Real-time detection method of processing quality of processing equipment and related equipment

Abstract

A method for real-time detection of processing quality of processing equipment and related equipment, the method comprising: real-time acquisition of sound data of processing equipment during processing; extraction of sound features from sound data to obtain initial sound features; inputting the initial sound features into a real-time detection model for processing quality of processing equipment, and outputting reconstructed sound features; calculating the abnormality scores of the reconstructed sound features and the initial sound features, and comparing the abnormality scores of the reconstructed sound features and the initial sound features with the processing abnormality threshold of the real-time detection model for processing quality of processing equipment; if the abnormality scores of the reconstructed sound features and the initial sound features are greater than the processing abnormality threshold, outputting equipment abnormality prompt information. The present application can perform contactless real-time detection of processing equipment, improve the efficiency of real-time detection of processing quality of processing equipment, reduce the loss of defective raw materials and labor costs, and improve the processing quality of processing equipment.

Description

Technical Field

[0001] The present application relates to the technical field of detection of processing quality of numerical control processing equipment, and in particular to a real-time detection method for processing quality of processing equipment based on autoencoder acoustic sensing technology and related equipment. Background Art

[0002] In automated production, the normal operation of processing equipment (such as grinders, lathes, etc.) plays a vital role in processing quality. At present, the detection efficiency of processing equipment is low, and abnormalities cannot be discovered in time during the processing, resulting in waste of processing materials and increased production costs for enterprises. Summary of the invention

[0003] The embodiment of the present application provides a real-time detection method for the processing quality of processing equipment and related equipment, which can perform non-contact real-time detection of the processing status of the processing equipment, reduce the loss of defective raw materials and labor costs, and improve the processing quality.

[0004] The first aspect of the present application provides a method for real-time detection of processing quality of processing equipment, the method comprising:

[0005] Acquire the sound data of processing equipment in real time;

[0006] Extracting sound features from the sound data to obtain initial sound features;

[0007] The initial sound feature is input into the real-time detection model of processing quality of processing equipment, and a reconstructed sound feature is output. The real-time detection model of processing quality of processing equipment is a trained dual-channel adaptive feature fusion autoencoder network, and the dual-channel adaptive feature fusion autoencoder network includes a time feature compression module, a frequency domain feature compression module, a feature fusion module and a decoder module, wherein:

[0008] The time series feature compression module is used to compress and map the time series features in the initial sound features to obtain a time series feature vector;

[0009] The frequency domain feature compression module is used to compress and map the frequency domain features in the initial sound features to obtain a frequency domain feature vector;

[0010] The feature fusion module is used to determine the weight of the time series feature vector according to the time series feature vector and the frequency domain feature vector;

[0011] The decoder module is used to decode the time series feature vector weighted according to the weight value and the weighted frequency domain feature vector to obtain a reconstructed sound feature;

[0012] Calculating anomaly scores of the reconstructed sound features and the initial sound features, and comparing the anomaly scores of the reconstructed sound features and the initial sound features with a processing anomaly threshold of the processing quality real-time detection model of the processing equipment;

[0013] If the abnormality scores of the reconstructed sound feature and the initial sound feature are greater than the processing abnormality threshold, device abnormality prompt information is output.

[0014] The real-time detection method for processing quality of processing equipment provided in the embodiment of the present application obtains the sound data of the processing equipment, and detects the processing status of the processing equipment according to the sound data of the processing equipment. Compared with the traditional manual sampling method, the real-time detection method for processing quality of processing equipment provided in the embodiment of the present application can perform non-contact real-time detection of processing equipment, improve detection efficiency, reduce the loss of defective raw materials and labor costs, and improve the processing quality of processing equipment.

[0015] In some optional embodiments, the temporal feature compression module is a bidirectional recurrent network, the frequency domain feature compression module is a fully connected encoding network, the feature fusion module is a gated compression network, and the decoder module is a fully connected decoding network.

[0016] In some optional embodiments, the training method of the dual-channel adaptive feature fusion autoencoder network includes:

[0017] Acquiring sound data of a processing process of the processing equipment, and marking the sound data set;

[0018] Extracting sound features from the sound data set to obtain sound samples, and dividing the sound samples into a training set and a validation set;

[0019] The dual-channel adaptive feature fusion autoencoder model is trained using the training set, and the performance of the dual-channel adaptive feature fusion autoencoder model is verified using the verification set to obtain the real-time detection model for processing quality of the processing equipment.

[0020] In some optional implementations, the method for calculating the processing abnormality threshold includes:

[0021] Inputting each sound sample in the training set into the real-time detection model for processing quality of processing equipment to obtain an abnormality score for each sound sample in the training set;

[0022] Statistically calculating the Γ distribution parameters of the abnormal scores of all sound samples in the training set;

[0023] The abnormal score at the Γ distribution percentage function β∈(0,1) is taken as the processing abnormality threshold.

