A sanding belt wear prediction method and system that fuses acoustic and acoustic emission signals
By fusing acoustic signals and acoustic emission signals, an LSTM neural network was used to establish a belt wear model, which allows for real-time monitoring and adjustment of process parameters. This solves the problem of real-time monitoring and prediction of belt wear in complex blade machining, and improves machining accuracy and quality.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- WUHAN UNIV
- Filing Date
- 2024-05-23
- Publication Date
- 2026-06-23
AI Technical Summary
Existing technologies cannot monitor and predict the wear state of the abrasive belt during the machining of complex blades in real time and accurately, which makes it difficult to guarantee machining accuracy and quality, and cannot deal with the damage caused by severe abrasive belt wear in a timely manner.
By fusing acoustic signals and acoustic emission signals, a nonlinear model of abrasive belt wear is established using an LSTM neural network. Combined with probabilistic statistical methods, the state is estimated, and the abrasive belt wear is monitored in real time and the process parameters are adjusted accordingly to prevent excessive tool wear.
It enables real-time and precise monitoring of abrasive belt wear during the machining of complex blades by robots, improving machining efficiency and quality, preventing damage caused by severe abrasive belt wear, and is highly adaptable with accurate prediction results.
Smart Images
Figure CN118673453B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of belt abrasion prediction technology, and more specifically, relates to a belt abrasion prediction method and system that integrates acoustic signals and acoustic emission signals. Background Technology
[0002] Blades, as a key component in many important mechanical devices, are widely found in large and complex components such as aircraft engines and wind turbines. Their intricate thin-walled structure makes their manufacturing process extremely difficult. Blade materials are primarily difficult to machine, such as titanium alloys and nickel-based superalloys. After forging into blanks, milling and polishing are generally required to complete the blade manufacturing process. Currently, blade polishing mainly involves manual polishing and CNC belt grinding. Manual polishing results in low production efficiency, harsh production environments, and difficulty in controlling blade shape accuracy and surface consistency. Multi-axis CNC grinding machines have relatively fixed processing modes, complex equipment, and are expensive, generally being specialized machine tools unsuitable for processing a wide range of products. Robotic belt grinding, as a new type of automated processing method, is gradually being applied to the precision polishing of complex blades due to its programmability, high flexibility, and strong adaptability. However, due to the complex structure and high material properties of blades, belt grinding, as a contact machining technology, presents significant challenges in precision machining. Uneven wear of the abrasive grains can severely impact blade machining quality. Therefore, online monitoring of the belt wear is often necessary to ensure machining accuracy and quality. This involves real-time monitoring of the wear state during the robotic belt grinding process to adjust and optimize machining parameters in real time, including grinding pressure, grinding speed, and robot feed rate. This extends the stable wear phase of the belt, improving production efficiency, ensuring blade machining accuracy and surface quality, and, to some extent, extending the service life of the belt. Therefore, real-time online monitoring plays a crucial role in the precision machining of blades.
[0003] In the prior art, patent document CN114406807A discloses a method for predicting abrasive belt wear based on machine learning and image processing. This method involves acquiring images of the abrasive belt, identifying regions of interest (ROIs) within the images, calculating multi-dimensional feature parameters, performing normalization, and inputting the normalized parameters into a material removal rate prediction model. Repeated operations yield the actual material removal rate, and this method can accurately determine the remaining grinding capacity of the abrasive belt. Additionally, patent document CN112643486A discloses a method for predicting abrasive belt wear during the grinding process of complex curved workpieces. This method determines the variation law of the abrasive belt wear coefficient with various parameters in the grinding process, establishes an elastic grinding contact model to obtain the shape and pressure distribution of the contact area during the abrasive belt wheel grinding process, constructs a contact area sweep body based on the change of the contact area over time in the local coordinate system, and proposes a prediction method for the non-uniform wear coefficient change of the abrasive belt using the sweep body.
