TWI853549B - Processing condition monitoring method based on sound signal and system thereof - Google Patents

Abstract

A processing condition monitoring method based on sound signal is proposed. The tool wear prediction method based on sound signal includes a signal extracting step, a signal preprocessing step, a feature extracting step, a remaining life predicting step and a model updating step. The signal extracting step includes obtaining the sound signal of the tool wear during processing. The signal preprocessing step includes converting the sound signal into an audio file. The feature extracting step includes extracting the characteristic frequency band of the audio file and performing calculations with the audio characteristic analysis parameters to generate a plurality of characteristic indicators. The remaining life predicting step includes inputting the sound file and the plurality of characteristic indicators into a tool wear processing condition prediction model to generate the tool wear processing condition prediction result corresponding to the sound signal. The model updating step includes obtaining the audio file and comparing the audio file with a plurality of reference audio to generate a feedback results, and marking the audio characteristic analysis parameters according to the feedback results, and making the tool wear processing condition prediction model to reinforce and update a plurality of standard feature.

Description

Processing state monitoring method and system based on sound signal

The present invention relates to a processing state monitoring method and a system thereof, and in particular to a processing state monitoring method and a system thereof based on sound signals.

The condition of the tool has a great impact on the processing quality, especially the process condition of the tool directly affects the processing accuracy and surface roughness of the workpiece. Tool damage not only affects the quality of the workpiece, but in serious cases it may even affect the operation of the entire processing system, causing huge losses. The current common method for judging the timing of tool replacement in the industry is to replace the tool when the count reaches a fixed number of processed parts in mass production, or to replace the tool based on the experience of technicians through processing sound, workpiece surface or naked eye observation of the tool condition.

In addition to being relatively conservative in using tools and easily replacing tools early, resulting in a waste of tool costs, the aforementioned method also increases labor costs by using manual judgment. It can be seen from this that there is currently a lack of a low-cost, cost-effective and certain accurate processing process status monitoring method and system on the market, so relevant industries are looking for solutions.

The purpose of the present invention is to provide a processing state monitoring method and system based on sound signals, which obtains tool process state prediction results by analyzing the sound signals of tool processing and continuously updating the tool process state prediction model. In addition to low equipment cost, it can also accurately predict the remaining life of the tool.

According to an implementation method of the method aspect of the present invention, a processing process state monitoring method based on sound signals is provided, which includes a signal acquisition step, a signal preprocessing step, a feature extraction step, a life prediction step and a model update step. The signal acquisition step includes driving a sensor to obtain a sound signal of a tool processing process. The signal preprocessing step includes driving a computing processor to convert the sound signal into an audio file and store it in a memory. The feature extraction step includes driving the computing processor to capture a feature frequency band of the audio file and perform calculations with an audio feature analysis parameter to generate a plurality of feature indicators. The life prediction step includes driving a computing processor to input the audio file and each feature index into a tool process state prediction model to compare each feature index with a plurality of reference features, thereby generating a tool process state prediction result corresponding to the sound signal. The model update step includes an audio backtracking model update step, wherein the audio backtracking model update step includes driving a computing processor to obtain an audio file from a memory, and compare the audio file with a plurality of reference audios to generate a feedback result. The computing processor labels the audio feature analysis parameters according to the feedback result, and causes the tool process state prediction model to execute a deep learning program to reinforce and update each reference feature.

According to one implementation method of the structural aspect of the present invention, a processing process state monitoring system based on sound signals is provided, comprising a sensor, a memory and an operation processor. The sensor is used to obtain a sound signal of a tool processing process. The memory stores an audio feature analysis parameter, a plurality of baseline features and a plurality of reference audios. The operation processor is electrically connected to the sensor and the memory, and the operation processor is configured to implement an operation comprising the following steps: a signal preprocessing step, a feature extraction step, a life prediction step and a model updating step. The signal preprocessing step comprises converting the sound signal into an audio file and storing it in the memory. The feature extraction step includes capturing a feature frequency band of the audio file and performing calculations with audio feature analysis parameters to generate a plurality of feature indicators. The life prediction step inputs the audio file and each feature indicator into a tool process state prediction model, and the tool process state prediction model compares each feature indicator with each benchmark feature to generate a tool process state prediction result corresponding to the sound signal. The model updating step includes an audio backtracking model updating step, which includes obtaining an audio file from a memory and comparing the audio file with each reference audio to generate a feedback result, labeling the audio feature analysis parameters according to the feedback result, and causing the tool process state prediction model to execute a deep learning program to reinforce and update each benchmark feature.

