Search

CN-122020357-A - Cutter state monitoring method based on multisource information fusion

CN122020357ACN 122020357 ACN122020357 ACN 122020357ACN-122020357-A

Abstract

The invention discloses a cutter state monitoring method based on multi-source information fusion, which comprises the steps of collecting vibration signals, sound signals, current signals and temperature signals of a cutter in different states, respectively carrying out pretreatment such as denoising, standardization, resampling and the like aiming at the characteristics of each mode signal so as to eliminate interference and unify data references, generating a time-frequency diagram through continuous wavelet transformation on vibration and sound and high-frequency dynamic signals, extracting local impact characteristics by combining a 2D convolution neural network, realizing the collaborative capture of time-frequency domain global information and time domain local details, converting one-dimensional time sequence data into a time sequence-characteristic matrix for current and temperature low-frequency slowly-varying signals, and finally inputting the fusion characteristics into a full-connection classifier to finish multi-source domain joint diagnosis. According to the method, the cutter abrasion state can be reflected more accurately through the customized feature extraction and intelligent fusion strategy aiming at different signal physical properties, and the robustness and accuracy of monitoring are effectively improved.

Inventors

  • CHEN JIE
  • ZHANG LUYAO
  • YIN ZHENKUN
  • ZHANG JIN

Assignees

  • 南京工大数控科技有限公司

Dates

Publication Date
20260512
Application Date
20251224

Claims (7)