[0024] In some optional implementations, extracting sound features from the sound dataset includes:

[0025] Performing discrete Fourier transform on the sound data set to obtain a frequency spectrum of the sound data set;

[0026] Calculating a discrete energy spectrum of the sound data set according to the frequency spectrum of the sound data set;

[0027] Passing the discrete energy spectrum of the sound data set through a triangular filter bank uniformly distributed on Mel frequencies;

[0028] Performing a logarithmic operation on the output of the triangular filter bank to obtain a logarithmic power spectrum of the sound data set;

[0029] Perform discrete cosine transform on the logarithmic power spectrum of the sound data set to obtain Mel-frequency cepstral coefficient features.

[0030] In some optional implementations, the sound features corresponding to each n frames of the sound data set are taken as a sound sample, and the number of sound samples obtained by the sound data set per second is:

[0031]

[0032] in, Indicates rounding down, fs is the sampling frequency, n_fft is the number of Fourier transform points per frame, n is the number of frames corresponding to each sound sample, and n is an integer greater than or equal to 1.

[0033] In some optional implementations, the training of the dual-channel adaptive feature fusion autoencoder model using the training set includes:

[0034] Input each sound sample in the training set into the time series feature compression module and the frequency domain feature compression module simultaneously;

[0035] The time series feature compression module compresses and maps the time series features in the sound sample to obtain a time series feature vector of the sound sample;

[0036] The frequency domain feature compression module compresses and maps the frequency domain features in the sound sample to obtain a frequency domain feature vector of the sound sample;

[0037] The time series feature vector and the frequency domain feature vector are input into the feature fusion module to obtain the weight of the time series feature vector of the sound sample.

[0038] In some optional implementations, the weight of the time series feature vector is determined according to the following formula:

[0039]

[0040] Among them, α is the weight of the time series feature vector, represents the feature fusion module, [,] represents matrix concatenation, z t represents the time series feature vector, z f represents the frequency domain eigenvector, is the network parameter of the feature fusion module, and σ is the sigmoid function.

[0041] In some optional embodiments, the loss function used in training the dual-channel adaptive feature fusion autoencoder network is a mean square error function, and the anomaly score is the square of the difference between the reconstructed sound feature and the initial sound feature.

[0042] In some optional implementations, before extracting sound features from the sound data, the method further includes:

[0043] Pre-emphasize the sound data; and / or

[0044] Framing the sound data; and / or

[0045] Acquire effective sound data from the sound data.

[0046] In some optional embodiments, the method further comprises:

[0047] Determine whether the abnormality scores of the reconstructed sound feature and the initial sound feature are greater than the processing abnormality threshold for a preset number of consecutive times;

[0048] If the abnormality scores of the reconstructed sound feature and the initial sound feature are greater than the processing abnormality threshold for a preset number of consecutive times, the processing equipment is controlled to stop running, and an alarm message of equipment abnormality is sent to the user.

[0049] The second aspect of the present application provides a real-time detection device for processing quality of processing equipment, which includes a sound data acquisition module, a main control analysis module and a communication early warning module, wherein:

[0050] The sound collection module is used to collect the sound data of the processing equipment in real time during processing;

[0051] The main control analysis module includes:

[0052] The data analysis unit is used to extract sound features from the sound data to obtain initial sound features, input the initial sound features into a real-time detection model for processing quality of a processing equipment, and output a reconstructed sound feature. The real-time detection model for processing quality of a processing equipment is a trained dual-channel adaptive feature fusion autoencoder network. The dual-channel adaptive feature fusion autoencoder network includes a time series feature compression module, a frequency domain feature compression module, a feature fusion module, and a decoder module, wherein: the time series feature compression module is used to compress and map the time series features in the initial sound features to obtain a time series feature vector; the frequency domain feature compression module is used to compress and map the frequency domain features in the initial sound features to obtain frequency domain feature vector; the feature fusion module is used to determine the weight of the time series feature vector according to the time series feature vector and the frequency domain feature vector; the decoder module is used to decode the time series feature vector weighted according to the weight and the weighted frequency domain feature vector to obtain the reconstructed sound feature; the data analysis unit is also used to calculate the abnormality score of the reconstructed sound feature and the initial sound feature, and compare the abnormality score of the reconstructed sound feature and the initial sound feature with the processing abnormality threshold of the real-time detection model of the processing quality of the processing equipment, if the abnormality score of the reconstructed sound feature and the initial sound feature is greater than the processing abnormality threshold, output the equipment abnormality prompt information;

[0053] A data storage unit, used for storing the sound data and the real-time detection model of processing quality of the processing equipment;

[0054] A display unit, used to display abnormal prompt information of the device;

[0055] A control unit, used to control the operation of the data analysis unit, the data storage unit and the display unit;

[0056] The communication warning module is used to send a control signal to the processing equipment to stop the operation of the processing equipment when the abnormality scores of the reconstructed sound feature and the initial sound feature are greater than the processing abnormality threshold.