Thus, the present invention obtains tool process state prediction results by analyzing the sound signal of tool processing and using a tool process state prediction model that can be continuously updated by audio backtracking and prediction results, thereby having the effect of low equipment cost, being cost-effective and being able to accurately predict the remaining life of the tool.

The following will describe several embodiments of the present invention with reference to the drawings. For the sake of clarity, many practical details will be described together in the following description. However, it should be understood that these practical details should not be used to limit the present invention. That is to say, in some embodiments of the present invention, these practical details are not necessary. In addition, in order to simplify the drawings, some commonly used structures and components will be shown in the drawings in a simple schematic manner; and repeated components may be represented by the same number.

In addition, in this article, when a certain component (or unit or module, etc.) is "connected" to another component, it may refer to that the component is directly connected to the other component, or it may refer to that the component is indirectly connected to the other component, that is, there are other components between the component and the other component. When it is clearly stated that a certain component is "directly connected" to another component, it means that there are no other components between the component and the other component. The terms first, second, third, etc. are only used to describe different components, and there is no restriction on the components themselves. Therefore, the first component can also be renamed as the second component. Moreover, the combination of components/units/circuits in this article is not a generally known, conventional or known combination in this field. Whether the components/units/circuits themselves are known cannot be used to determine whether their combination relationship is easy to be completed by ordinary knowledgeable people in the technical field.

Please refer to FIG. 1 and FIG. 2 together. FIG. 1 is a schematic diagram of a processing state monitoring system 100 based on sound signals according to a first embodiment of the present invention; and FIG. 2 is a flow chart of a processing state monitoring method 200 based on sound signals according to a second embodiment of the present invention. The processing state monitoring system 100 based on sound signals is configured to implement the processing state monitoring method 200 based on sound signals to monitor a tool and predict the remaining life of the tool. It should be noted that the processing state monitoring method 200 based on sound signals according to the present invention is not limited to being implemented through the processing state monitoring system 100 based on sound signals according to the present invention. In the embodiment of the present invention, the machining process status monitoring is tool wear monitoring.

In FIG. 1 , a processing state monitoring system 100 based on sound signals includes a sensor 110, a memory 120, and an operation processor 130. The sensor 110 is used to obtain a sound signal A of a tool processing process. The memory 120 stores an audio feature analysis parameter 121, a plurality of benchmark features 122, and a plurality of reference audios 123. The operation processor 130 is electrically connected to the sensor 110 and the memory 120, and the operation processor 130 is configured to implement a processing state monitoring method 200 based on sound signals. The sensor 110 can receive frequencies of 20 Hz to 30 kHz. The sensor 110 can be a microphone or other electronic device that can capture audio, but the present invention is not limited thereto. The memory 120 can be a random access memory (RAM) or other type of dynamic storage device that can store information and instructions for the computing processor 130 to execute, but the present invention is not limited thereto. The computing processor 130 can be a processor, a microprocessor, a central processing unit (CPU), a computer, a mobile device processor, a cloud processor, or other electronic computing processors, but the present invention is not limited thereto.

In FIG. 2 , the processing state monitoring method 200 based on sound signals includes a signal acquisition step S01, a signal preprocessing step S02, a feature extraction step S03, a life prediction step S04, and a model updating step S05. The signal acquisition step S01 includes driving the sensor 110 to obtain the sound signal A of the tool processing process. The signal preprocessing step S02 includes driving the computing processor 130 to convert the sound signal A into an audio file B and store it in the memory 120. The signal pre-processing step S02 includes driving the computational processor 130 to perform waveform displacement, harmonic distortion, resampling or signal-to-noise ratio control (SNR Control) on the sound signal A to generate an audio file B.