  1. 1. The cutter state monitoring method based on multi-source information fusion is characterized by comprising the following steps of: S1, collecting vibration signals, sound signals, current signals and temperature signals of a cutter in different states through different sensors, and providing an original data basis for subsequent analysis; S2, respectively preprocessing aiming at physical characteristic differences of signals of different modes; S3, generating a time-frequency diagram through continuous wavelet transformation aiming at vibration and sound high-frequency dynamic signals, and reserving the time domain-frequency domain joint characteristics; S4, carrying out feature extraction by adopting a 2D convolutional neural network (2D CNN) aiming at the time-frequency diagram generated in the S3; S5, constructing a time sequence-feature matrix by a sliding window method aiming at low-frequency slow-change signals of current and temperature, calculating statistics such as mean value, variance, rising rate and the like in a window, and converting one-dimensional time sequence data into structural two-dimensional features; S6, adopting a two-dimensional characteristic of current and temperature to adopt a two-way long-short-term memory network, namely, two-way LSTM modeling forward and backward time sequence dependence, introducing a focusing key wear stage trend mutation point of an attention mechanism, and extracting trend characteristics reflecting load accumulation change; S7, introducing a cross-mode attention mechanism, respectively distributing dynamic weights for high-frequency features and low-frequency features, and learning the importance of each mode under different cutting working conditions through a full-connection layer, so as to inhibit the weight of a noise interference mode and realize the self-adaptive information interaction and fusion of the two paths of features; S8, inputting the fused feature vectors into a full-connection classifier, outputting a classification result of the cutter wear state through a softmax activation function, and completing multi-source domain information joint diagnosis to realize precise identification of the cutter state.
  2. 2. The method for monitoring the cutter state based on multi-source information fusion as claimed in claim 1, wherein in S2, preprocessing operations are respectively performed according to physical characteristic differences of different mode signals, and the method mainly comprises the following steps: S2.1, denoising the vibration signal by adopting a wavelet threshold value to eliminate mechanical interference. Wavelet decomposing the vibration signal, selecting Symlet8 wavelet base, and processing the vibration signal Performing N-layer decomposition (in general ) Obtaining approximation coefficients And detail coefficient The decomposition formula is as follows: ; ; Wherein the method comprises the steps of The scale function and the wavelet function are respectively adopted, and j is the number of decomposition layers. The original signal is decomposed into approximate coefficients through a decomposition formula, low-frequency trend and detail coefficients are reflected, high-frequency noise and impact characteristics are reflected, then the detail coefficients are subjected to threshold processing, and a general threshold is adopted: wherein Is the standard deviation of noise, N is the signal length, and the coefficient after thresholding Reconstructing a signal The mechanical interference is effectively eliminated, key characteristics related to cutter abrasion in the vibration signal are reserved, and the formula is as follows: ; And S2.2, suppressing the environmental noise on the sound signal through band-pass filtering. The main frequency range of sound is determined when the cutter is in a working state, environmental noise is avoided, a Butterworth band-pass filter (Butterworth Bandpass Filter) is selected for filtering, and the transfer function is as follows: ; Wherein the method comprises the steps of As a result of the center frequency, As a quality factor, the quality factor is, Is the cut-off frequency. Performing convolution operation on the sound signal through the filter, retaining the effective cutting sound signal in the main frequency range, and inhibiting the interference of environmental noise; And S2.3, carrying out Kalman filtering smoothing power grid fluctuation processing on the current signal. The state equation and the observation equation are established as follows: the equation of state: wherein Is process noise; Observation equation: wherein To observe noise. The filtering process comprises prediction and updating, and finally outputting a smoothed current signal sequence; s2.4, for the possible instantaneous jump variation constant value of the temperature sensor, adopting Detecting and correcting the criterion, calculating local mean value of temperature sequence T (T) by window W And standard deviation If the temperature T (T) at a certain time satisfies |T (T) - |>3 The abnormal value is determined and corrected. The correction formula is: ; wherein k is the window radius; S2.5, respectively performing z-score standardization on the preprocessed vibration, sound, current and temperature signals, and eliminating dimension differences: , wherein, An i-th data point of the original signal, And The mean and standard deviation of the signal respectively, The method comprises the steps of obtaining a normalized value, realizing resampling through time stamp alignment, unifying a data sampling rate, adopting downsampling for vibration signals and sound signals with high sampling rate, adopting linear interpolation complement for current and temperature signals with low sampling rate, realizing time domain alignment of multi-mode signals through time stamp matching, and ensuring that features at the same moment can be associatively analyzed.
  3. 3. The method for monitoring the cutter state based on multi-source information fusion according to claim 1, wherein in the step S3, a time-frequency diagram is generated by continuous wavelet transformation aiming at vibration and an audio high-frequency dynamic signal, and a time-frequency domain joint characteristic is reserved, and the method mainly comprises the following steps: And S3.1, carrying out continuous wavelet transformation Continuous Wavelet Transform on the preprocessed vibration signal and sound signal to mine implicit time-frequency localization characteristics. As high frequency dynamic signals, the instantaneous impact of the vibration signal, such as intermittent contact of the tool with the workpiece and frequency drift of the sound signal, such as cutting noise dominant frequency drift caused by increased wear, all have significant time-varying characteristics, while continuous wavelet transformation can achieve refined decomposition of the signal at different time and frequency scales while maintaining good time-domain and frequency-domain localization capabilities by convolving the signal with a series of wavelet basis functions generated by mother wavelet stretching and translating. The Continuous Wavelet Transform (CWT) formula is: ; Wherein: Is the signal after the pre-processing of the signal, Is a complex conjugate form of a wavelet function, defined as: , is the scale factor (controlling the frequency resolution), Is a panning factor (controlling the temporal resolution), Is a mother wavelet function. Results Reflecting signals at different scales And time of The amplitude of the component corresponds to the energy intensity of the position in the time-frequency domain, so that a time-frequency diagram capable of intuitively reflecting the time-frequency distribution characteristics of the signal is constructed, and a structured time-frequency information support is provided for the subsequent 2D CNN extraction of local impact characteristics; And S3.2, drawing a time-frequency diagram by using imshow functions, drawing the transformation coefficients coeffientions into a two-dimensional image, wherein the x-axis represents time and the y-axis represents frequency. The drawing method is as follows: ; the extracted energy features, statistical features, spectrum entropy and the like form feature vectors ; S3.3, trying different scale ranges and color mapping, setting the scale ranges according to the main frequency characteristics of the signals, and ensuring that a target frequency segment is positioned in the central area of the image; And S3.4, under the condition of higher resolution requirement, increasing the scale quantity can improve the frequency resolution, but the calculated quantity can also be increased, and the frequency resolution is dynamically adjusted according to the diagnosis precision requirement and the real-time requirement.