[0004] However, the belt wear prediction method based on machine learning and image processing disclosed in patent document CN114406807A relies on capturing images of the belt surface. During blade machining, the belt rotates at high speed, which may prevent clear images of its shape from being captured during grinding, thus affecting the prediction results. This method only proposes using grinding capacity as an equivalent substitute for belt wear volume without specifying the conversion relationship between the two, and it cannot achieve real-time accurate wear prediction. CN112643486A discloses a belt wear prediction method during the grinding of complex curved surface workpieces, but its applicability is narrow and it cannot guarantee machining accuracy. Both of these methods fail to consider the actual machining environment of complex thin-walled blade-like components, cannot provide real-time and accurate prediction of belt wear, and cannot effectively address the situation caused by severe belt wear. Summary of the Invention
[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method and system for predicting abrasive belt wear by fusing acoustic signals and acoustic emission signals. This invention establishes a relationship model between the acoustic signals and acoustic emission signals generated during robotic grinding by fusing the characteristics of these signals. This allows for real-time and accurate monitoring of the abrasive belt wear state during the robotic machining of complex blades, and timely feedback adjustments to prevent damage to complex blades due to severe abrasive belt wear during robotic machining.
[0006] According to a first aspect of the present invention, a method for predicting belt wear by fusing acoustic signals and acoustic emission signals is provided, comprising the following steps:
[0007] S100, installs an acoustic signal sensor and an acoustic emission sensor to collect signals, performs time-domain analysis and frequency-domain analysis on the two signals, and extracts the time-domain features and frequency-domain features of the signals respectively;
[0008] S200, using an LSTM neural network to fuse the time-domain and frequency-domain features of the two signals, establishes a nonlinear model of belt wear during the blade robot grinding process;
[0009] S300, based on the aforementioned nonlinear model, uses state estimation techniques in probabilistic statistics to predict the probability of stable wear and severe wear occurring in the sand belt at the next moment;
[0010] S400, based on the probability of stable and severe wear of the abrasive belt at the next moment, feeds back to adjust and control the control process parameters during the grinding process to prevent excessive tool wear from causing quality degradation phenomena such as burns on the blades.
[0011] Furthermore, in step S100,
[0012] The time-domain analysis method includes the time-domain analysis method;
[0013] The time-domain features include: average amplitude, amplitude variance, peak amplitude, and kurtosis coefficient;
[0014] The frequency domain analysis method includes: continuous wavelet transform (CWT); and obtaining the characteristics of acoustic signals and acoustic emission signals generated during belt grinding by a blade robot at different scales and times by sliding wavelet basis functions in the time and scale domains.
[0015] The frequency domain features include: wavelet energy, wavelet entropy, wavelet peak value, and wavelet frequency band energy ratio for each frequency band.
[0016] Specifically, the continuous wavelet transform method includes:
[0017] The initial acoustic signal and the acoustic emission signal are convolved at different scales and locations using a family of wavelet basis functions to obtain a time-scale representation of the signal. The mathematical model is as follows:
[0018]
[0019] Where a is the scale factor, b is the translation factor, and ψ * (t) is the complex conjugate of the mother wavelet function;
[0020] When performing wavelet transform, it is necessary to first define a, b, and ψ(t); and then repeatedly convolve the parameters at different scales and translations to obtain wavelet transform coefficients at different frequencies and time scales.
[0021] After wavelet transform processing, the acoustic signals and acoustic emission signals generated during the belt grinding process of the blade robot can be divided into three characteristic frequency bands: low frequency band, mid frequency band, and high frequency band.
[0022] Furthermore, in step S200,
[0023] The Long Short-Term Memory (LSTM) network method is used to fuse the time-frequency domain features of acoustic signals and acoustic emission signals during the grinding process of the blade robot. LSTM is used to process long sequences and capture the long-term dependence between the fused time-frequency domain features and the abrasive belt wear state, thereby fitting the nonlinear relationship between the two and establishing a nonlinear model between the time-frequency domain features after signal fusion and the abrasive belt wear during the complex blade robot processing.
[0024] Specifically, in the LSTM method,
[0025] LSTM includes a forget gate, an input gate, an output gate, and a cell state; the LSTM signal update steps include:
[0026] Step 1: Calculate the forget gate;
[0027] The forget gate determines which information from the time-frequency domain characteristics of the acoustic signal and acoustic emission signal in the previous unit state to discard. It receives the input x at the current time step. t The hidden state h of the previous time step t-1 It outputs a numerical vector between 0 and 1, representing the degree to which information in the cell state should be preserved.