Specifically, referring to FIGS. 1 to 7B, FIG. 3A is a diagram showing the relationship between the amplitude and time of the sound signal A according to FIG. 2; FIG. 3B is a diagram showing the relationship between the frequency and time of the sound signal A according to FIG. 2; FIG. 4 is a diagram showing the relationship between the amplitude and time of the sound file B subjected to waveform displacement according to the sound signal A according to FIG. 3A; FIG. 5A is a diagram showing the relationship between the amplitude and time of the sound file B subjected to harmonic distortion according to the sound signal A according to FIG. 3A; FIG. 5B is a diagram showing the relationship between the amplitude and time of the sound file B subjected to harmonic distortion according to the sound signal A according to FIG. 3B. FIG6A is a diagram showing the relationship between the amplitude and time of the audio file B resampled according to the sound signal A of FIG3A; FIG6B is a diagram showing the relationship between the frequency and time of the audio file B resampled according to the sound signal A of FIG3B; FIG7A is a diagram showing the relationship between the amplitude and time of the audio file B with the signal-to-noise ratio controlled according to the sound signal A of FIG3A; and FIG7B is a diagram showing the relationship between the frequency and time of the audio file B with the signal-to-noise ratio controlled according to the sound signal A of FIG3B. In the signal pre-processing step S02, the operation processor 130 obtains the sound signal A from the sensor 110, and performs a frequency band visual analysis on the sound signal A to generate an audio file B, thereby achieving the effect of data enhancement and data volume increase, thereby improving the robustness of a tool process state prediction model M. In FIG. 4, the operation processor 130 performs time displacement adjustment on the sound signal A (shown in FIG. 3A), shifting the entire waveform forward or backward as required. In FIG. 5A to FIG. 5B, the operation processor 130 performs harmonic distortion on the sound signal A (shown in FIG. 3A), and enhances the vibration and frequency of the main waveform of the sound signal A. In FIGS. 6A to 6B, the computing processor 130 resamples the sound signal A to make the spectrum of the sound signal A more detailed. In FIGS. 7A to 7B, the computing processor 130 controls the signal-to-noise ratio of the sound signal A and adds noise to make the tool process state prediction model M have anti-noise capability.

Referring to FIG. 1 and FIG. 2, the feature extraction step S03 includes driving the computation processor 130 to capture a feature frequency band of the audio file B and performing algorithm feature computation and feature classification with the audio feature analysis parameter 121 pre-stored in the memory 120, and generating a plurality of feature indicators C through convolution operation. The feature indicators C include root mean square, arithmetic mean, standard deviation, maximum value, kurtosis and skewness, but the present invention is not limited thereto.

Referring to FIG. 1 and FIG. 2, the life prediction step S04 includes driving the computing processor 130 to input the audio file B and each characteristic index C into the tool process state prediction model M (shown in FIG. 8), and comparing each characteristic index C with the reference characteristic 122 pre-stored in the memory 120, thereby generating a tool process state prediction result corresponding to the sound signal A. In the second embodiment, the reference characteristic 122 pre-stored in the memory 120 is the audio characteristic index C of the tool under different process processing. In this way, it is possible to confirm whether the tool is close to the critical life according to the tool process state prediction result, so as to stop the machine immediately and replace the tool. The following is a detailed embodiment to illustrate the details of the above tool process state prediction model M.

With reference to FIG. 8, FIG. 8 is a schematic diagram showing the structure of the tool process state prediction model M of the third embodiment of the present invention. The tool process state prediction model M includes a convolutional neural network (CNN) model and a recurrent neural network (RNN) model, wherein the convolutional neural network model and the recurrent neural network model are end-to-end models. In FIG. 8, the convolutional neural network model includes a convolution layer L1 and a classification layer L2. The convolution layer L1 includes a plurality of maximum pooling units L11 to perform convolution and pooling operations and transform with activation functions, thereby filtering out the noise of the audio file B and extracting the feature index C, thereby reducing the data size and reducing the computational loss. The classification layer L2 includes a plurality of fully connected (Dense) units L21 to classify whether the audio file B is abnormal and output the tool process state prediction result. The recurrent neural network model includes a gated loop layer L3, and the gated loop layer L3 includes a gated loop (GRU) unit L31 to capture the time features of the feature index C extracted by the maximum pooling unit L11 and unify the size of time series of different lengths. In the third embodiment, the gated loop unit L31 of the gated loop layer L3 is connected between the multiple maximum pooling units L11 of the convolution layer L1 and the multiple fully connected units L21 of the classification layer L2. After the audio file B and each feature index C are input into the tool process state prediction model M, the tool process state prediction result is obtained by sequentially passing through the multiple maximum pooling units L11, the gated loop unit L31 and the multiple fully connected units L21.