[0057] It should be understood that the real-time detection device for processing quality of the processing equipment of the second aspect provided above corresponds to the method of the first aspect mentioned above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding method provided above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a schematic diagram of the application environment of the real-time detection method of processing quality of processing equipment provided in an embodiment of the present application.

[0059] Figure 2It is a flow chart of a method for real-time detection of processing quality of processing equipment provided in an embodiment of the present application.

[0060] Figure 3 It is a schematic diagram of a real-time detection model for processing quality of processing equipment used in an embodiment of the present application. DETAILED DESCRIPTION

[0061] To facilitate understanding, some illustrations of concepts related to the embodiments of the present application are given by way of example for reference.

[0062] It should be noted that in this application, "at least one" means one or more, and "more than one" means two or more than two. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. The terms "first", "second", "third", "fourth", etc. (if any) in the specification, claims and drawings of this application are used to distinguish similar objects, rather than to describe a specific order or sequence.

[0063] In order to better understand the real-time detection method for processing quality of processing equipment and related equipment provided in the embodiments of the present application, the application scenario of the real-time detection method for processing quality of processing equipment of the present application is first described below.

[0064] Figure 1 It is a schematic diagram of the application environment of the real-time detection method of processing quality of processing equipment provided in an embodiment of the present application.

[0065] like Figure 1 As shown, the real-time detection method for processing quality of processing equipment provided in the embodiment of the present application performs real-time detection of the processing quality of processing equipment 11 through a real-time detection device 10 for processing quality of processing equipment. The real-time detection device 10 for processing quality of processing equipment includes a sound data acquisition module 101, a main control analysis module 102, and a communication warning module 103. The sound data acquisition module 101, the processing equipment 11, and the communication warning module 103 are all connected to the main control analysis module 102. The communication warning module 103 is connected to the processing equipment 11.

[0066] The sound data acquisition module 10 can be installed on or near the processing equipment 11, and is used to collect sound data during processing by the processing equipment 11 (including sound data used for model training and sound data used for real-time detection), and transmit the collected sound data to the main control analysis module 102.

[0067] The sound data acquisition module 101 may include a sensor unit 1011, a signal amplification unit 1012, an analog-to-digital conversion unit 1013, and a data transmission unit 1014. The sensor unit 1011 is used to sense the sound signal during processing by the processing device 11. The signal amplification unit 1012 is used to amplify the sound signal sensed by the sensor unit 1011. The analog-to-digital conversion unit 1013 is used to perform analog-to-digital conversion on the amplified sound signal to obtain the sound data during processing by the processing device 11. The data transmission unit 1014 is used to transmit the sound data to the main control analysis module 102.

[0068] The main control analysis module 102 is used to extract features, analyze and model the sound data collected by the sound data collection module 101 and control other modules (ie, the sound data collection module 101 and the communication warning module 103 ) to communicate with the processing equipment 11 .

[0069] The main control analysis module 102 may include a data storage unit 1021, a control unit 1022, a data analysis unit 1023 and a display unit 1024. The control unit 1022 is connected to the data analysis unit 1023, the data storage unit 1021 and the display unit 1024. The control unit 1022 is used to control the operation of other units in the main control analysis module 102 (i.e., the data storage unit 1021, the data analysis unit 1023 and the display unit 1024). The data analysis unit 1023 is used to train and verify the performance of the dual-channel adaptive feature fusion autoencoder network, obtain a real-time detection model for the processing quality of the processing equipment, and use the real-time detection model for the processing quality of the processing equipment to analyze the sound data collected in real time to obtain the processing state detection result of the processing equipment 11. The data storage unit 1021 is used to store sound data and the trained dual-channel adaptive feature fusion autoencoder network (i.e., the real-time detection model for the processing quality of the processing equipment). The display unit 1024 is used to display the processing state detection result of the processing equipment 11, such as the abnormal equipment prompt information of the processing equipment 11.

[0070] When performing real-time detection, the data analysis unit 1023 extracts sound features from the sound data acquired in real time to obtain an initial sound feature, inputs the initial sound feature into the real-time detection model for processing quality of the processing equipment, outputs a reconstructed sound feature, calculates the abnormality scores of the reconstructed sound feature and the initial sound feature, and compares the abnormality scores of the reconstructed sound feature and the initial sound feature with the processing abnormality threshold of the real-time detection model for processing quality of the processing equipment. If the abnormality scores of the reconstructed sound feature and the initial sound feature are greater than the processing abnormality threshold, output equipment abnormality prompt information.

[0071] The real-time detection model of processing quality of processing equipment is a trained dual-channel adaptive feature fusion autoencoder network, which includes a time series feature compression module, a frequency domain feature compression module, a feature fusion module and a decoder module. The time series feature compression module is used to compress and map the time series features in the initial sound features to obtain a time series feature vector; the frequency domain feature compression module is used to compress and map the frequency domain features in the initial sound features to obtain a frequency domain feature vector. The feature fusion module is used to determine the weight of the time series feature vector based on the time series feature vector and the frequency domain feature vector. The decoder module is used to decode the time series feature vector weighted according to the weight and the weighted frequency domain feature vector to obtain the reconstructed sound feature.