[0028] f t =σ(W f ·[h t-1 ,x t ]+b f )
[0029] Among them, f t W indicates the degree to which information about the cell state should be preserved. f and b f These are the weight matrix and bias term of the forget gate, respectively, where σ is the sigmoid function, [h t-1 ,x t The symbol ] represents concatenating the hidden state of the previous time step and the input of the current time step into a single vector;
[0030] Step 2: Input gate calculation;
[0031] The input gate determines which new acoustic signals and their time-frequency domain characteristics will be updated in the cell state, and it also receives the input x at the current time step. t The hidden state h of the previous time step t-1 It outputs a numerical vector between 0 and 1, representing the degree to which the cell state should be updated:
[0032] i t =σ(Wi ·[h t-1 ,x t ]+b i )
[0033] Among them, i t W represents a numerical vector whose values are all between 0 and 1. i Let b represent the weight matrix of the input gate. i This represents the bias term of the input gate;
[0034] Step 3: Update candidate values;
[0035] The update candidate value indicates how much the cell state should be updated based on the input at the current time step; using the input x at the current time step t The hidden state h of the previous time step t-1 The calculation yielded:
[0036]
[0037] Among them, W c With b c These are the weight matrix and bias term of the cell state, respectively. Candidate values representing cell states, where tanh is the hyperbolic tangent function;
[0038] Step 4: Update cell status;
[0039] By combining the decisions of the forget gate and the input gate, the cell state is updated, information is passed between different time steps, and long-term dependencies are recorded. Based on the forget gate, the input gate, and the update candidate values, the cell state C at the current time step can be calculated. t ,as follows:
[0040]
[0041] Step 5: Output gate calculation;
[0042] The output gate determines the hidden state h at the current time step. t What information should it include, and what input x should it receive at the current time step? t The hidden state h of the previous time step t-1 And based on the cell state C at the current time step t Calculate the degree of opening of the output gate and output the hidden state at the current time step;
[0043] o t =σ(W o ·[h t-1 ,x t ]+b o )
[0044] ht =o t ·tanh(C t )
[0045] Among them, W f W i W C W o These are the weight matrices for the forget gate, input gate, candidate update gate, and output gate, respectively. f b i b c b o σ is the corresponding bias term, sigmoid function, and tanh is hyperbolic tangent function.
[0046] Furthermore, in step S300,
[0047] Based on the aforementioned nonlinear model, the acoustic signals and acoustic emission signals generated during grinding are collected and analyzed to set the initial processing state of the blade robot. During the processing, the state estimation technique in the probabilistic statistical method is used to predict the wear state of the abrasive belt and the probability of severe wear occurring at the next moment.
[0048] The initial blade robot processing state includes: for the processing of complex blades, grinding a brand-new abrasive belt to quickly bring it into a stable wear state; for the monitoring of the blade robot abrasive belt grinding process, it is assumed that the abrasive belt has already entered a stable wear state at the beginning.
[0049] The wear state of the abrasive belt is divided into three categories: initial wear state, stable wear state, and severe wear state. Further, in step S400,
[0050] As the robotic machining of complex blades continues, different operations are performed on belt abrasion under different conditions; including:
[0051] When the abrasive belt wear is in an ideal stable wear state, processing continues and the cycle is entered.
[0052] When the abrasive belt shows signs of severe wear or when it is predicted that the grinding belt will enter an unstable grinding state or a severe wear state in the next moment, it is necessary to adjust and optimize the process parameters of the blade robot grinding process in real time based on the parameter combination of the process database in order to extend the stable wear stage.
[0053] When the abrasive belt wear reaches a critical value and it no longer has grinding capacity, it is necessary to replace the abrasive belt in time to complete the precision grinding of the blades. This is to prevent over-grinding or burning of thin-walled blades, which could lead to blade scrapping and ensure the processing quality of the blades.
[0054] According to a second aspect of the present invention, the present invention also provides a belt abrasion prediction system that integrates acoustic signals and acoustic emission signals, comprising: a robot system module, a grinding and polishing machine system module, a signal processing system module, and a feedback control system module;
[0055] The robot system module is used to mount force sensors and grinding force control systems on the robot, which clamps the complex thin-walled blades to be processed by the robot end gripper, and sets up the robot control system to perform overall control and execution.
[0056] The grinding and polishing machine system module is used to arrange the contact wheel and sanding belt. The rotation of the contact wheel connected to it is controlled by the rotation of the motor, which in turn drives the sanding belt above to rotate and controls and adjusts the speed of the grinding and polishing machine control system.
