Part machining quality monitoring method based on multi-sensor fusion

Abstract

The invention discloses a part machining quality monitoring method based on multi-sensor fusion, which comprises the following steps: collecting vibration, acoustic emission, cutting force and spindle current signals, and forming a standardized sequence through alignment and normalization; performing time-frequency transformation on the sequence to obtain a time-frequency spectrum; applying a combined Teager-Kaiser energy operator on the time-frequency spectrum, and extracting a first type of features; calculating an energy operator response on the sparse dictionary coefficient, and extracting a second type of features; calculating an energy cross-correlation spectrum for the cross-channel signal, and extracting a third type of features; extremum detection is carried out on the three types of features, an energy peak column is generated, statistics are calculated, and graph embedding features are obtained in combination with collaborative energy indexes and graph Laplacian constraints; and inputting the statistical characteristics, the collaborative energy index and the graph embedding characteristics into a quality evaluation model, and outputting roughness estimation, an abnormal position and a quality label. According to the invention, multi-channel fusion monitoring is realized, and the anomaly detection and quality evaluation precision is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a method for monitoring the processing quality of parts based on multi-sensor fusion. Background Technology

[0002] With the continuous development of intelligent manufacturing and industrial processing, real-time monitoring of part machining quality has become a crucial aspect of improving product performance and reducing production costs. Traditional machining quality monitoring methods mainly rely on manual inspection or single sensor signal analysis, such as manual surface roughness detection, single-channel vibration signal analysis, and spindle current monitoring to determine the machining status. These methods suffer from insufficient real-time performance, reliance on experience for monitoring results, and limited information dimensions, making it difficult to meet the demands for high-precision quality monitoring in complex machining environments. With the widespread adoption of CNC machine tools and automated machining, single signals can no longer fully reflect the dynamic changes during the machining process, making multi-source information fusion an important direction for solving this problem.

[0003] In recent years, multi-sensor monitoring has been increasingly applied in the field of machining. Examples include using vibration signals for tool wear identification, acoustic emission signals for crack or spalling detection, cutting force signals for cutting stability analysis, and spindle current signals for load estimation. However, these studies largely focus on feature extraction from single or a few channels, lacking a systematic multi-sensor fusion mechanism. Existing methods commonly include Fourier transform, wavelet analysis, and envelope demodulation, but these methods have limited ability to capture transient energy changes in non-stationary signals, making it difficult to accurately characterize sudden anomalies and complex multimodal relationships during machining. Furthermore, traditional statistical features often rely on low-order indices such as mean, variance, and power spectral density, failing to effectively utilize cross-channel correlations.

[0004] Regarding the study of energy characteristics, some scholars have proposed using the Teager-Kaiser energy operator for signal energy analysis. This operator can reflect transient energy in the time domain, but its scalability in multi-sensor environments is insufficient. Existing techniques are mostly limited to single-channel applications, lacking in-depth research on energy calculations for time-frequency domain signals, sparse coefficients, and cross-modal signals. Furthermore, energy operator calculations often remain at the point value level, failing to combine local extremum detection and peak statistics to establish a complete time-frequency evolution structure, nor to model the correlation relationships of energy peaks using graph structure analysis. This results in limited accuracy of existing methods for anomaly identification and quality status determination in complex processing environments.

[0005] In the research direction of multi-sensor fusion, existing technologies mostly employ simple data splicing or feature-level concatenation, lacking modeling mechanisms for the relationships between different signal channels. Cross-modal feature cross-correlation calculations have been attempted in some studies, but most remain at the signal level, failing to model and extend at the energy operator response level. Furthermore, existing multi-channel collaborative feature extraction methods are relatively simplistic, lacking systematic quantitative indicators for the synchronicity of multi-channel peak values, and failing to form a unified framework suitable for process quality monitoring.

[0006] Therefore, how to provide a part processing quality monitoring method based on multi-sensor fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a multi-sensor fusion-based method for monitoring the machining quality of parts. This invention fully utilizes the joint acquisition of vibration signals, acoustic emission signals, cutting force signals, and spindle current signals. It combines time-frequency domain coupled Teager-Kaiser energy operators, sparse dictionary learning-driven energy operators, cross-modal dual energy cross-correlation analysis, and peak-array modeling with graph Laplace constraints. The invention details the entire process of extracting energy features from multi-sensor signals, constructing a collaborative energy index, and inputting it into a quality assessment model to achieve machining quality monitoring. This method possesses advantages such as strong characterization capabilities, high anomaly detection accuracy, and refined machining quality assessment.