[0072] The communication warning module 103 is used to send a control signal to the processing device 11 to stop the operation of the processing device 11 when a processing abnormality is detected in the processing device 11 (i.e., the abnormality score of the reconstructed sound feature and the initial sound feature is greater than the processing abnormality threshold). The communication warning module 103 can also be used to send the processing status detection result of the processing device 11 to the user.

[0073] When a processing abnormality of the processing equipment 11 is detected, the main control analysis module 102 can output a prompt message of the processing abnormality, or control the processing equipment 11 to stop running.

[0074] The main control analysis module 102 can perform human-computer interaction with the user through a keyboard, mouse, remote control, touch pad or voice control device.

[0075] The processing equipment 11 may be a grinder, a lathe or a milling machine, or may be other processing equipment.

[0076] According to the real-time detection method for processing quality of processing equipment provided in the embodiment of the present application, the real-time detection device 10 for processing quality of processing equipment obtains the sound data of the processing equipment 11, and detects the processing state of the processing equipment 11 according to the sound data of the processing equipment 11. Compared with the traditional manual sampling method, the real-time detection method for processing quality of processing equipment provided in the embodiment of the present application can perform non-contact real-time detection of the processing equipment 11, improve the detection efficiency, reduce the loss of defective raw materials and labor costs, and improve the processing quality of the processing equipment 11.

[0077] Figure 2 Flow chart of the method for real-time detection of processing quality of processing equipment provided in the embodiment of the present application. The method for real-time detection of processing quality of processing equipment provided in the embodiment of the present application is applied to the real-time detection of processing quality of processing equipment (see Figure 1 Real-time detection equipment for processing quality of processing equipment 10).

[0078] 201, the main control analysis module obtains the sound data of the processing process of the processing equipment, and marks it to form a sound data set.

[0079] The sound data set is the sound data of the processing status of the corresponding processing equipment that has been labeled. The sound data set includes positive sample sound data and negative sample sound data. The positive sample sound data is the sound data when the processing equipment is processing normally, and the negative sample sound data is the sound data when the processing equipment is processing abnormally.

[0080] When the processing equipment is processing normally, the sound data acquisition module can be controlled to collect sound data and mark the corresponding processing status of the processing equipment to obtain positive sample sound data. When the processing equipment is processing abnormally, the sound data acquisition module can be controlled to collect sound data and mark the corresponding processing status of the processing equipment to obtain negative sample sound data.

[0081] 202, the main control analysis module extracts sound features from the sound data set to obtain sound samples, and divides the sound samples into a training set and a verification set.

[0082] The sound features extracted from sound data (such as a sound data set) may include time series features, and may also include transform domain features such as linear prediction coefficient (Linear Prediction Coefficient, LPC), linear prediction cepstrum coefficient (Linear Prediction Cepstrum Coefficient, LPCC), Mel Frequency Cepstrum Coefficient (Mel Frequency Cepstrum Coefficient, MFCC), and perceptual linear prediction (Perceptual Linear Predictive, PLP).

[0083] In one embodiment of the present application, MFCC features are extracted from sound data (e.g., a sound data set), specifically including: performing a discrete Fourier transform (DFT, such as a fast Fourier transform) on the sound data to obtain a spectrum of the sound data; calculating a discrete energy spectrum of the sound data based on the spectrum of the sound data; passing the discrete energy spectrum of the sound data through a triangular filter group uniformly distributed on the Mel frequency; performing a logarithmic operation on the output of the triangular filter group to obtain a logarithmic power spectrum of the sound data; performing a discrete cosine transform (DCT) on the logarithmic power spectrum of the sound data to obtain MFCC features.

[0084] In one embodiment of the present application, when performing discrete Fourier transform on sound data, the number of Fourier transform points is 1024.

[0085] In the embodiment of the present application, the sound feature corresponding to one frame of sound data can be used as a sound sample, or the sound features corresponding to multiple frames of sound data can be used as a sound sample. If the sound feature corresponding to every n frames of sound data is used as a sound sample, then m sound samples can be obtained per second of sound data:

[0086]

[0087] in, Indicates rounding down, fs is the sampling frequency, n_fft is the number of Fourier transform points per frame, n is the number of frames corresponding to each sound sample, and n is an integer greater than or equal to 1.

[0088] For example, if the number of Fourier transform points per frame is 1024 and n is 5, then

[0089] In one embodiment of the present application, the sampling frequency is 16 kHz.

[0090] In the embodiment of the present application, before extracting sound features from sound data (eg, a sound data set), the main control analysis module pre-processes the sound data.

[0091] Preprocessing the sound data may include pre-emphasizing the sound data.