  4. 4. The method for monitoring the cutter state based on the multi-source information fusion of claim 1, wherein in the step S4, a 2D convolutional neural network (2D CNN) is adopted for feature extraction aiming at the time-frequency diagram generated in the step S3, and the method mainly comprises the following steps of: S4.1, designing a network input layer, standardizing a time-frequency diagram to a [0,1] interval, inputting the time-frequency diagram into the network, setting the input size to 256 multiplied by 1 according to the resolution of the time-frequency diagram, and directly transmitting original time-frequency distribution characteristics without compression treatment for the input layer, wherein a high-frequency impact area is represented as a local high-gray value plaque, a low-frequency trend area is represented as a continuous texture, and providing clear characteristic anchor points for subsequent convolution operation; s4.2, shallow layer feature extraction, namely adopting a stacked structure of convolution, batch normalization, activation and pooling, focusing basic visual features such as edges, textures and the like of a time-frequency diagram, and corresponding to a primary signal mode of a cutter state: 1 st convolution block 13 x3 convolution layers, 64 filters, step size 1, padding set to "same" to keep the size. After Batch Normalization is accelerated and converged, the nonlinear expression is enhanced through a ReLU activation function, and then a 2X 2 maximum pooling layer is adopted, the step size 2 compresses the feature map to 128X 64, and the obvious edge features are reserved; the 2 nd convolution block is formed by 1 3X 3 convolution layer, 128 filters, further capturing texture features formed by edge combination, and compressing to 64X 128 through 2X 2 max pooling after batch normalization and ReLU activation; The 3 rd convolution block is composed of 13×3 convolution layers and 256 filters, extracts more complex local modes corresponding to the evolution trend of frequency components in the tool abrasion process, and the dimension of the pooled characteristic diagram is 32×32×256; S4.3, after the basic visual characteristics are obtained through shallow characteristic extraction, further mining abstract semantic characteristics which are strongly related to the cutter state through a deep network, and mining high-order characteristics which are strongly related to the abrasion state in a time-frequency diagram by expanding a filter receptive field and increasing the number of channels: The 4 th convolution block adopts 1 5X 5 convolution layer, 512 filters, step length 1, padding 'same', a receptive field covering a time-frequency area with a larger range, capturing cross-time-frequency associated features, and after batch normalization and ReLU activation, reserving key semantic information in a 32X 512 feature map through global average pooling GAP compression space dimension, and outputting 1X 512 features; The 5 th convolution block is used for introducing a1 multiplied by 1 convolution layer and 256 filters, compressing channel dimensions, reducing redundancy characteristics, simultaneously suppressing overfitting through Dropout with rate=0.5, and finally outputting 256-dimensional feature vectors, wherein the 256-dimensional feature vectors comprise a high-order mode directly related to a cutter abrasion state in a time-frequency diagram; and S4.4, specifically designing feature extraction, aiming at a vibration signal time-frequency diagram, mainly capturing the shape change of a high-frequency impact area, extracting plaque edges through shallow layer 3×3 convolution, and capturing the spatial correlation of plaque distribution through deep layer 5×5 convolution. Aiming at the texture continuity of the sound signal time-frequency diagram, the local detail and the global trend are balanced through multi-scale convolution. The attention mechanism is introduced to complement the time-frequency space features extracted by the 2D CNN and the time-domain local impact features extracted by the 1D CNN, so that a multi-dimensional feature system of 'time-domain instantaneous features-time-frequency space modes' is constructed together, and the discrimination of the cutter abrasion state is improved.
  5. 5. The method for monitoring the cutter state based on multi-source information fusion according to claim 1, wherein in the step S5, a time sequence-feature matrix is constructed by a sliding window method aiming at low-frequency slow-varying signals of current and temperature, statistics such as mean value, variance and rising rate are calculated in a window, and one-dimensional time sequence data are converted into structural two-dimensional features, and the method mainly comprises the following steps: S5.1, inputting the preprocessed current signal And a temperature signal Dividing each signal sequence by using a sliding window with fixed length, setting the window length L according to signal characteristics and diagnosis requirements, and converting the window length L into points by combining sampling frequency fs, wherein the formula is as follows: ; To control the time resolution, the sliding step S takes 20% of L: ; s5.2, dividing each signal into a plurality of non-overlapping subsequences, wherein the formula is as follows: ; Is obtained together with A window, calculating statistics of basic statistics, trend features, morphological features and the like in the window, converting one-dimensional time sequence data into structured two-dimensional features, 。
  6. 6. The method for monitoring the cutter state based on multi-source information fusion according to claim 1, wherein in the step S6, a two-dimensional characteristic of current and temperature is adopted to adopt a two-way long-short-term memory network, namely, a two-way LSTM modeling forward and backward time sequence dependence, a attention mechanism is introduced to focus on trend mutation points of a key abrasion stage, and trend characteristics reflecting load accumulation change are extracted, and the method mainly comprises the following steps: S6.1, inputting a time sequence-feature matrix constructed by sliding windows, wherein the dimension is T multiplied by F, T is the number of windows, and F is the statistical feature number of each window; s6.2 Using bidirectional LSTM modeling, the matrix is input into a bidirectional LSTM (Bidirectional LSTM) network, the matrix is then input into a network Input bidirectional LSTM network: ; ; ; Outputting a hidden state sequence of all time steps Wherein H is the number of hidden units of the single-layer LSTM; s6.3, introducing an attention mechanism, giving different weights to the hidden state sequence, and focusing on a key time window with obvious wear mutation; Calculating attention weight: ; ; the parameters may be learned and the parameters may be used, Is a randomly initialized context vector; Weighted context vector: ; outputting feature vectors for focus critical wear stage ; And S6.4, finally outputting a characteristic, wherein the context vector c is a global trend representation of a current or temperature mode, reflects the evolution characteristic of a signal along with time and emphasizes a key turning period.
  7. 7. The method for monitoring the cutter state based on multi-source information fusion according to claim 1, wherein in the step S7, a cross-mode attention mechanism is introduced, dynamic weights are respectively distributed for high-frequency features and low-frequency features, the importance of each mode under different cutting working conditions is learned through a full-connection layer, the weights of noise interference modes are restrained, and the self-adaptive information interaction and fusion of the two paths of features are realized, and the method mainly comprises the following steps: s7.1, the time-frequency diagram characteristic of the high-frequency mode extracted in the S4 is marked as The feature vector of the low-frequency mode extracted in the S6 is recorded as ; S7.2 mapping the two types of features to the same dimension d through the fully connected layer: ; For a pair of And (3) with Each assigned a learnable weight The method represents the importance of the depression of each mode under the current working condition, realizes the adaptive fusion of modes, weakens the interference of the mode with low signal to noise ratio on the judging result, and has the following formula: ; ; S7.3, inputting the features fused in the S7.2 into the full connection layer for further information integration The feature vectors after self-adaptive fusion are as follows: Where K is the fusion feature dimension.