[0057] The signal processing system module is used to acquire acoustic signals and acoustic emission signals during the grinding process, and to perform time-frequency domain analysis on the two types of signals to extract signal features;
[0058] The feedback control system module is used to predict the wear state of the abrasive belt based on the fusion of two signal features using an LSTM neural network, and to perform feedback adjustment and control of process parameters.
[0059] Preferably, in the signal processing system, the acoustic signal is acquired through a microphone; and the acoustic emission signal is acquired through an acoustic emission sensor.
[0060] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:
[0061] 1. The method of the present invention collects acoustic signals and acoustic emission signals during the precision grinding process of a complex blade robot, processes and extracts features from these two signals, and then uses a signal fusion method to effectively fuse these two signals to establish a relationship model between them and the wear of the abrasive belt. This allows for more effective real-time monitoring of the grinding state of the abrasive belt during the robot grinding process. By fusing these two different signals, it can accurately and effectively respond to unexpected situations in the precision grinding process of the robot and adapt to more complex situations.
[0062] 2. The method of the present invention, by real-time monitoring of the complex thin-walled blade robot grinding process, can optimize and adjust the process parameters in a timely manner when the abrasive belt enters different wear stages, thereby improving processing efficiency and processing quality. At the same time, when the abrasive belt is severely worn, it can be replaced in time to prevent burns and other phenomena that damage the workpiece, further ensuring the quality of precision grinding.
[0063] 3. The method of the present invention, by using an LSTM neural network to fuse the time-domain and frequency-domain features of two signals, establishes and predicts the nonlinear model of abrasive belt wear during the blade robot grinding process. This allows for a more comprehensive and accurate analysis of abrasive belt wear, stronger adaptability to different processing conditions and workpiece characteristics, higher robustness, and more accurate prediction results. Attached Figure Description
[0064] Figure 1 This is a flowchart of a method for predicting belt wear by fusing acoustic signals and acoustic emission signals according to an embodiment of the present invention;
[0065] Figure 2 This is a flowchart illustrating the online monitoring and real-time parameter optimization of abrasive belt wear during the machining of complex blades using a robot, as described in this embodiment of the invention.
[0066] Figure 3 This is a flowchart of the time-frequency domain analysis and feature extraction of acoustic signals and acoustic emission signals according to an embodiment of the present invention;
[0067] Figure 4 This is a flowchart of the sand belt wear modeling and state prediction based on the fusion of acoustic signal and acoustic emission signal features according to an embodiment of the present invention;
[0068] Figure 5 This is a structural diagram of an LSTM unit according to an embodiment of the present invention;
[0069] Figure 6 This is a graph showing the variation in belt grinding capability according to an embodiment of the present invention;
[0070] Figure 7 This is a network structure diagram for real-time monitoring and parameter optimization of complex blade robot belt grinding in an embodiment of the present invention;
[0071] Figure 8 This is a schematic diagram of the blade robot grinding equipment and signal acquisition equipment according to an embodiment of the present invention. Detailed Implementation
[0072] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0073] like Figure 1-2 As shown, this invention provides a method for predicting belt wear by fusing acoustic signals and acoustic emission signals, comprising the following steps:
[0074] S100, installs an acoustic signal sensor and an acoustic emission sensor to collect signals, performs time-domain analysis and frequency-domain analysis on the two signals, and extracts the time-domain features and frequency-domain features of the signals respectively;
[0075] S200, using an LSTM neural network to fuse the time-domain and frequency-domain features of the two signals, establishes a nonlinear model of belt wear during the blade robot grinding process;
[0076] S300, based on the nonlinear model, using the state estimation technique in the probabilistic statistical method, predicts the probability of stable wear and severe wear of the sand belt at the next moment;
[0077] S400, based on the probability of stable and severe wear of the abrasive belt at the next moment, feeds back to adjust and control the process parameters during the grinding process to prevent excessive tool wear from causing quality degradation phenomena such as burns on the blades.