[0092] The purpose of pre-emphasis is to enhance the high-frequency components in the sound data and make the signal spectrum flat. Pre-emphasis is generally implemented using a high-pass filter, and the transfer function of the high-pass filter can be:

[0093] H(z)=1-κz -1 , 0.9≤κ≤1.0.

[0094] Here, κ is the pre-emphasis coefficient, and the preferred value is between 0.94 and 0.97.

[0095] Preprocessing the sound data may also include dividing the sound data into frames.

[0096] Framing is achieved by windowing the sound data. In order to avoid excessive changes in two adjacent frames, the frames need to overlap. In one embodiment of the present application, each frame is 25 milliseconds, and there is a 15 millisecond overlap between two adjacent frames, that is, a frame is taken every 10 milliseconds.

[0097] Commonly used window functions include rectangular window, Hamming window and Hanning window.

[0098] Preprocessing the sound data may also include obtaining valid sound data from the sound data.

[0099] The purpose of obtaining effective sound data from sound data is to remove invalid sounds from the sound data, so as to reduce the computational complexity and accuracy of feature extraction, shorten the time of sound recognition, and improve the recognition rate. Effective sound detection can be performed based on the short-time energy and short-time zero-crossing rate of sound data.

[0100] In another embodiment, valid sound data may be obtained from sound data by the following method: framing the sound data to obtain sound frames of the sound data;

[0101] Perform discrete Fourier transform on the sound frame to obtain the spectrum of the sound frame;

[0102] Calculate the cumulative energy of each frequency band according to the spectrum of the sound frame;

[0103] Performing logarithmic operation on the accumulated energy of each frequency band to obtain the logarithmic value of the accumulated energy of each frequency band;

[0104] The accumulated energy logarithm of each frequency band is compared with a preset threshold to obtain valid sound data.

[0105] 203, the main control analysis module trains the dual-channel adaptive feature fusion autoencoder network using the training set, verifies the performance of the dual-channel adaptive feature fusion autoencoder network using the verification set, obtains the real-time detection model of the processing quality of the processing equipment, and determines the processing anomaly threshold of the real-time detection model of the processing quality of the processing equipment.

[0106] The input of the dual-channel adaptive feature fusion autoencoder network is the initial sound feature (e.g., sound sample), and the output is the reconstructed sound feature (e.g., reconstructed sound sample). Training the dual-channel adaptive feature fusion autoencoder network is to minimize the difference between the input (i.e., initial sound sample) and the output (i.e., reconstructed sound sample) of the dual-channel adaptive feature fusion autoencoder network.

[0107] In one embodiment of the present application, the dual-channel adaptive feature fusion autoencoder network includes a time series feature compression module (which can be recorded as G), a frequency domain feature compression module (which can be recorded as F), and a feature fusion module (which can be recorded as ) and the decoder module (which can be recorded as The time series feature compression module is used to compress and map (i.e., encode) the time series features in the input sound features to obtain a time series feature vector (which can be recorded as z t The frequency domain feature compression module is used to compress and map the frequency domain features (such as MFCC features) in the input sound features to obtain the frequency domain feature vector (which can be recorded as z fThe feature fusion module is used to determine the weight α of the time series feature vector (the frequency domain feature weight is 1-α) according to the time series feature vector and the frequency domain feature vector. The decoder module is used to t ) and the weighted frequency domain eigenvector (i.e. (1-α)·z f ) is decoded to obtain the reconstructed sound features (which can be recorded as ).

[0108] The feature fusion module determines the weight α of the time series feature vector based on the time series feature vector and the frequency domain feature vector, which can be expressed as:

[0109]

[0110] Among them, [,] represents matrix concatenation, is the network parameter of the feature fusion module, and σ is the sigmoid function.

[0111] The decoder module is based on the weighted temporal feature vector (i.e., α·z t ) and the weighted frequency domain eigenvector (i.e. (1-α)·z f ) is decoded, and the reconstructed sound features can be expressed as:

[0112]

[0113] in, are the network parameters of the decoder module.

[0114] Figure 3 It is a schematic diagram of a real-time detection model for processing quality of processing equipment used in an embodiment of the present application.

[0115] like Figure 3 As shown, the sound feature z (for example, each sound sample in the training set) is simultaneously input into the time series feature compression module and the frequency domain feature compression module. The time series feature compression module compresses and maps the time series features in the sound features to obtain the time series feature vector z t The frequency domain feature compression module compresses and maps the frequency domain features in the sound features to obtain the frequency domain feature vector z f . Time series feature vector z t and frequency domain eigenvector z f Input the feature fusion module to obtain the weight α of the time series feature vector. Time series feature vector z t and frequency domain eigenvector z f The input is weighted and then fed into the decoder module to obtain the reconstructed sound features. According to the reconstructed sound characteristics and sound features z can calculate the anomaly score ( Figure 3 The anomaly score is the reconstructed sound feature and the square of the difference between the sound feature z).