Referring to FIGS. 1 to 2, the model updating step S05 includes an audio backtracking model updating step S051 and a prediction result model updating step S052. The audio backtracking model updating step S051 includes driving the computing processor 130 to obtain the audio file B from the memory 120, and comparing the audio file B with each reference audio 123 stored in the memory 120 to generate a feedback result. The computing processor 130 annotates the audio feature analysis parameter 121 according to the feedback result, and causes the tool process state prediction model M to execute a deep learning program to reinforce and update each benchmark feature 122 of the memory 120. The prediction result model updating step S052 includes driving the computing processor 130 to enhance the acoustic feature analysis parameter 121 according to the tool process state prediction result, and making the tool process state prediction model M execute a deep learning program to enhance and update each benchmark feature 122 of the memory 120. In this way, the tool process state prediction model M can update the benchmark feature 122 by two methods, namely, audio backtracking training and tool process state prediction result training, so as to enhance the accuracy of model analysis.

From the above implementation, it can be seen that the present invention has the following advantages: First, the tool process state prediction result can be obtained by analyzing the sound signal, and the equipment cost is low and cost-effective. Second, the tool process state prediction model is trained by using both audio backtracking and prediction results, which can obtain the tool process state prediction result more accurately, thereby avoiding tool cost waste.

Although the present invention has been disclosed as above by the embodiments, they are not intended to limit the present invention. Any person with ordinary knowledge in the relevant technical field can make some changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be defined by the scope of the attached patent application.

100: Process state monitoring system based on sound signal 110: Sensor 120: Memory 121: Audio feature analysis parameters 122: Benchmark feature 123: Reference audio 130: Calculation processor 200: Process state monitoring method based on sound signal S01: Signal acquisition step S02: Signal preprocessing step S03: Feature extraction step S04: Life prediction step S05: Model update step S051: Audio backtracking model update step S052: Prediction result model update step A: Sound signal B: Audio file C: Feature index L1: Convolutional layer L11: Max pooling unit L2: Classification layer L21: Fully connected unit L3: Gated loop layer L31: Gated loop unit M: Tool process state prediction model

Figure 1 is a schematic diagram of a processing state monitoring system based on sound signals of the first embodiment of the present invention; Figure 2 is a schematic diagram of a process flow of a processing state monitoring method based on sound signals of the second embodiment of the present invention; Figure 3A is a schematic diagram showing the relationship between the amplitude and time of the sound signal according to Figure 2; Figure 3B is a schematic diagram showing the relationship between the frequency and time of the sound signal according to Figure 2; Figure 4 is a schematic diagram showing the relationship between the amplitude and time of the audio file of the sound signal according to Figure 3A for waveform displacement; Figure 5A is a schematic diagram showing the relationship between the amplitude and time of the audio file of the sound signal according to Figure 3A for harmonic distortion; FIG. 5B is a diagram showing the relationship between the frequency and time of the sound file harmonically distorted according to the sound signal of FIG. 3B; FIG. 6A is a diagram showing the relationship between the amplitude and time of the sound file resampled according to the sound signal of FIG. 3A; FIG. 6B is a diagram showing the relationship between the frequency and time of the sound file resampled according to the sound signal of FIG. 3B; FIG. 7A is a diagram showing the relationship between the amplitude and time of the sound file for which the signal-to-noise ratio is controlled according to the sound signal of FIG. 3A; FIG. 7B is a diagram showing the relationship between the frequency and time of the sound file for which the signal-to-noise ratio is controlled according to the sound signal of FIG. 3B; and FIG. 8 is a diagram showing the framework of the tool process state prediction model of the third embodiment of the present invention.