[0008] A method for monitoring the machining quality of parts based on multi-sensor fusion according to an embodiment of the present invention includes the following steps: The system collects vibration, acoustic emission, cutting force, and spindle current signals during the part machining process, performs time alignment and amplitude normalization, and forms a multi-sensor standardized sequence. The time-frequency transformation of the multi-sensor normalized sequence is performed to obtain the multi-sensor time spectrum; By applying the time-frequency domain joint Teager-Kaiser energy operator to the multi-sensor time spectrum, transient energy-frequency ridges are extracted to generate the first type of energy features. Sparse dictionary learning is performed on the multi-sensor normalized sequence, and the Teager-Kaiser energy operator response is calculated on the atomic coefficient set to generate the second type of energy feature; For multi-sensor normalized sequences, cross-modal dual Teager–Kaiser energy operator responses are calculated by pairing sensor channels, dual energy cross-correlation spectra are extracted, and third-type energy features are generated. Local extremum detection is performed on the energy features of the first, second and third categories. Energy peaks are extracted and their amplitude, interval and distribution statistics are calculated. The cooperative energy index is calculated based on the synchronicity of peaks in different channels. A time-weighted graph is constructed based on the energy peaks and graph Laplace constraints are applied to obtain the statistical features of energy peaks and graph-embedded energy features. By fusing statistical features of energy peaks, co-energy indices, and graph-embedded energy features, and inputting them into a quality assessment model, the output includes estimates of surface roughness, location of anomalies, and quality status labels for the machined parts.

[0009] Optionally, the formation of the multi-sensor normalization sequence specifically includes: Vibration signals acquired by an accelerometer, acoustic emission signals acquired by an acoustic emission sensor, cutting force signals acquired by a power sensor, and spindle current signals acquired by a current sensor are collected. A unified time reference is set for vibration signals, acoustic emission signals, cutting force signals and spindle current signals, and time alignment is completed according to the machine tool encoder pulse sequence; The vibration signal, acoustic emission signal, cutting force signal and spindle current signal are resampled under a unified time reference to obtain a data sequence with the same sampling interval; Amplitude normalization is performed on the resampled vibration signal, acoustic emission signal, cutting force signal, and spindle current signal. The vibration signal, acoustic emission signal, cutting force signal and spindle current signal are combined after time alignment, resampling and amplitude normalization to form a multi-sensor standardized sequence set.

[0010] Optionally, obtaining the multi-sensor time spectrum specifically includes: The multi-sensor standardized sequence is divided into continuous data segments according to a fixed-length time window; Perform a Fourier transform within each data segment to obtain the frequency distribution results at the corresponding time position; The frequency distribution results of all time periods are combined in chronological order to form a frequency spectrum that changes over time. Time spectrum diagrams are generated separately for different sensing channels, and the time spectrum diagrams of vibration channel, acoustic emission channel, cutting force channel and spindle current channel are combined to form a multi-sensor time spectrum data set.

[0011] Optionally, the processing of the time-frequency domain combined with the Teager-Kaiser energy operator specifically includes: In the multi-sensor time spectrum, a time index and a frequency index are set for each sensing channel to form a complex spectral coefficient matrix containing three-dimensional information of channel, time and frequency. By applying the energy operator to the complex spectral coefficient sequence along the time index at a fixed frequency index, the transient energy distribution that varies with time is obtained. By applying the energy operator to the complex spectral coefficient sequence along the frequency index at a fixed time index, the transient energy distribution that varies with frequency is obtained. The time-frequency energy distribution is combined with the frequency-frequency energy distribution to form a time-frequency energy matrix: ; in, In the channel Time Index Frequency Index The energy operator response under the given conditions Indicates channel The complex spectral coefficients, Represents the modulus of a complex number. and These represent indices at adjacent times. and Complex spectral coefficients at the location, To perform the operation of taking the real part, For complex conjugates, this formula is derived from the extended form of the discrete Teager–Kaiser energy operator over the time-spectral coefficients; Local extremum detection is performed in the time-frequency energy matrix to generate a set of energy peaks, which are then connected according to the frequency deviation threshold under adjacent time indices to form transient energy-frequency ridges. The amplitude sequence, frequency sequence, and time sequence of the transient energy-frequency ridge are recorded as first-class energy features, and combined across all channels to form a set of first-class energy features.

[0012] Optionally, the process of performing sparse dictionary learning combined with the Teager-Kaiser energy operator on the multi-sensor normalized sequence specifically includes: The multi-sensor standardized sequence is divided into signal segments, and the signal segments are used as training samples; The dictionary matrix is ​​obtained by iteratively updating the training samples through the sparse dictionary learning process. The dictionary matrix consists of a preset number of dictionary atoms, and each dictionary atom is used to represent the local pattern of the signal segment. The trained dictionary matrix is ​​used to perform sparse decomposition on the signal segment to generate a sparse coefficient vector. Each element in the sparse coefficient vector corresponds to the coefficient of the signal segment in the dictionary atom. The Teager–Kaiser energy operator response is calculated on the sparse coefficient vector by taking the square of the coefficient at the current index position and subtracting the product of the coefficients at the previous and next index positions. Arrange the energy operator responses of the sparse coefficient vectors sequentially to form a sparse domain energy sequence. The sparse domain energy sequences of all signal segments are summarized to generate a second type of energy feature, which is then combined in the vibration channel, acoustic emission channel, cutting force channel, and spindle current channel to form a set of second type of energy features.