Description

Cutter state monitoring method based on multisource information fusion Technical Field The invention relates to the field of cutter state monitoring of numerical control gear milling machines, in particular to a cutter state monitoring method based on multi-source information fusion. Background With the progress of industrial automation, the role of the numerical control gear milling cutter in the manufacturing industry is becoming more and more critical, and the performance of the numerical control gear milling cutter directly determines the processing quality and efficiency. However, tool fault diagnosis faces double challenges, namely on one hand, the cutting process is influenced by multiple factors such as material properties, cutting parameters, environmental conditions and the like, so that the abrasion and fault expression forms are various, the diagnosis complexity is increased, and on the other hand, the strong vibration and noise in the working environment easily mask the real fault characteristics, so that the signal analysis is greatly hindered. In this context, the multi-sensor information fusion (MSIF) technique becomes an important means to break through the bottleneck. The technology is born in the beginning of the 70 th century of the 20 th century, and generates complete and accurate comprehensive information by collecting and integrating multi-source and multi-format information, and is initially applied to target tracking and pose detection in the military field. With the development of information fusion theory, statistics and artificial intelligence technology, the method becomes a research hot spot in the fields of complex industrial monitoring, equipment fault diagnosis and the like. In the process state monitoring, MSIF effectively overcomes the limitation of a single sensor by virtue of information redundancy, complementarity and synergy. Compared with a single sensor, the multi-source sensor has the advantages of richer multi-source information and higher reliability, and provides theoretical and technical support for analyzing the processing state causes and improving the quality level. At present, the technology is applied to milling, turning, drilling and other processing modes, and the stable operation of the processing process is ensured by fusing multimode signals such as vibration, sound and the like. However, the existing research still has a plurality of limitations, the heterogeneous fusion depth of multi-mode data is insufficient, most of the research stays in shallow fusion stages such as feature splicing and the like, physical association excavation of vibration and current signals is insufficient, cross-mode deep semantic association is difficult to form, model generalization capability is obviously restricted by working conditions, when cutting parameters are greatly changed, sensor signal distribution is easy to deviate, model precision collapse based on fixed working condition training is caused, and a dynamic working condition self-adaption mechanism is lacked. Disclosure of Invention The invention provides a cutter state monitoring method based on multi-source information fusion, which aims to overcome the defects in the prior art and starts from a state monitoring principle of multi-sensor information fusion. The invention is realized by the following technology: a cutter state monitoring method based on multi-source information fusion comprises the following steps: Step S1, collecting vibration signals, sound signals, current signals and temperature signals of a cutter in different states through different sensors, and providing an original data base for subsequent analysis; Step S2, respectively preprocessing the physical characteristic differences of the signals of different modes; S3, generating a time-frequency diagram through continuous wavelet transformation aiming at vibration and sound high-frequency dynamic signals, and reserving time domain-frequency domain joint characteristics; s4, performing feature extraction by adopting a 2D convolutional neural network (2D CNN) aiming at the time-frequency diagram generated in the S3; S5, aiming at low-frequency slow-varying signals of current and temperature, constructing a time sequence-feature matrix by a sliding window method, calculating statistics such as mean value, variance, rising rate and the like in a window, and converting one-dimensional time sequence data into structural two-dimensional features; s6, adopting a two-dimensional characteristic of current and temperature to adopt a two-way long-short-term memory network, namely, two-way LSTM modeling front-rear time sequence dependence, introducing a focusing key wear stage trend mutation point of an attention mechanism, and extracting trend characteristics reflecting load accumulation change; s7, introducing a cross-mode attention mechanism, respectively distributing dynamic weights for the high-frequency features and the low-frequency features, an