200: Processing state monitoring method based on sound signals

S01: Signal collection step

S02: Signal pre-processing step

S03: Feature extraction step

S04: Life expectancy prediction steps

S05: Model update step

S051: Audio backtracking model update step

S052: Prediction result model update step

A: Sound signal

B: Audio file

C: Characteristic indicators

Claims (6)

A method for monitoring the state of a machining process based on sound signals comprises: a signal acquisition step, comprising driving a sensor to obtain a sound signal of a tool machining process; a signal pre-processing step, comprising driving a computing processor to convert the sound signal into an audio file and store it in a memory; a feature extraction step, comprising driving the computing processor to capture a feature of the audio file. The characteristic frequency band is calculated with an audio characteristic analysis parameter to generate a plurality of characteristic indicators; a life prediction step includes driving the calculation processor to input the audio file and the characteristic indicators into a tool process state prediction model to compare the characteristic indicators with a plurality of benchmark characteristics to generate a tool process state prediction result corresponding to the sound signal; and a model update step , comprising: an audio backtracking model updating step, comprising driving the computing processor to obtain the audio file from the memory, and comparing the audio file with a plurality of reference audios to generate a feedback result, the computing processor annotating the audio feature analysis parameters according to the feedback result, and causing the tool process state prediction model to execute a deep learning program to strengthen and update the benchmark features The tool process state prediction model includes: a convolutional neural network model, including: a convolutional layer, including a plurality of maximum pooling units; and a classification layer, including a plurality of fully connected units; and a recurrent neural network model, including: a gated loop layer, connected between the maximum pooling units and the fully connected units, and including a gated loop unit. The processing process state monitoring method based on sound signals as described in claim 1, wherein the model updating step further includes: a prediction result model updating step, including driving the computing processor to enhance the audio feature analysis parameters according to the tool process state prediction result, and making the tool process state prediction model execute the deep learning program to update the baseline features. As described in claim 1, the processing state monitoring method based on sound signals, wherein the signal pre-processing step further includes driving the computing processor to perform waveform shift, harmonic distortion, resampling or signal-to-noise ratio control on the sound signal to generate the audio file. A processing process status monitoring system based on sound signals includes: a sensor for obtaining a sound signal of a tool processing process; a memory for storing an audio feature analysis parameter, multiple benchmark features and multiple reference audios; and an operation processor electrically connected to the sensor and the memory, the operation processor being configured to implement operations including the following steps: A signal preprocessing step , including converting the sound signal into an audio file and storing it in the memory; a feature extraction step, including capturing a feature frequency band of the audio file and performing calculations with the audio feature analysis parameters to generate a plurality of feature indicators; a life prediction step, inputting the audio file and the feature indicators into a tool process state prediction model, and the tool process state prediction model compares the feature indicators with the benchmarks The invention relates to a method for updating a model of a tool manufacturing process state according to the characteristics of the sound signal; and a model updating step, comprising: an audio backtracking model updating step, comprising obtaining the audio file from the memory, and comparing the audio file with the reference audios to generate a feedback result, annotating the audio characteristic analysis parameters according to the feedback result, and making the tool manufacturing process state prediction model execute a deep learning program. To strengthen and update the baseline features; wherein the tool process state prediction model includes: a convolutional neural network model, including: a convolutional layer, including a plurality of maximum pooling units; and a classification layer, including a plurality of fully connected units; and a recurrent neural network model, including: a gated loop layer, connected between the maximum pooling units and the fully connected units, and including a gated loop unit. The processing process state monitoring system based on sound signals as described in claim 4, wherein the model updating step further includes: A prediction result model updating step, including enhancing the audio feature analysis parameters according to the tool process state prediction result, and making the tool process state prediction model execute the deep learning program to update the baseline features. As described in claim 4, the processing state monitoring system based on sound signals, wherein the signal pre-processing step further includes driving the computing processor to perform waveform shift, harmonic distortion, resampling or signal-to-noise ratio control on the sound signal to generate the audio file.