter the system as a piece of data, then the update formula (17) has the following simple form:

specifically, the FPGA implementation method of the online sequential increment extreme learning machine module in S4 is as follows, as shown in fig. 4, the FPGA implementation method of the online sequential increment extreme learning machine module training portion that can be obtained by equation (18) is implemented in a pipeline manner, and includes four steps of multiplication and addition, subtraction, activation function calculation, matrix transposition calculation, and generalized inverse matrix solution, and includes the following steps:

s4.6: data in RAM (OSELM) and RAM when a rising clock edge is imminent_wThe data in the data storage module are simultaneously transmitted to a multiplication and addition module and temporarily stored in a 16-bit memory;

s4.7: when next clock rising edge comes, data in the memory and RAM_bThe data in the formula (6) are simultaneously transmitted to an addition module, the calculation result is directly input to an activation function solving module to obtain an H matrix in the formula (6), and the matrix is stored in an RAM_HA memory;

s4.8: when the rising edge of the clock comes, the data in the memory is input into the transposition calculation module, and the result is stored in the RAM_HTIn a memory, and RAM_HData RAM in memory_HTThe data in the memory is simultaneously transmitted to the multiplication module to obtain the training part parameter K, and the result is stored in the RAM_kA memory;

s4.9: when the rising edge of the clock comes, the RAM is started_kData in (1), RAM_HTData and RAM in_TThe data in the RAM are simultaneously transmitted to the multiplication module, and the result is stored in the RAM_βA memory;

s4.10: when the next clock rising edge comes, RAM is started_βData and RAM in_TThe data in the step (a) are simultaneously transmitted to a subtraction module, and the result is stored in a memory;

s4.11: when the next clock is onWhen rising edge comes, RAM is driven_kData of (5) and data after data is input into inversion module, and RAM_HTThe data in (1) and the data in the memory are simultaneously input into the multiplication module, and the result and the RAM are simultaneously input into the multiplication module_βThe data in the memory are simultaneously input into the addition module to obtain the calculated beta⁽²⁾；

S4.12: and circulating the above steps to finally obtain the output weight beta after training.

Specifically, the FPGA implementation method of the on-line sequential increment extreme learning machine module monitoring part in S4 is as follows, as shown in fig. 5, implemented in a pipeline manner, and includes 5 steps of multiply-add, sum, and activation function calculation, where the steps include:

s4.13: when the rising edge of the clock comes, RAM₁The data in the buffer memory is transmitted to a multiplication and addition module and temporarily stored in a 16-bit memory, and when the data amount reaches the number of input layer neurons, the data amount is stored in a RAM (random access memory) from a value memory_wReading corresponding weight values, simultaneously calculating a plurality of groups of data, and temporarily storing the calculation results in a memory;

s4.14: when the next rising clock edge comes, the data in the memory is supplied to the summing module while the RAM is running from the threshold memory_bReading the corresponding threshold value, carrying out summation calculation, directly inputting the calculation result into an activation function solving module, and storing the solving result in a memory;

s4.15: when the rising edge of the clock comes, the result in the memory is stored in the output weight value memory RAM_βThe corresponding output weight beta in the step (b) is simultaneously input into a multiplication and addition module;

s4.16: and when the next clock rising edge comes, summing all the results to obtain the online abrasion loss of the cutter.

Specifically, the FPGA implementation method for online sequential incremental limit learning machine activation function calculation in S4 is as follows:

the solution of the activation function is realized by adopting a table look-up method, and the activation function is y-1/(1 + e)^-x) When x is-5, y is 0.0067, when x is 5, y is 0.9933, because only the final result needs to be retained when calculating the activation functionTwo decimal places, therefore, it can be set that when x is greater than 5, y is 1, when x is less than-5, y is 0, will [ -5,5]The interval is divided equally into 100 equal parts at intervals of 0.1, the function value corresponding to each point is calculated using MATLAB, and the function value of + -0.05 for that point is regarded as the function value for that point, for example, [ -4.95, -4.85]The function value of (c) is defined as a function value of-4.9, which is [ -4.85, -4.75 [ ]]The function value of (a) is defined as a function value of-4.8. And storing all the intervals and the corresponding function values in the FPGA in a form of a lookup table, and outputting the corresponding function values when the input values are in a fixed range.

Specifically, the calculation method of the inverse matrix solution in S4 is as follows: setting An n-dimensional matrix An × n, and performing LU decomposition on An × n to obtain:

from formula (19)

u_1j＝a_1j,j＝1,2,…,n，l_i1＝a_i1/u₁₁,i＝2,3,…,n (20)

Further obtained by multiplication of matrices

The ith row element of U can be obtained

By

If u_ijNot equal to 0, available

Solving for L and U matrices by recursive solutions, e.g.

Finally, all elements of the L and U matrixes are obtained.