[0013] Optionally, the processing of the cross-modal dual Teager–Kaiser energy operator specifically includes: The vibration channel is paired with the acoustic emission channel, and the cutting force channel is paired with the spindle current channel to form a cross-modal signal pair; Within each cross-modal signal pair, the Teager-Kaiser energy operator response is calculated for the normalized sequence of each channel by taking the square of the current sample at each time index and subtracting the product of the previous time index sample and the next time index sample to obtain the energy sequence. Construct a dual energy cross-correlation function on the energy sequence: ; in, Indicates channel With channel In the delay index The dual energy cross-correlation value under the following conditions Indicates channel In time index The energy sequence value below, Indicates channel In time index The energy sequence value below, The data is taken from the vibration channel, acoustic emission channel, cutting force channel, and spindle current channel. For time indexing, For time delay indexing, the form of the cross-correlation function is derived from the standard definition of cross-correlation in discrete-time signal processing; Within the time delay index range Perform a scan to generate a dual energy cross-correlation spectrum; The amplitude sequence, time delay sequence, and channel index are extracted from the dual energy cross-correlation spectrum and recorded as the third type of energy feature. The third type of energy feature set is then combined within the vibration channel, acoustic emission channel, cutting force channel, and spindle current channel.

[0014] Optionally, the energy peak detection and co-energy index calculation specifically include: Local extremum point detection is performed on the first type of energy feature, the second type of energy feature, and the third type of energy feature respectively to generate energy peak sequences; The amplitude of each peak, the time interval between adjacent peaks, and the duration of the peak in the time index are recorded in the energy peak series to form an energy peak series parameter set; Statistical measures are calculated on the set of energy peak parameters, including the mean and variance of peak amplitude, the mean and variance of time intervals, and the concentration of the time distribution. Peak synchronicity is calculated among the energy peaks of the vibration channel, acoustic emission channel, cutting force channel, and spindle current channel, and the proportion of peaks appearing simultaneously in all channels under the same time index is used as the cooperative energy index. A time-weighted graph is constructed based on the distribution of energy peaks on the time index, and the edge weights of the time-weighted graph are determined by the time interval between peaks. Applying graph Laplacian constraints to the time-weighted graph yields graph embedding energy features; The statistical characteristics of energy peaks, the co-energy index, and the graph-embedded energy characteristics are output as input features for quality assessment.

[0015] Optionally, the processing of the quality assessment model specifically includes: The statistical characteristics of energy peaks, the co-energy index, and the graph-embedded energy characteristics are combined to form an input vector for quality assessment. In the input vector, the amplitude statistics include the average amplitude and amplitude variance, the interval statistics include the average interval and interval variance, the distribution statistics include the concentration of peaks on the time index, the co-energy index is the proportion of all channels that have peaks at the same time index, and the graph embedding energy feature is composed of the embedding vector generated by the time-weighted graph under the graph Laplace constraint. The input vector is fed into the quality assessment model, which includes a feature input layer, a feature mapping layer, and an output layer. The output layer generates the processing quality result. The processing quality results include surface roughness estimates, time index locations of anomalies, and quality status labels.

[0016] The beneficial effects of this invention are: This invention constructs a multi-sensor standardized sequence by introducing the joint acquisition of vibration, acoustic emission, cutting force, and spindle current signals during part machining, overcoming the limitation of existing methods that rely on a single signal and thus lack sufficient information. Based on this, it utilizes the time-frequency domain combined with the Teager-Kaiser energy operator to extract transient energy and frequency ridge features. Combined with the energy operator response under sparse dictionary learning and the cross-modal dual energy cross-correlation spectrum, it achieves comprehensive modeling of signal energy patterns at the time-frequency domain, sparse domain, and cross-channel levels, improving the ability to characterize the non-stationary characteristics and inter-channel coupling relationships of complex machining processes.

[0017] This invention further constructs an energy peak sequence through local extremum detection, recording amplitude, interval, and distribution characteristics, and introduces a cooperative energy index to quantify the synchronicity of different channels. A time-weighted graph is established based on the energy peak sequence, and graph-embedded energy features are extracted under graph Laplacian constraints, allowing the energy evolution law to be expressed globally. Finally, the statistical characteristics of the energy peak sequence, the cooperative energy index, and the graph-embedded energy features are input into a quality assessment model, which can output surface roughness estimates, abnormal time locations, and quality status labels, achieving refined monitoring of processing quality. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is a flowchart of a part processing quality monitoring method based on multi-sensor fusion proposed in this invention; Figure 2 This is a schematic diagram of multi-sensor signal acquisition and standardized processing for a part processing quality monitoring method based on multi-sensor fusion proposed in this invention. Figure 3 This is a schematic diagram of the input and output of a quality assessment model for a multi-sensor fusion-based part processing quality monitoring method proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1-3 A method for monitoring the machining quality of parts based on multi-sensor fusion includes the following steps: The system collects vibration, acoustic emission, cutting force, and spindle current signals during the part machining process, performs time alignment and amplitude normalization, and forms a multi-sensor standardized sequence. The time-frequency transformation of the multi-sensor normalized sequence is performed to obtain the multi-sensor time spectrum; By applying the time-frequency domain joint Teager-Kaiser energy operator to the multi-sensor time spectrum, transient energy-frequency ridges are extracted to generate the first type of energy features. Sparse dictionary learning is performed on the multi-sensor normalized sequence, and the Teager-Kaiser energy operator response is calculated on the atomic coefficient set to generate the second type of energy feature; For multi-sensor normalized sequences, cross-modal dual Teager–Kaiser energy operator responses are calculated by pairing sensor channels, dual energy cross-correlation spectra are extracted, and third-type energy features are generated. Local extremum detection is performed on the energy features of the first, second and third categories. Energy peaks are extracted and their amplitude, interval and distribution statistics are calculated. The cooperative energy index is calculated based on the synchronicity of peaks in different channels. A time-weighted graph is constructed based on the energy peaks and graph Laplace constraints are applied to obtain the statistical features of energy peaks and graph-embedded energy features. By fusing statistical features of energy peaks, co-energy indices, and graph-embedded energy features, and inputting them into a quality assessment model, the output includes estimates of surface roughness, location of anomalies, and quality status labels for the machined parts.