[0116] In one embodiment of the present application, the time series feature compression module is a bidirectional recurrent network (for example, a bidirectional gated recurrent unit (BiGRU) network), the frequency domain feature compression module is a fully connected encoding network, the feature fusion module is a gated compression network, and the decoder module is a fully connected decoding network.

[0117] It should be understood that in other embodiments of the present application, the time series feature compression module, the frequency domain feature compression module, the feature fusion module, and the decoder module may be other types of networks.

[0118] When the training set is used to train the real-time detection model of processing quality of processing equipment, the loss function can be defined as the mean square error, and the training algorithm can adopt the RMSprop algorithm. RMSprop is an improved stochastic gradient descent algorithm.

[0119] In the embodiment of the present application, each sound sample in the training set can be input into the real-time detection model of the processing quality of the processing equipment to obtain the abnormal score of each sound sample in the training set; the Γ distribution of the abnormal scores of all sound samples in the training set is statistically calculated; the abnormal score at the percentage function (i.e., ppf function) β∈(0,1) of the Γ distribution is taken as the processing abnormality threshold. If the abnormal score is greater than the processing abnormality threshold, it is determined that the processing of the processing equipment is abnormal. If the abnormal score is less than or equal to the processing abnormality threshold, it is determined that the processing of the processing equipment is normal.

[0120] The performance of the real-time detection model for processing quality of processing equipment is verified by using the verification set, that is, the sound samples in the verification set are input into the trained real-time detection model for processing quality of processing equipment to verify whether the trained real-time detection model for processing quality of processing equipment can correctly detect the processing status of the processing equipment. If the trained real-time detection model for processing quality of processing equipment cannot correctly detect the processing status of the processing equipment, the training set is used again to train the real-time detection model for processing quality of processing equipment.

[0121] 204 , the main control analysis module obtains in real time the sound data of the processing equipment during processing (hereinafter referred to as the sound data to be tested).

[0122] When performing actual detection, the main control analysis module may send a data collection command to the sound data collection module. In response to the data collection command, the sound data collection module collects the sound data to be tested and returns the sound data to be tested to the main control analysis module.

[0123] Alternatively, the sound data acquisition module can automatically send the sound data to be tested collected in real time during processing by the processing equipment to the main control analysis module at a preset frequency (for example, once every 5 seconds).

[0124] The main control analysis module can store the received sound data to be tested in a memory built into the main control analysis module, or in a memory external to the main control analysis module.

[0125] 205 , the main control analysis module extracts sound features from the sound data to be tested, and obtains the sound features to be tested (ie, initial sound features to be tested).

[0126] As mentioned above, the sound features extracted from the sound data to be measured may include time series features, and may also include transform domain features such as LPC, LPCC, MFCC, and PLP.

[0127] 206 , the main control analysis module inputs the sound feature to be measured into the real-time detection model of processing quality of the processing equipment, and outputs a reconstructed sound feature to be measured.

[0128] When training and verifying the real-time detection model for processing quality of processing equipment, if the sound features corresponding to one frame of sound data are used as a sound sample, then in actual detection, the sound features to be tested corresponding to one frame of sound data to be tested are input into the real-time detection model for processing quality of processing equipment each time. When training and verifying the real-time detection model for processing quality of processing equipment, if the sound features corresponding to multiple frames (for example, 5 frames) of sound data are used as a sound sample, then in actual detection, the sound features to be tested corresponding to multiple frames of sound data to be tested are input into the real-time detection model for processing quality of processing equipment each time.

[0129] The real-time detection model of processing quality of processing equipment is a trained dual-channel adaptive feature fusion autoencoder network, which includes a time feature compression module, a frequency domain feature compression module, a feature fusion module and a decoder module, wherein the time feature compression module is used to compress and map the time features in the initial sound features to be tested to obtain a time feature vector; the frequency domain feature compression module is used to compress and map the frequency domain features in the initial sound features to be tested to obtain a frequency domain feature vector; the feature fusion module is used to determine the weight of the time feature vector according to the time feature vector and the frequency domain feature vector; the decoder module is used to decode the time feature vector weighted according to the weight and the weighted frequency domain feature vector to obtain a reconstructed sound feature to be tested.

[0130] The specific process of the real-time detection model of the processing quality of the processing equipment inputting the initial sound feature to be measured and outputting a reconstructed sound feature to be measured can be referred to the relevant description of 203, which will not be repeated here.

[0131] 207 , the main control analysis module calculates the abnormality scores of the reconstructed sound feature to be tested and the initial sound feature to be tested, and compares the abnormality scores of the reconstructed sound feature to be tested and the initial sound feature to be tested with the processing abnormality threshold.

[0132] The abnormality score between the reconstructed sound feature to be measured and the initial sound feature to be measured may be the square of the difference between the reconstructed sound feature to be measured and the initial sound feature to be measured.