According to A^-1＝(LU)^-1＝U^-1L^-1Then, it is required to obtain U^-1，L^-1Namely, the method can be used for preparing the anti-cancer medicine,

let the inverse matrix of U be V, then

Can obtain the product

All elements of the V matrix are calculated by using a recurrence method, and the L matrix is only transposed when the inverse matrix of the L is calculated, and then the calculation is carried out according to the method.

Specifically, the FPGA implementing method of inverse matrix solution in S4 is as follows, as shown in fig. 6, including the following:

s4.17: will be stored in RAM_aCorresponds to a_1jElements of j-1, 2, …, n are directly transferred to RAM_uCorresponds to u_1jJ is stored in RAM at 1,2, …, n address_aCorresponds to a_i12,3, …, n and the element stored in RAM_uU in₁₁Is divided to obtain l_i1And is combined withStore it into RAM_lWithin the corresponding address of (a);

s4.18: using RAM after update_uAnd RAM_lMultiplying known elements in the RAM by the product of the known elements in the RAM_aThe corresponding elements in the sequence are subtracted to obtain the corresponding u_iiStoring the result in RAM_uIn the corresponding address;

s4.19: repeating the above steps by using different elements in the memory until a RAM is obtained_uAnd RAM_lAll corresponding elements in the list.

From equation (21), the inverse of the upper triangular matrix using the FPGA is:

s4.20: will be stored in RAM_uThe upper triangular matrix U in the V matrix is subjected to reciprocal transformation by using an inversion module to obtain V in the V matrix_iiElements, and storing them in RAM_vPerforming the following steps;

s4.21: will be stored in RAM_uIn corresponds to u_ikAnd elements stored in RAM_vIn (1) corresponds to v_kjThe elements of (a) are subjected to multiplication and addition operation and finally are summed with v_iiPerforming product operation to obtain a V matrix;

s4.22: will be stored in RAM_lFirstly, transposing the corresponding lower triangular matrix L;

s4.23: the above is repeated.

S5: the SDN technology layer is realized by the following method that the facility comprises an SND controller containing OpenFlow software and an OpenFlow switch, wherein the SND controller communicates with the OpenFlow-compatible switch by using an OpenFlow protocol running on a transport layer security protocol, and forwards the received tool wear data to a cloud computing layer along a specified data forwarding path by using rules defined by SND application.

S6: the cloud computing layer is mainly used for storing monitoring data in a cloud space by using a cloud service system and establishing a cloud service standard access protocol, wherein an application program mainly comprises 3 application functions of visual display, real-time early warning and decision assistance, wherein the visual display means that the real-time cutter abrasion loss is displayed on a liquid crystal display screen as a waveform graph and the like; the real-time early warning means that the tool wear loss is used for setting a warning trigger of fixed parameters in advance, and when the wear loss is not in a normal range, the warning trigger is triggered; the decision assistance means that the tool abrasion loss is used for performing auxiliary adjustment on the processing parameters of the processing equipment, so that the adjustment of the processing parameters of the processing equipment is more reasonable.

Whether the intelligent learning tool selects reasonably has great influence on the monitoring efficiency and the diagnosis service level of the tool abrasion loss. The intelligent learning tools for monitoring the tool wear amount in the prior art all adopt a batch learning method and an off-line training mode, so that the requirements of on-line monitoring are met, previous learning achievements need to be abandoned, and the learning and training need to be repeated according to newly added data and all past data. Relearning all of the data results in a significant amount of time and space resources being consumed. As the scale of online monitoring data increases, the demand for time and space increases rapidly, and eventually, the learning speed cannot catch up with the data updating speed. In addition, the application of the intelligent learning tool for monitoring the tool wear amount in the prior art also assumes that the tool wear state during the cutting process is monitored in a mode known in advance. However, in practical applications, the training samples are usually not available all at once, various tool wear patterns are often obtained gradually over time, and the information reflected by the samples may change over time. Therefore, in order to enable the intelligent learning tool for monitoring the cutter abrasion loss to have strong learning capacity, a sufficient number of cutter abrasion patterns can be learned and can be communicated; the tool wear model learning system also has continuous self-improvement and updating capability, and can continuously relearn a new tool wear model in the cutting process, and gradually improve the knowledge and understanding of the tool wear state in the cutting process; meanwhile, the system also has the capability of detecting and analyzing the wear state of the cutter in the cutting process fast enough; the technology is based on the idea of edge intelligence, adopts an online incremental learning strategy, takes a single-layer feedforward neural network as a base, establishes a rapid online sequential neural network model, and uses the model as an intelligent learning tool to monitor the cutter abrasion in the cutting process online in real time, thereby improving the learning speed and the accuracy of the prediction result.

The above-described embodiments are merely illustrative of the preferred embodiments of the present technology, and do not limit the scope of the present technology, and various modifications and improvements made to the technical solutions of the present technology by those skilled in the art without departing from the spirit of the present technology should fall within the protection scope defined by the claims of the present technology.