[0022] This invention constructs a multi-sensor standardized sequence by acquiring vibration, acoustic emission, cutting force, and spindle current signals, combined with time alignment and amplitude normalization processing. This method ensures consistency in both time and amplitude scales for data from different sensors, thereby eliminating deviations caused by differences in sampling rates or amplitude range mismatches. This unified data foundation effectively supports subsequent time-frequency analysis, energy feature extraction, and cross-channel correlation calculations, laying a reliable foundation for the fusion analysis of multimodal signals and enabling a comprehensive characterization of the dynamic features of complex machining processes.

[0023] In this embodiment, the formation of the multi-sensor normalization sequence specifically includes: Vibration signals acquired by an accelerometer, acoustic emission signals acquired by an acoustic emission sensor, cutting force signals acquired by a power sensor, and spindle current signals acquired by a current sensor are collected. A unified time reference is set for vibration signals, acoustic emission signals, cutting force signals and spindle current signals, and time alignment is completed according to the machine tool encoder pulse sequence; The vibration signal, acoustic emission signal, cutting force signal and spindle current signal are resampled under a unified time reference to obtain a data sequence with the same sampling interval; Amplitude normalization is performed on the resampled vibration signal, acoustic emission signal, cutting force signal, and spindle current signal. The vibration signal, acoustic emission signal, cutting force signal and spindle current signal are combined after time alignment, resampling and amplitude normalization to form a multi-sensor standardized sequence set.

[0024] This invention divides a multi-sensor standardized sequence into time windows and performs Fourier transform to obtain the time-frequency spectrum that varies with time. Then, it combines the spectra of vibration, acoustic emission, cutting force, and spindle current channels into a multi-sensor time-frequency spectrum data set. This method can reflect the energy variation of different signal channels with time and frequency within a unified framework, helping to capture dynamic characteristics such as tool wear, abnormal vibration, or sudden load changes during machining, and providing rich time-frequency information support for subsequent energy operator-based analysis.

[0025] In this embodiment, obtaining the multi-sensor time spectrum specifically includes: The multi-sensor standardized sequence is divided into continuous data segments according to a fixed-length time window; Perform a Fourier transform within each data segment to obtain the frequency distribution results at the corresponding time position; The frequency distribution results of all time periods are combined in chronological order to form a frequency spectrum that changes over time. Time spectrum diagrams are generated separately for different sensing channels, and the time spectrum diagrams of vibration channel, acoustic emission channel, cutting force channel and spindle current channel are combined to form a multi-sensor time spectrum data set.

[0026] This invention applies a combined time-frequency domain Teager-Kaiser energy operator to the multi-sensor time spectrum, calculating transient energy distributions along both the time and frequency indices, and then combining them to form a time-frequency energy matrix. This method can simultaneously characterize the energy evolution of a signal in both the time and frequency directions, avoiding the problem of traditional one-dimensional energy analysis failing to capture time-frequency coupling effects. By detecting local extrema in the time-frequency energy matrix and connecting them to form transient energy-frequency ridges, abnormal fluctuations in non-stationary signals can be accurately identified, achieving high-precision extraction of the first type of energy feature.

[0027] In this embodiment, the processing of the time-frequency domain combined with the Teager-Kaiser energy operator specifically includes: In the multi-sensor time spectrum, a time index and a frequency index are set for each sensing channel to form a complex spectral coefficient matrix containing three-dimensional information of channel, time and frequency. By applying the energy operator to the complex spectral coefficient sequence along the time index at a fixed frequency index, the transient energy distribution that varies with time is obtained. By applying the energy operator to the complex spectral coefficient sequence along the frequency index at a fixed time index, the transient energy distribution that varies with frequency is obtained. The time-frequency energy distribution is combined with the frequency-frequency energy distribution to form a time-frequency energy matrix: ; in, In the channel Time Index Frequency Index The energy operator response under the given conditions Indicates channel The complex spectral coefficients, Represents the modulus of a complex number. and These represent indices at adjacent times. and Complex spectral coefficients at the location, To perform the operation of taking the real part, For complex conjugates, this formula is derived from the extended form of the discrete Teager–Kaiser energy operator over the time-spectral coefficients; Local extremum detection is performed in the time-frequency energy matrix to generate a set of energy peaks, which are then connected according to the frequency deviation threshold under adjacent time indices to form transient energy-frequency ridges. The amplitude sequence, frequency sequence, and time sequence of the transient energy-frequency ridge are recorded as first-class energy features, and combined across all channels to form a set of first-class energy features.