[0078] Specifically, please see Figure 3 As shown, in step S100,
[0079] The time-domain analysis method includes the time-domain analysis method;
[0080] The time-domain features include: average amplitude, amplitude variance, peak amplitude, and kurtosis coefficient;
[0081] The frequency domain analysis method includes: continuous wavelet transform (CWT); by sliding wavelet basis functions in the time domain and scale domain, the characteristics of the acoustic signals and acoustic emission signals generated during the belt grinding process of the blade robot at different scales and times are obtained;
[0082] The frequency domain features include: wavelet energy, wavelet entropy, wavelet peak value, and wavelet frequency band energy ratio for each frequency band.
[0083] Continuous wavelet transform (CWT) includes:
[0084] The initial acoustic signal and the acoustic emission signal are convolved at different scales and locations using a family of wavelet basis functions to obtain a time-scale representation of the signal. The mathematical model is as follows:
[0085]
[0086] Where a is the scale factor, b is the translation factor, and ψ * (t) is the complex conjugate of the mother wavelet function;
[0087] When performing wavelet transform, it is necessary to first define a, b, and ψ(t); and then repeatedly convolve the parameters at different scales and translations to obtain wavelet transform coefficients at different frequencies and time scales.
[0088] After wavelet transform processing, the acoustic signals and acoustic emission signals generated during the belt grinding process of the blade robot can be divided into three characteristic frequency bands: low frequency band, mid frequency band, and high frequency band.
[0089] Specifically, please see Figure 4 As shown, in step S200,
[0090] The Long Short-Term Memory (LSTM) network method is used to fuse the time-frequency domain features of acoustic signals and acoustic emission signals during the grinding process of the blade robot. LSTM is used to process long sequences and capture the long-term dependence between the fused time-frequency domain features and the abrasive belt wear state, thereby fitting the nonlinear relationship between the two and establishing a nonlinear model between the time-frequency domain features after signal fusion and the abrasive belt wear during the complex blade robot processing.
[0091] The LSTM method includes:
[0092] LSTM includes a forget gate, an input gate, an output gate, and a cell state, such as Figure 5 As shown; the LSTM signal update steps include:
[0093] Step 1: Calculate the forget gate;
[0094] The forget gate determines which information from the time-frequency domain characteristics of the acoustic signal and acoustic emission signal in the previous unit state to discard. It receives the input x at the current time step. t The hidden state h of the previous time step t-1 It outputs a numerical vector between 0 and 1, representing the degree to which information in the cell state should be preserved.
[0095] f t =σ(W f ·[h t-1 ,x t ]+b f )
[0096] Among them, f t W indicates the degree to which information about the cell state should be preserved. f and b f These are the weight matrix and bias term of the forget gate, respectively, where σ is the sigmoid function, [h t-1 ,x t The symbol ] represents concatenating the hidden state of the previous time step and the input of the current time step into a single vector;
[0097] Step 2: Input gate calculation;
[0098] The input gate determines which new acoustic signals and their time-frequency domain characteristics will be updated in the cell state, and it also receives the input x at the current time step.t The hidden state h of the previous time step t-1 It outputs a numerical vector between 0 and 1, representing the degree to which the cell state should be updated:
[0099] i t =σ(W i ·[h t-1 ,x t ]+b i )
[0100] Among them, i t W represents a numerical vector whose values are all between 0 and 1. i Let b represent the weight matrix of the input gate. i This represents the bias term of the input gate;
[0101] Step 3: Update candidate values;
[0102] The update candidate value indicates how much the cell state should be updated based on the input at the current time step; using the input x at the current time step t The hidden state h of the previous time step t-1 The calculation yielded:
[0103]
[0104] Among them, W c With b c These are the weight matrix and bias term of the cell state, respectively. Candidate values representing cell states, where tanh is the hyperbolic tangent function;
[0105] Step 4: Update cell status;
[0106] By combining the decisions of the forget gate and the input gate, the cell state is updated, information is passed between different time steps, and long-term dependencies are recorded. Based on the forget gate, the input gate, and the update candidate values, the cell state C at the current time step can be calculated. t ,as follows:
[0107]
[0108] Step 5: Output gate calculation;
[0109] The output gate determines the hidden state h at the current time step. t What information should it include, and what input x should it receive at the current time step? t The hidden state h of the previous time step t-1 And based on the cell state C at the current time step t Calculate the degree of opening of the output gate and output the hidden state at the current time step;
[0110] o t =σ(W o ·[h t-1 ,x t ]+b o )
[0111] h t =o t ·tanh(C t )
[0112] Among them, W f W i W C W o These are the weight matrices for the forget gate, input gate, candidate update gate, and output gate, respectively. f b i b c b o σ is the corresponding bias term, sigmoid function, and tanh is hyperbolic tangent function.