Claims (9)

1. An intelligent monitoring method for tool wear based on edge calculation is characterized by comprising the following steps:

s1: establishing an intelligent monitoring system for the edge of the cutter;

s2: the method comprises the steps that a three-way cutting force sensor of a sensing layer in an intelligent cutter edge monitoring system is used for collecting cutting force signals of an X axis, a Y axis and a Z axis in the machining process of machining equipment in real time, and meanwhile, a three-way vibration sensor and an acoustic emission sensor are used for respectively collecting vibration signals and acoustic emission signals in the machining process of the machining equipment;

s3: a ZigBee module, a Wi-Fi module and a Bluetooth module in the edge gateway are utilized to form a wireless data transmission network, a domain controller is utilized to control a data transmission path, and finally, data are transmitted from a sensing layer to an edge computing layer;

s4: preprocessing data by using an FPGA hardware system of an edge calculation layer to obtain a training sample, determining an internal topological structure of the on-line sequential increment extreme learning machine by combining the feature number and the sample number of the training sample, and finally transmitting the data acquired by the sensor to the edge calculation layer after training under the same working condition;

s5: sending a real-time monitoring result to a cloud computing layer by utilizing an SDN technology layer in the tool edge intelligent monitoring system;

s6: and applying the tool wear monitoring result by using the application service in the cloud computing layer.

2. The intelligent monitoring method for tool wear based on edge calculation according to claim 1, wherein in S3, a domain controller is used to control a data transmission path, and the data transmission path is optimized by using edge transmission and internal transmission methods to realize adjustment of the data transmission path between the sensing layer and the edge calculation layer.

3. The intelligent monitoring method for tool wear based on edge calculation according to claim 1, wherein in S4, a self-encoder feature extraction module in an FPGA hardware system is used to perform feature extraction on information collected by the sensor in real time.

4. The intelligent monitoring method for tool wear based on edge calculation according to claim 1, wherein in S4, an online sequential incremental extreme learning module in an FPGA hardware system is used to perform online sequential incremental learning on the information after feature extraction.

5. The intelligent tool wear monitoring method based on edge computing according to claim 1, wherein in S5, the SND controller uses an OpenFlow protocol running on top of a transport layer security protocol to communicate with an OpenFlow compatible switch, and forwards the received tool wear amount data to the cloud computing layer along a specified data forwarding path using rules defined by the SND application.

6. The intelligent monitoring method for tool wear based on edge calculation as claimed in claim 1, wherein in S6, the application service includes a visual display function, and the real-time tool wear amount is displayed as a waveform graph on the liquid crystal display screen by using the visual display.

7. The intelligent monitoring method for tool wear based on edge calculation as claimed in claim 1, wherein in S6, the application service includes a real-time warning function, a warning trigger for setting fixed parameters in advance is implemented by using the real-time warning function, and the warning trigger is triggered when the wear amount is not in a normal range.

8. The intelligent monitoring method for tool wear based on edge calculation as claimed in claim 1, wherein in S6, the application service includes a decision-making auxiliary function, and the decision-making auxiliary function is utilized to perform auxiliary adjustment on the machining parameters of the machining equipment in combination with the tool wear amount, so that the adjustment of the machining parameters of the machining equipment is more reasonable.

9. The intelligent tool wear monitoring method based on edge calculation as claimed in claim 1, wherein the intelligent tool edge monitoring system comprises:

the sensing layer comprises a three-way cutting force sensor, a three-way vibration sensor and an acoustic emission sensor, the three-way cutting force sensor is used for collecting cutting force signals of an X axis, a Y axis and a Z axis in the machining process of the machining equipment in real time, the three-way vibration sensor is used for collecting vibration signals in the machining process of the machining equipment, and the acoustic emission sensor is used for collecting acoustic emission signals in the machining process of the machining equipment;

the data transmission layer comprises a plurality of edge gateways and a domain controller, wherein the edge gateways comprise a ZigBee module, a Wi-Fi module and a Bluetooth module, form a wireless data transmission network by utilizing the ZigBee module, the Wi-Fi module and the Bluetooth module, and control a data transmission path by utilizing the domain controller;

the edge calculation layer comprises a programmable logic gate array hardware system, the hardware function module comprises a self-encoder feature extraction module and an online sequential increment extreme learning machine module, the self-encoder feature extraction module is used for preprocessing data to obtain a training sample, the internal topological structure of the online sequential increment extreme learning machine module is determined by combining the feature number and the sample number of the training sample, and finally the data acquired by the sensor is transmitted to the edge calculation layer after training under the same working condition;

the software defined network technology layer comprises a plurality of SND controllers, the SND controllers are used for adjusting a transmission path between the edge computing layer and the cloud computing layer in real time, and real-time monitoring results are sent to the cloud computing layer in an optimal path;

and the cloud computing layer is used for finishing various specific applications including real-time early warning, visual display and decision assistance to users and industries.

Images