[0028] This invention obtains a dictionary matrix composed of dictionary atoms by performing sparse dictionary learning on multi-sensor normalized sequences, and then calculates the Teager-Kaiser energy operator response on the sparse coefficient set to generate a second type of energy feature. This method utilizes sparse representation to highlight key structural patterns in the signal, and combined with energy operator analysis, it can mine local energy characteristics in the sparse domain. Compared with directly extracting features from the original signal, sparse domain features have stronger noise robustness and higher expressive power, and can more sensitively reflect complex states such as tool wear and cutting instability.

[0029] In this embodiment, the process of performing sparse dictionary learning combined with the Teager-Kaiser energy operator on the multi-sensor normalized sequence specifically includes: The multi-sensor standardized sequence is divided into signal segments, and the signal segments are used as training samples; The dictionary matrix is ​​obtained by iteratively updating the training samples through the sparse dictionary learning process. The dictionary matrix consists of a preset number of dictionary atoms, and each dictionary atom is used to represent the local pattern of the signal segment. The trained dictionary matrix is ​​used to perform sparse decomposition on the signal segment to generate a sparse coefficient vector. Each element in the sparse coefficient vector corresponds to the coefficient of the signal segment in the dictionary atom. The Teager–Kaiser energy operator response is calculated on the sparse coefficient vector by taking the square of the coefficient at the current index position and subtracting the product of the coefficients at the previous and next index positions. Arrange the energy operator responses of the sparse coefficient vectors sequentially to form a sparse domain energy sequence. The sparse domain energy sequences of all signal segments are summarized to generate a second type of energy feature, which is then combined in the vibration channel, acoustic emission channel, cutting force channel, and spindle current channel to form a set of second type of energy features.

[0030] This invention pairs vibration with acoustic emission channels and cutting force with spindle current channels respectively, calculates cross-modal dual energy cross-correlation functions and generates cross-correlation spectra, and extracts third-type energy features. This method not only considers the energy changes of single-channel signals but also characterizes the dynamic correlation between different physical quantities. By analyzing the time delay and energy coupling modes between different channels, it can effectively reveal the cross-channel linkage characteristics in phenomena such as tool failure and machining anomalies, achieving in-depth utilization of multi-sensor collaborative information and enhancing the comprehensive discrimination capability of monitoring methods.

[0031] In this embodiment, the processing of the cross-modal dual Teager–Kaiser energy operator specifically includes: The vibration channel is paired with the acoustic emission channel, and the cutting force channel is paired with the spindle current channel to form a cross-modal signal pair; Within each cross-modal signal pair, the Teager-Kaiser energy operator response is calculated for the normalized sequence of each channel by taking the square of the current sample at each time index and subtracting the product of the previous time index sample and the next time index sample to obtain the energy sequence. Construct a dual energy cross-correlation function on the energy sequence: ; in, Indicates channel With channel In the delay index The dual energy cross-correlation value under the following conditions Indicates channel In time index The energy sequence value below, Indicates channel In time index The energy sequence value below, The data is taken from the vibration channel, acoustic emission channel, cutting force channel, and spindle current channel. For time indexing, For time delay indexing, the form of the cross-correlation function is derived from the standard definition of cross-correlation in discrete-time signal processing; Within the time delay index range Perform a scan to generate a dual energy cross-correlation spectrum; The amplitude sequence, time delay sequence, and channel index are extracted from the dual energy cross-correlation spectrum and recorded as the third type of energy feature. The third type of energy feature set is then combined within the vibration channel, acoustic emission channel, cutting force channel, and spindle current channel.

[0032] This invention performs local extremum detection on three types of energy features, extracts energy peak sequences, and calculates amplitude, interval, and distribution statistics. Simultaneously, it constructs a cooperative energy index based on the synchronicity of peak values ​​across different channels and combines time-weighted graphs and graph Laplace constraints to obtain graph-embedded energy features. This method achieves unified modeling of local energy features and global graph structure features, quantifying the synchronicity between multiple channels and revealing the overall laws of energy evolution through graph structure, providing a more refined feature representation for anomaly identification in the processing process.