[0133] 208 , if the abnormality scores of the reconstructed sound feature to be tested and the initial sound feature to be tested are greater than the processing abnormality threshold, the main control analysis module outputs device abnormality prompt information.

[0134] The main control analysis module can send out device abnormality prompt information through sound, text, etc.

[0135] If the anomaly scores of the reconstructed sound feature to be tested and the initial sound feature to be tested are greater than the processing anomaly threshold, the main control analysis module can control the processing equipment to stop running.

[0136] In one embodiment of the present application, the main control analysis module determines whether the abnormality scores of the sound features to be tested and the initial sound features to be tested reconstructed for a preset number of times (for example, 3 times) are greater than the processing abnormality threshold. If the abnormality scores of the sound features to be tested and the initial sound features to be tested reconstructed for a preset number of times are greater than the processing abnormality threshold, the processing equipment is controlled to stop running and an alarm message of equipment abnormality is sent to the user. The communication warning module can send an alarm message of equipment abnormality to a preset user via email, SMS, or instant messaging tools.

[0137] Figure 2 In the illustrated embodiment, the main control analysis module obtains the sound data of the processing equipment, and detects the processing status of the processing equipment based on the sound data of the processing equipment. Sound data is an important source of information reflecting the processing status of the processing equipment, and sound data has the advantages of convenient collection, non-contact measurement, and fast processing speed. Detecting the processing status of the processing equipment by identifying the processing operation sound of the processing equipment has broad application prospects. Compared with the traditional manual sampling method, the real-time detection method for the processing quality of the processing equipment provided in the embodiment of the present application can perform contactless real-time detection of the processing equipment 11, improve detection efficiency, reduce the loss of defective production materials and labor costs, and improve the processing quality of the processing equipment 11.

[0138] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and that the present application can be implemented in other specific forms without departing from the spirit or basic features of the present application. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the present application is limited by the attached claims rather than the above description, so it is intended to include all changes that fall within the meaning and scope of the equivalent elements of the claims in the present application. Any attached figure mark in the claims should not be regarded as limiting the claims involved. In addition, it is obvious that the word "comprising" does not exclude other modules or steps, and the singular does not exclude the plural. Multiple modules or devices stated in the system claim can also be implemented by one module or device through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any particular order.

Claims

1. A real-time detection method for processing quality of processing equipment, characterized in that: The method comprises: Acquire the sound data of processing equipment in real time; Extracting sound features from the sound data to obtain initial sound features; The initial sound feature is input into the real-time detection model of processing quality of processing equipment, and a reconstructed sound feature is output. The real-time detection model of processing quality of processing equipment is a trained dual-channel adaptive feature fusion autoencoder network, and the dual-channel adaptive feature fusion autoencoder network includes a time feature compression module, a frequency domain feature compression module, a feature fusion module and a decoder module, wherein: The time series feature compression module is used to compress and map the time series features in the initial sound features to obtain a time series feature vector; The frequency domain feature compression module is used to compress and map the frequency domain features in the initial sound features to obtain a frequency domain feature vector; The feature fusion module is used to determine the weight of the time series feature vector according to the time series feature vector and the frequency domain feature vector; The decoder module is used to decode the time series feature vector weighted according to the weight value and the weighted frequency domain feature vector to obtain a reconstructed sound feature; Calculating anomaly scores of the reconstructed sound features and the initial sound features, and comparing the anomaly scores of the reconstructed sound features and the initial sound features with a processing anomaly threshold of the processing quality real-time detection model of the processing equipment; If the abnormality scores of the reconstructed sound feature and the initial sound feature are greater than the processing abnormality threshold, device abnormality prompt information is output.

2. The real-time detection method for processing quality of processing equipment according to claim 1, characterized in that: The temporal feature compression module is a bidirectional recurrent network, the frequency domain feature compression module is a fully connected encoding network, the feature fusion module is a gated compression network, and the decoder module is a fully connected decoding network.

3. The real-time detection method for processing quality of processing equipment according to claim 1, characterized in that: The training method of the dual-channel adaptive feature fusion autoencoder network includes: Acquiring sound data of a processing process of the processing equipment, and marking the sound data set; Extracting sound features from the sound data set to obtain sound samples, and dividing the sound samples into a training set and a validation set; The dual-channel adaptive feature fusion autoencoder model is trained using the training set, and the performance of the dual-channel adaptive feature fusion autoencoder model is verified using the verification set to obtain the real-time detection model for processing quality of the processing equipment.

4. The real-time detection method for processing quality of processing equipment according to claim 3, characterized in that: The calculation method of the processing abnormality threshold includes: Inputting each sound sample in the training set into the real-time detection model for processing quality of processing equipment to obtain an abnormality score for each sound sample in the training set; Statistically calculating the Γ distribution parameters of the abnormal scores of all sound samples in the training set; The abnormal score at the Γ distribution percentage function β∈(0,1) is taken as the processing abnormality threshold.