[0113] Specifically, in step S300,
[0114] An LSTM neural network was used to fit a nonlinear relationship model between the time-frequency domain characteristics of the acoustic signal and acoustic emission signal and the wear state of the abrasive belt; the acoustic signal and acoustic emission signal generated during grinding were collected and analyzed to predict the wear state of the abrasive belt.
[0115] For machining complex blades, brand-new abrasive belts are used for grinding to quickly bring them into a stable wear state, thus avoiding any impact on the blade's machining quality. For monitoring the robotic belt grinding process for blades, it is assumed that the abrasive belt has already entered a stable wear state from the outset. Figure 6 As shown, the wear states of the abrasive belt include: initial wear state, stable wear state, and severe wear state.
[0116] Furthermore, in step S400,
[0117] As the complex blade robotic machining continues, if the abrasive belt wear remains in an ideal stable wear state, machining continues and enters the prediction cycle. If the abrasive belt wear shows signs of severe wear or if it is predicted that the grinding belt will soon enter an unstable grinding state or a severe wear state, the process parameters of the blade robotic grinding process need to be adjusted and optimized in real time based on the parameter combination in the process database to extend the stable wear stage. If the abrasive belt wear has reached a critical value and the abrasive belt no longer has grinding capacity, the abrasive belt needs to be replaced in time to complete the precision grinding of the blade, preventing over-grinding, burning, and other situations that could lead to blade scrapping and ensuring the processing quality of the blade.
[0118] like Figure 7 As shown, the present invention also provides a belt abrasion prediction system that integrates acoustic signals and acoustic emission signals, including: a robot system module, a grinding and polishing machine system module, a signal processing system module, and a feedback control system module;
[0119] The robot system module is used to mount force sensors and grinding force control systems on the robot, which clamps the complex thin-walled blades to be processed by the robot end gripper, and sets up the robot control system to perform overall control and execution.
[0120] The grinding and polishing machine system module is used to arrange the contact wheel and sanding belt. The rotation of the contact wheel connected to it is controlled by the rotation of the motor, which in turn drives the sanding belt above to rotate and controls and adjusts the speed of the grinding and polishing machine control system.
[0121] The signal processing system module is used to acquire acoustic signals and acoustic emission signals during the grinding process, and to perform time-frequency domain analysis on the two types of signals to extract signal features;
[0122] The feedback control system module is used to predict the wear state of the abrasive belt based on the fusion of two signal features using the LSTM neural network, and to perform feedback adjustment and control of process parameters.
[0123] Furthermore, in the signal processing system,
[0124] like Figure 8 As shown, the acoustic signal is acquired through a microphone; the acoustic emission signal is acquired through an acoustic emission sensor;
[0125] The time-domain and frequency-domain features of the acoustic signal generated during the belt grinding process of the blade robot were extracted using time-domain analysis and continuous wavelet transform, respectively.