[0033] In this embodiment, the energy peak detection and co-energy index calculation specifically include: Local extremum point detection is performed on the first type of energy feature, the second type of energy feature, and the third type of energy feature respectively to generate energy peak sequences; The amplitude of each peak, the time interval between adjacent peaks, and the duration of the peak in the time index are recorded in the energy peak series to form an energy peak series parameter set; Statistical measures are calculated on the set of energy peak parameters, including the mean and variance of peak amplitude, the mean and variance of time intervals, and the concentration of the time distribution. Peak synchronicity is calculated among the energy peaks of the vibration channel, acoustic emission channel, cutting force channel, and spindle current channel, and the proportion of peaks appearing simultaneously in all channels under the same time index is used as the cooperative energy index. A time-weighted graph is constructed based on the distribution of energy peaks on the time index, and the edge weights of the time-weighted graph are determined by the time interval between peaks. Applying graph Laplacian constraints to the time-weighted graph yields graph embedding energy features; The statistical characteristics of energy peaks, the co-energy index, and the graph-embedded energy characteristics are output as input features for quality assessment.

[0034] This invention establishes a unified computational framework from features to results by combining statistical characteristics of energy peaks, co-energy indices, and graph-embedded energy features into a quality assessment model. This method fully utilizes the local statistical information, cross-channel synchronization, and global graph structure features of multi-channel signals within the model, enhancing its ability to represent complex processing states. The output processing quality results include surface roughness estimation, anomaly occurrence time and location, and quality status labels, meeting the needs for real-time and refined quality monitoring in intelligent manufacturing environments.

[0035] In this embodiment, the processing of the quality assessment model specifically includes: The statistical characteristics of energy peaks, the co-energy index, and the graph-embedded energy characteristics are combined to form an input vector for quality assessment. In the input vector, the amplitude statistics include the average amplitude and amplitude variance, the interval statistics include the average interval and interval variance, the distribution statistics include the concentration of peaks on the time index, the co-energy index is the proportion of all channels that have peaks at the same time index, and the graph embedding energy feature is composed of the embedding vector generated by the time-weighted graph under the graph Laplace constraint. The input vector is fed into the quality assessment model, which includes a feature input layer, a feature mapping layer, and an output layer. The output layer generates the processing quality result. The processing quality results include surface roughness estimates, time index locations of anomalies, and quality status labels.

[0036] This invention constructs a quality assessment model comprising a feature input layer, a feature mapping layer, and an output layer, ensuring that statistical features, co-exponential indices, and graph embedding features in the input vector are fully processed. The model output includes surface roughness estimates, time index locations of anomalies, and quality status labels, achieving a quantitative assessment of processing quality. This method can provide stable prediction results in complex signal environments, effectively overcoming the shortcomings of traditional methods in terms of accuracy and robustness, and providing reliable support for the dynamic monitoring of processing quality.

[0037] Example 1: To verify the feasibility of this invention in practice, it was applied to a real-world production scenario for monitoring the quality of machined parts. The machined objects were metal parts, including shaft parts and disc parts. The machining quality was monitored in real time during turning and milling. The experimental environment was equipped with vibration sensors, acoustic emission sensors, dynamic sensors, and current sensors to collect vibration signals, acoustic emission signals, cutting force signals, and spindle current signals during the machining process. All sensor signals were time-aligned using machine tool encoder pulses, resampled, and normalized under a unified time reference to form a multi-sensor standardized sequence, providing a consistent data foundation for subsequent feature extraction.

[0038] In this scenario, traditional machining quality monitoring mainly relies on single-channel signals, such as using only spindle current to detect machining load, or relying only on vibration signals to judge tool wear. However, this method has obvious shortcomings: a single signal cannot reflect the multi-source coupling effect in complex machining processes, resulting in a low anomaly detection rate and easy misjudgment. This invention uses multi-sensor fusion to unify the physical quantities reflected by different sensors under a standardized framework, avoiding information loss.

[0039] In the experiment, the original vibration signal frequency range was 0–10 kHz, and the sampling rate was set to 20 kHz; the acoustic emission signal frequency range was 50–400 kHz, and the sampling rate was 1 MHz; the cutting force signal frequency range was 0–5 kHz, and the sampling rate was 10 kHz; the spindle current signal frequency range was 0–500 Hz, and the sampling rate was 2 kHz. After normalization, the signals of each channel were divided into 0.5-second time windows as analysis units. After Fourier transform of these time windows, the multi-sensor time spectrum was obtained. Then, the time-frequency domain joint Teager-Kaiser energy operator was applied to the time spectrum to extract the transient energy-frequency ridge feature. This feature can accurately characterize the sudden high-frequency energy fluctuations that occur when the tool enters the cutting state or when slight chipping occurs.

[0040] Meanwhile, sparse dictionary learning is performed on the multi-sensor normalized sequence, and the signal segments are represented as sparse combinations of dictionary atoms. Using 500 signal segments as training samples, a dictionary matrix containing 256 dictionary atoms is learned. The Teager-Kaiser energy operator response is calculated on the dictionary coefficients to form a second type of energy feature. Compared with the traditional power spectrum feature, this feature improves the accuracy of identifying slight tool wear by about 12%.

[0041] Furthermore, by pairing vibration with acoustic emission signals and cutting force with spindle current signals respectively, the cross-modal dual energy cross-correlation spectrum was calculated to obtain the third type of energy characteristics. Experiments show that when machining anomalies (such as tool flank wear) occur, the peak value of the cross-correlation between vibration and acoustic emission energy sequences increases significantly, and the delay time is shortened from 2.5 milliseconds to 0.5 milliseconds, indicating that the energy synchronization between channels is enhanced, which provides a clear basis for anomaly identification.