5. The real-time detection method for processing quality of processing equipment according to claim 3, characterized in that: The extracting sound features from the sound data set comprises: Performing discrete Fourier transform on the sound data set to obtain a frequency spectrum of the sound data set; Calculating a discrete energy spectrum of the sound data set according to the frequency spectrum of the sound data set; Passing the discrete energy spectrum of the sound data set through a triangular filter bank uniformly distributed on Mel frequencies; Performing a logarithmic operation on the output of the triangular filter bank to obtain a logarithmic power spectrum of the sound data set; Perform discrete cosine transform on the logarithmic power spectrum of the sound data set to obtain Mel-frequency cepstral coefficient features.

6. The real-time detection method for processing quality of processing equipment according to claim 5, characterized in that: Taking the sound features corresponding to each n frames of the sound data set as a sound sample, the number of sound samples obtained by the sound data set per second is: in, Indicates rounding down, fs is the sampling frequency, n_fft is the number of Fourier transform points per frame, n is the number of frames corresponding to each sound sample, and n is an integer greater than or equal to 1.

7. The real-time detection method for processing quality of processing equipment according to claim 3, characterized in that: The training of the dual-channel adaptive feature fusion autoencoder model using the training set includes: Input each sound sample in the training set into the time series feature compression module and the frequency domain feature compression module simultaneously; The time series feature compression module compresses and maps the time series features in the sound sample to obtain a time series feature vector of the sound sample; The frequency domain feature compression module compresses and maps the frequency domain features in the sound sample to obtain a frequency domain feature vector of the sound sample; The time series feature vector and the frequency domain feature vector are input into the feature fusion module to obtain the weight of the time series feature vector of the sound sample.

8. The real-time detection method for processing quality of processing equipment according to claim 1, characterized in that: The weight of the time series feature vector is determined according to the following formula: Among them, α is the weight of the time series feature vector, represents the feature fusion module, [,] represents matrix concatenation, z t represents the time series feature vector, z f represents the frequency domain eigenvector, is the network parameter of the feature fusion module, and σ is the sigmoid function.

9. The real-time detection method for processing quality of processing equipment according to claim 1, characterized in that: The loss function used in the training of the dual-channel adaptive feature fusion autoencoder network is a mean square error function, and the anomaly score is the square of the difference between the reconstructed sound feature and the initial sound feature.

10. The real-time detection method for processing quality of processing equipment according to claim 1, characterized in that: Before extracting the sound features from the sound data, the method further includes: Pre-emphasize the sound data; and / or Framing the sound data; and / or Acquire effective sound data from the sound data.

11. The real-time detection method for processing quality of processing equipment according to claim 1, characterized in that: The method further comprises: Determine whether the abnormality scores of the reconstructed sound feature and the initial sound feature are greater than the processing abnormality threshold for a preset number of consecutive times; If the abnormality scores of the reconstructed sound feature and the initial sound feature are greater than the processing abnormality threshold for a preset number of consecutive times, the processing equipment is controlled to stop running, and an alarm message of equipment abnormality is sent to the user.

12. A real-time detection device for processing quality of processing equipment, characterized in that: It includes a sound data acquisition module, a main control analysis module and a communication warning module, wherein: The sound collection module is used to collect the sound data of the processing equipment in real time during processing; The main control analysis module includes: The data analysis unit is used to extract sound features from the sound data to obtain initial sound features, input the initial sound features into a real-time detection model for processing quality of a processing equipment, and output a reconstructed sound feature. The real-time detection model for processing quality of a processing equipment is a trained dual-channel adaptive feature fusion autoencoder network. The dual-channel adaptive feature fusion autoencoder network includes a time series feature compression module, a frequency domain feature compression module, a feature fusion module, and a decoder module, wherein: the time series feature compression module is used to compress and map the time series features in the initial sound features to obtain a time series feature vector; the frequency domain feature compression module is used to compress and map the frequency domain features in the initial sound features to obtain frequency domain feature vector; the feature fusion module is used to determine the weight of the time series feature vector according to the time series feature vector and the frequency domain feature vector; the decoder module is used to decode the time series feature vector weighted according to the weight and the weighted frequency domain feature vector to obtain the reconstructed sound feature; the data analysis unit is also used to calculate the abnormality score of the reconstructed sound feature and the initial sound feature, and compare the abnormality score of the reconstructed sound feature and the initial sound feature with the processing abnormality threshold of the real-time detection model of the processing quality of the processing equipment, if the abnormality score of the reconstructed sound feature and the initial sound feature is greater than the processing abnormality threshold, output the equipment abnormality prompt information; A data storage unit, used for storing the sound data and the real-time detection model of processing quality of the processing equipment; A display unit, used to display abnormal prompt information of the device; A control unit, used to control the operation of the data analysis unit, the data storage unit and the display unit; The communication warning module is used to send a control signal to the processing equipment to stop the operation of the processing equipment when the abnormality scores of the reconstructed sound feature and the initial sound feature are greater than the processing abnormality threshold.

Images