[0126] The time-domain features include: average amplitude, amplitude variance, peak amplitude, and kurtosis coefficient;
[0127] The frequency domain features include: wavelet energy, wavelet entropy, wavelet peak value, and wavelet band energy ratio for each frequency band;
[0128] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for predicting belt wear by fusing acoustic signals and acoustic emission signals, characterized in that, Includes the following steps: S100, installs an acoustic signal sensor and an acoustic emission sensor to collect signals, performs time-domain analysis and frequency-domain analysis on the two signals, and extracts the time-domain features and frequency-domain features of the signals respectively; S200, using an LSTM neural network to fuse the time-domain and frequency-domain features of the two signals, a nonlinear model of belt wear during the blade robot grinding process is established; S300, based on the nonlinear model, using the state estimation technique in the probabilistic statistical method, predicts the probability of stable wear and severe wear of the sand belt at the next moment; S400, based on the probability of stable wear and severe wear of the abrasive belt at the next moment, feeds back to adjust and control the process parameters during the grinding process to prevent excessive tool wear from causing a decline in blade quality; The LSTM method includes the following steps: Step 1: Calculate the forget gate; The forget gate determines which time-domain and frequency-domain features of the acoustic signal and acoustic emission signal from the previous unit state are discarded; it receives the input from the current time step. The hidden state of the previous time step It outputs a numerical vector between 0 and 1, representing the degree to which information in the cell state should be preserved; in, This indicates the degree to which information about the cell state should be preserved. These are the weight matrix and bias term of the forget gate, respectively, and σ is the sigmoid function. This means concatenating the hidden state of the previous time step and the input of the current time step into a vector; Step 2: Input gate calculation; The input gate determines which new input acoustic signals, along with the time-domain and frequency-domain characteristics of the emitted acoustic signals, will be updated to the cell state. It also receives the input at the current time step. The hidden state of the previous time step It outputs a numerical vector between 0 and 1, representing the degree to which the cell state should be updated. in, This represents a numerical vector whose values are all between 0 and 1. This represents the weight matrix of the input gate. This represents the bias term of the input gate; The LSTM method includes the following steps: Step 3: Update candidate values; The update candidate value indicates how much the cell state should be updated based on the input at the current time step; using the input at the current time step... The hidden state of the previous time step The calculation yielded: t in, These are the weight matrix and bias term of the cell state, respectively. t represents the cell state. Candidate values, where tanh is the hyperbolic tangent function; Step 4: Update cell status; By combining the decisions of the forget gate and the input gate, the cell state is updated, information is passed between different time steps, and long-term dependencies are recorded; based on the forget gate, the input gate, and the update candidate values, the cell state at the current time step is calculated. : t Step 5: Output gate calculation; The output gate determines the hidden state at the current time step. What information should it include, and what input should it receive at the current time step? The hidden state of the previous time step And based on the cell state at the current time step. Calculate the degree of opening of the output gate and output the hidden state at the current time step; in, These are the weight matrices for the forget gate, input gate, candidate value update gate, and output gate, respectively. σ is the corresponding bias term, sigmoid function, and tanh is hyperbolic tangent function.
2. The method for predicting belt wear by fusing acoustic signals and acoustic emission signals according to claim 1, characterized in that, In step S100, The time-domain analysis method includes the time-domain analysis method; The time-domain features include: average amplitude, amplitude variance, peak amplitude, and kurtosis coefficient; The frequency domain analysis method includes: continuous wavelet transform (CWT); and obtaining the characteristics of acoustic signals and acoustic emission signals generated during belt grinding by a blade robot at different scales and times by sliding wavelet basis functions in the time and scale domains. The frequency domain features include: wavelet energy, wavelet entropy, wavelet peak value, and wavelet frequency band energy ratio for each frequency band.
3. The method for predicting belt wear by fusing acoustic signals and acoustic emission signals according to claim 2, characterized in that, The continuous wavelet transform method includes: The initial acoustic signal and the acoustic emission signal are convolved at different scales and locations using a family of wavelet basis functions to obtain a time-scale representation of the signal: Where a is the scaling factor, b is the translation factor, and ψ*(t) is the complex conjugate of the mother wavelet function; When performing wavelet transform, it is necessary to first define a, b, and ψ(t); and then repeatedly convolve the parameters at different scales and translations to obtain wavelet transform coefficients at different frequencies and time scales. After wavelet transform processing, the acoustic signals and acoustic emission signals generated during the belt grinding process of the blade robot are spectrated into three characteristic frequency bands: low frequency band, mid frequency band, and high frequency band.
4. The method for predicting belt wear by fusing acoustic signals and acoustic emission signals according to claim 1, characterized in that, In step S200, The Long Short-Term Memory (LSTM) network method is used to fuse the time-domain and frequency-domain features of acoustic signals and acoustic emission signals during the blade robot grinding process; LSTM is used to process long sequences and capture the long-term dependence between the fused time-domain features and frequency-domain features and the wear state of the abrasive belt, thereby fitting the nonlinear relationship between the two. Based on this, a nonlinear model is established between the fused time-domain features and frequency-domain features and the abrasive belt wear during the machining of complex blades by a robot.