[0042] Local extremum detection was performed on the energy features of the first, second, and third categories to obtain energy peak sequences. The peak amplitude, interval, and duration were recorded in the peak sequences. The cooperative energy index was calculated based on the synchronicity of the peaks in different channels. The results showed that the cooperative energy index was about 0.35 under normal machining conditions, while it rose to 0.72 when tool abnormalities occurred, significantly distinguishing the two states. At the same time, by constructing a time-weighted graph and applying graph Laplacian constraints, graph embedding energy features were further obtained to capture the global laws of energy evolution.

[0043] Finally, the statistical characteristics of energy peaks, the co-energy index, and the graph-embedded energy characteristics were input into the quality assessment model. This model adopts a three-layer structure: the input layer receives the fused features, the mapping layer completes the feature dimension transformation, and the output layer generates the processing quality results, including surface roughness estimates, abnormal time index locations, and quality status labels. The comparative experiment used the traditional monitoring method based on single-channel vibration signals as a reference. The comparison results are shown in Table 1. Table 1. Performance Comparison of Different Monitoring Methods in Parts Machining Quality Assessment

[0044] As can be seen from the table, the method of this invention significantly outperforms traditional methods in terms of surface roughness prediction, anomaly detection accuracy, anomaly localization accuracy, and quality status recognition rate. Specifically, the surface roughness prediction error is reduced to 0.29 μm, the anomaly detection accuracy reaches 94%, the average anomaly localization error is only 3.6 milliseconds, and the overall quality status recognition accuracy reaches 95%. These results verify the feasibility and superiority of this invention in complex processing environments, effectively solving the problems of insufficient information utilization, inadequate energy feature expression, and lack of cross-channel relationship modeling in existing technologies.

[0045] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring the machining quality of parts based on multi-sensor fusion, characterized in that, Includes the following steps: The system collects vibration, acoustic emission, cutting force, and spindle current signals during the part machining process, performs time alignment and amplitude normalization, and forms a multi-sensor standardized sequence. The time-frequency transformation of the multi-sensor normalized sequence is performed to obtain the multi-sensor time spectrum; By applying the time-frequency domain joint Teager-Kaiser energy operator to the multi-sensor time spectrum, transient energy-frequency ridges are extracted to generate the first type of energy features. Sparse dictionary learning is performed on the multi-sensor normalized sequence, and the Teager-Kaiser energy operator response is calculated on the atomic coefficient set to generate the second type of energy feature; For multi-sensor normalized sequences, cross-modal dual Teager–Kaiser energy operator responses are calculated by pairing sensor channels, dual energy cross-correlation spectra are extracted, and third-type energy features are generated. Local extremum detection is performed on the energy features of the first, second and third categories. Energy peaks are extracted and their amplitude, interval and distribution statistics are calculated. The cooperative energy index is calculated based on the synchronicity of peaks in different channels. A time-weighted graph is constructed based on the energy peaks and graph Laplace constraints are applied to obtain the statistical features of energy peaks and graph-embedded energy features. By fusing statistical features of energy peaks, co-energy indices, and graph-embedded energy features, and inputting them into a quality assessment model, the output includes estimates of surface roughness, location of anomalies, and quality status labels for the machined parts.

2. The method for monitoring the machining quality of parts based on multi-sensor fusion according to claim 1, characterized in that, The formation of the multi-sensor normalization sequence specifically includes: Vibration signals acquired by an accelerometer, acoustic emission signals acquired by an acoustic emission sensor, cutting force signals acquired by a power sensor, and spindle current signals acquired by a current sensor are collected. A unified time reference is set for vibration signals, acoustic emission signals, cutting force signals and spindle current signals, and time alignment is completed according to the machine tool encoder pulse sequence; The vibration signal, acoustic emission signal, cutting force signal and spindle current signal are resampled under a unified time reference to obtain a data sequence with the same sampling interval; Amplitude normalization is performed on the resampled vibration signal, acoustic emission signal, cutting force signal, and spindle current signal. The vibration signal, acoustic emission signal, cutting force signal and spindle current signal are combined after time alignment, resampling and amplitude normalization to form a multi-sensor standardized sequence set.

3. The method for monitoring the machining quality of parts based on multi-sensor fusion according to claim 1, characterized in that, The acquisition of the multi-sensor time spectrum specifically includes: The multi-sensor standardized sequence is divided into continuous data segments according to a fixed-length time window; Perform a Fourier transform within each data segment to obtain the frequency distribution results at the corresponding time position; The frequency distribution results of all time periods are combined in chronological order to form a frequency spectrum that changes over time. Time spectrum diagrams are generated separately for different sensing channels, and the time spectrum diagrams of vibration channel, acoustic emission channel, cutting force channel and spindle current channel are combined to form a multi-sensor time spectrum data set.

4. The method for monitoring the machining quality of parts based on multi-sensor fusion according to claim 1, characterized in that, The processing of the time-frequency domain joint Teager-Kaiser energy operator specifically includes: In the multi-sensor time spectrum, a time index and a frequency index are set for each sensing channel to form a complex spectral coefficient matrix containing three-dimensional information of channel, time and frequency. By applying the energy operator to the complex spectral coefficient sequence along the time index at a fixed frequency index, the transient energy distribution that varies with time is obtained. By applying the energy operator to the complex spectral coefficient sequence along the frequency index at a fixed time index, the transient energy distribution that varies with frequency is obtained. The energy distribution in the time direction and the energy distribution in the frequency direction are combined to form a time-frequency energy matrix; Local extremum detection is performed in the time-frequency energy matrix to generate a set of energy peaks, which are then connected according to the frequency deviation threshold under adjacent time indices to form transient energy-frequency ridges. The amplitude sequence, frequency sequence, and time sequence of the transient energy-frequency ridge are recorded as first-class energy features, and combined across all channels to form a set of first-class energy features.

5. The method for monitoring the machining quality of parts based on multi-sensor fusion according to claim 1, characterized in that, The specific processing of the multi-sensor normalized sequence by performing sparse dictionary learning combined with the Teager-Kaiser energy operator includes: The multi-sensor standardized sequence is divided into signal segments, and the signal segments are used as training samples; The dictionary matrix is ​​obtained by iteratively updating the training samples through the sparse dictionary learning process. The dictionary matrix consists of a preset number of dictionary atoms, and each dictionary atom is used to represent the local pattern of the signal segment. The trained dictionary matrix is ​​used to perform sparse decomposition on the signal segment to generate a sparse coefficient vector. Each element in the sparse coefficient vector corresponds to the coefficient of the signal segment in the dictionary atom. The Teager–Kaiser energy operator response is calculated on the sparse coefficient vector by taking the square of the coefficient at the current index position and subtracting the product of the coefficients at the previous and next index positions. Arrange the energy operator responses of the sparse coefficient vectors sequentially to form a sparse domain energy sequence. The sparse domain energy sequences of all signal segments are summarized to generate a second type of energy feature, which is then combined in the vibration channel, acoustic emission channel, cutting force channel, and spindle current channel to form a set of second type of energy features.

6. The method for monitoring the machining quality of parts based on multi-sensor fusion according to claim 1, characterized in that, The processing of the cross-modal dual Teager–Kaiser energy operator specifically includes: The vibration channel is paired with the acoustic emission channel, and the cutting force channel is paired with the spindle current channel to form a cross-modal signal pair; Within each cross-modal signal pair, the Teager-Kaiser energy operator response is calculated for the normalized sequence of each channel by taking the square of the current sample at each time index and subtracting the product of the previous time index sample and the next time index sample to obtain the energy sequence. Construct dual energy cross-correlation functions on the energy sequence; Within the time delay index range Perform a scan to generate a dual energy cross-correlation spectrum; The amplitude sequence, time delay sequence, and channel index are extracted from the dual energy cross-correlation spectrum and recorded as the third type of energy feature. The third type of energy feature set is then combined within the vibration channel, acoustic emission channel, cutting force channel, and spindle current channel.

7. The method for monitoring the machining quality of parts based on multi-sensor fusion according to claim 1, characterized in that, The energy peak detection and synergistic energy index calculation specifically include: Local extremum point detection is performed on the first type of energy feature, the second type of energy feature, and the third type of energy feature respectively to generate energy peak sequences; The amplitude of each peak, the time interval between adjacent peaks, and the duration of the peak in the time index are recorded in the energy peak series to form an energy peak series parameter set; Statistical measures are calculated on the set of energy peak parameters, including the mean and variance of peak amplitude, the mean and variance of time intervals, and the concentration of the time distribution. Peak synchronicity is calculated among the energy peaks of the vibration channel, acoustic emission channel, cutting force channel, and spindle current channel, and the proportion of peaks appearing simultaneously in all channels under the same time index is used as the cooperative energy index. A time-weighted graph is constructed based on the distribution of energy peaks on the time index, and the edge weights of the time-weighted graph are determined by the time interval between peaks. Applying graph Laplacian constraints to the time-weighted graph yields graph embedding energy features; The statistical characteristics of energy peaks, the co-energy index, and the graph-embedded energy characteristics are output as input features for quality assessment.

8. The method for monitoring the machining quality of parts based on multi-sensor fusion according to claim 1, characterized in that, The processing of the quality assessment model specifically includes: The statistical characteristics of energy peaks, the co-energy index, and the graph-embedded energy characteristics are combined to form an input vector for quality assessment. In the input vector, the amplitude statistics include the average amplitude and amplitude variance, the interval statistics include the average interval and interval variance, the distribution statistics include the concentration of peaks on the time index, the co-energy index is the proportion of all channels that have peaks at the same time index, and the graph embedding energy feature is composed of the embedding vector generated by the time-weighted graph under the graph Laplace constraint. The input vector is fed into the quality assessment model, which includes a feature input layer, a feature mapping layer, and an output layer. The output layer generates the processing quality result. The processing quality results include surface roughness estimates, time index locations of anomalies, and quality status labels.

Images