5. The method for predicting belt wear by fusing acoustic signals and acoustic emission signals according to claim 1, characterized in that, In step S300, Based on the aforementioned nonlinear model, the acoustic signals and acoustic emission signals generated during grinding are collected and analyzed to set the initial processing state of the blade robot. During the processing, the state estimation technique in the probabilistic statistical method is used to predict the wear state of the abrasive belt and the probability of severe wear occurring at the next moment. The initial blade robot processing state includes: for blade processing, grinding the abrasive belt to bring it into a stable wear state; for monitoring the blade robot abrasive belt grinding process, it is assumed that the abrasive belt has already entered a stable wear state at the beginning; The wear state of the abrasive belt is divided into: initial wear state, stable wear state, and severe wear state.
6. A method for predicting belt wear by fusing acoustic signals and acoustic emission signals according to any one of claims 1-5, characterized in that, In step S400, As the robotic machining of complex blades continues, different operations are performed on belt abrasion under different conditions; including: When the abrasive belt wear is in an ideal stable wear state, processing continues and the cycle is entered. When the abrasive belt shows signs of severe wear or when it is predicted that the grinding belt will enter an unstable grinding state or a severe wear state in the next moment, it is necessary to adjust and optimize the process parameters of the blade robot grinding process in real time based on the parameter combination of the process database. When the sanding belt wear reaches a critical value, it needs to be replaced in time.
7. A belt wear prediction system that integrates acoustic signals and acoustic emission signals, characterized in that, include: Robot system module, grinding and polishing machine system module, signal processing system module, and feedback control system System module; The robot system module is used to mount force sensors and grinding force control systems on the robot, which clamps the complex thin-walled blades to be processed by the robot end gripper, and sets up the robot control system to perform overall control and execution. The grinding and polishing machine system module is used to arrange the contact wheel and sanding belt. The rotation of the contact wheel connected to it is controlled by the rotation of the motor, which in turn drives the sanding belt above to rotate and controls and adjusts the speed of the grinding and polishing machine control system. The signal processing system module is used to acquire acoustic signals and acoustic emission signals during the grinding process, and to perform time-frequency domain analysis on the two types of signals to extract signal features; The feedback control system module is used to predict the wear state of the abrasive belt based on the fusion of two signal features using an LSTM neural network, and to perform feedback adjustment and control of process parameters. The LSTM method includes the following steps: Step 1: Calculate the forget gate; The forget gate determines which time-domain and frequency-domain features of the acoustic signal and acoustic emission signal from the previous unit state are discarded; it receives the input from the current time step. The hidden state of the previous time step It outputs a numerical vector between 0 and 1, representing the degree to which information in the cell state should be preserved; in, This indicates the degree to which information about the cell state should be preserved. These are the weight matrix and bias term of the forget gate, respectively, and σ is the sigmoid function. This means concatenating the hidden state of the previous time step and the input of the current time step into a vector; Step 2: Input gate calculation; The input gate determines which new input acoustic signals, along with the time-domain and frequency-domain characteristics of the emitted acoustic signals, will be updated to the cell state. It also receives the input at the current time step. The hidden state of the previous time step It outputs a numerical vector between 0 and 1, representing the degree to which the cell state should be updated. in, This represents a numerical vector whose values are all between 0 and 1. This represents the weight matrix of the input gate. This represents the bias term of the input gate; The LSTM method includes the following steps: Step 3: Update candidate values; The update candidate value indicates how much the cell state should be updated based on the input at the current time step; using the input at the current time step... The hidden state of the previous time step The calculation yielded: t in, These are the weight matrix and bias term of the cell state, respectively. t represents the cell state. Candidate values, where tanh is the hyperbolic tangent function; Step 4: Update cell status; By combining the decisions of the forget gate and the input gate, the cell state is updated, information is passed between different time steps, and long-term dependencies are recorded; based on the forget gate, the input gate, and the update candidate values, the cell state at the current time step is calculated. : t Step 5: Output gate calculation; The output gate determines the hidden state at the current time step. What information should it include, and what input should it receive at the current time step? The hidden state of the previous time step And based on the cell state at the current time step. Calculate the degree of opening of the output gate and output the hidden state at the current time step; in, These are the weight matrices for the forget gate, input gate, candidate value update gate, and output gate, respectively. σ is the corresponding bias term, sigmoid function, and tanh is hyperbolic tangent function.
8. The belt wear prediction system based on the fusion of acoustic signals and acoustic emission signals according to claim 7, characterized in that, In the signal processing system The acoustic signal is acquired via a microphone; the acoustic emission signal is acquired via an acoustic emission sensor.