Method for detecting overloading of a truck based on acoustic signals

Abstract

The application discloses a truck overload detection method based on sound signals, specifically comprising the following steps: collecting original truck driving sounds of a target model, and preprocessing original truck driving sound audio signal data to extract independent truck driving sound signals; performing feature extraction on the truck driving sound signals; constructing a truck driving sound model, wherein the truck driving sound model comprises a target model truck driving sound model and an overload detection model, and the extracted features are taken as input to train a self-encoder network; and inputting truck driving sound signals to be detected into the trained self-encoder network to perform truck model detection and overload detection. The unsupervised self-encoder has strong feature learning and reconstruction capabilities, and the network structure is relatively simple, so the self-encoder is very suitable for sound detection tasks.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of sound signal processing, and particularly relates to a truck overload detection method based on sound signals. BACKGROUND

[0002] At present, the highways in China are in an important period of transition from construction to management, but the phenomenon of truck overload has been going on despite repeated prohibitions and has caused great harm. In China, overload control is a difficult problem, and the most widely used manual and static weighing detection has many defects such as poor reliability and low efficiency, while the existing dynamic weighing overload control system has high installation and maintenance costs and cannot be fully applied on national and provincial trunk lines. The above two types of contact overload detection methods are time-consuming and laborious, and do not match the current trend of intelligent information technology.

[0003] Among the non-contact overload detection technologies, there is currently a truck overload detection method based on machine vision, but the image or video quality is easily limited by environmental light and obstacle shielding, and when the weather is dark or the target is shielded, the performance of overload detection by machine vision will be greatly impaired. SUMMARY

[0004] In order to solve the above technical defects in the prior art, the present application proposes a truck overload detection method based on sound signals.

[0005] The technical solution for achieving the purpose of the present application is as follows: a truck overload detection method based on sound signals, comprising the following specific steps:

[0006] Collecting the original truck driving sound of the target model, and pre-processing the original truck driving sound audio signal data to extract independent truck driving sound signals;

[0007] Extracting features from the truck driving sound signals;

[0008] Constructing a truck driving sound model, wherein the truck driving sound model comprises a target model truck driving sound model and an overload detection model, and the extracted features are used as input to train an autoencoder network;

[0009] Inputting the truck driving sound signals to be detected into the trained autoencoder network for truck model detection and overload detection.

[0010] Preferably, the obtained audio signal data is pre-processed using a sparse constraint non-negative matrix factorization method, and the specific process is as follows:

[0011] Step 1.1, constructing the obtained original vehicle driving sound signal matrix V = WH, wherein W and H are respectively the truck driving sound signal and the background noise signal factor matrix of the target truck;

[0012] Step 1.2, update the truck running sound signal factor matrix W of the target truck by KL divergence bias, the specific formula is:

[0013]

[0014] In the formula, t represents the iteration number, i and j are the row number and column number of the matrix, l and k are parameters for iteration by KL divergence bias, V ij is the value of the observation matrix corresponding to the (i, j) subscript, W ik is the value of the source signal matrix W corresponding to the (i, k) subscript, H kj is the value of the noise matrix H corresponding to the (k, j) subscript, W il is the value of the source signal matrix W corresponding to the (i, l) subscript, H lj is the value of the noise matrix H corresponding to the (l, j) subscript.

[0015] Step 1.3, normalize the columns of W according to the following formula, and set the negative elements in W and H to zero:

[0016]

[0017] In the formula, W ik t+1 is the matrix value before normalization of the t+1 iteration, W ik ' t+1 is the value after normalization.

[0018] Step 1.4, update H:

[0019]

[0020] In the formula, a is the weight.

[0021] Step 1.5: calculate the objective function, if the value of the objective function is greater than the specified separation index, return to step 1.2, otherwise proceed to step 2.

[0022] Preferably, the objective function is specifically:

[0023]

[0024] Preferably, the specific method for feature extraction of the truck running sound signal is: performing fast Fourier transform on each frame of signal, converting the time domain signal to the frequency domain, and calculating the energy spectrum; performing Mel filtering on each frame of energy spectrum in the frequency domain, and converting the linear energy spectrum to a logarithmic energy spectrum to obtain the features.

[0025] Preferably, the specific method for feature extraction of the truck running sound signal is:

[0026] Step 2.1, FFT is performed on each frame of signal to convert time domain signal to frequency domain:

[0027]

[0028] wherein x(q) is the input signal, N is the number of FFT points, L is the frame length, n represents the number of signal frames, and p represents the serial number of spectral lines in the frequency domain;

[0029] Step 2.2, the energy of spectral lines is calculated for each frame of FFT data:

[0030] E n (p)=[X n (p)] 2 0≤p<L

[0031] Step 2.3, the energy spectrum is filtered through a Mel filter and summed to obtain the output energy:

[0032]

[0033] In the frequency domain, multiply each frame of energy spectrum with the frequency domain response of the Mel filter and add them up, m represents the serial number of the Mel filter, and M is the total number of Mel filters;

[0034] Step 2.4, take the logarithm of the energy of the Mel filter to convert the linear energy spectrum to the logarithmic energy spectrum, and obtain the feature to be extracted - the logarithmic Mel energy:

[0035] X(n, m) = log[S(n, m)] 0≤m<M.

[0036] Preferably, the specific steps for constructing the driving sound model of the target model truck are as follows:

[0037] Step 3.1, input the features X of the driving sound of the target model truck, pass through the encoder, and obtain the latent features D:

[0038] D = g(GX + b e )

[0039] wherein G is the weight matrix of the encoding layer and the input layer, b e is the node bias of the encoding layer, and g(.) is the node activation function;

[0040] Step 3.2, pass through the decoder to obtain the reconstruction of the input sample, wherein b d is the decoding layer bias:

[0041]

[0042] Step 3.3, calculate the loss function by the mean square error function of the reconstructed features and the input features:

[0043]

[0044] wherein x r is the input vector of the rth dimension, is the reconstruction vector of the rth dimension, n is the dimension of the feature matrix;

[0045] Step 3.4, the optimization algorithm adopts the gradient descent algorithm, and the weight and bias are updated to reduce the loss function, assuming that in the tth round of training, the weight G update formula is:

[0046]

[0047] The bias b update formula is:

[0048]

[0049] wherein η is the learning rate;

[0050] Step 3.5, judge whether the reconstruction error is less than the expected error If not, repeat steps 3.2 to 3.4, otherwise end the process.

[0051] Preferably, the construction process of the overload detection model is the same as that of the target model truck driving sound model, and the input is the feature of the normal load sample of the truck.

[0052] Preferably, the truck driving sound signal to be detected is input into the trained autoencoder network, and the specific method for truck model detection and overload detection is:

[0053] Step 4.1, input the sound feature to be detected into the trained target model truck sound autoencoder model, and obtain the reconstruction error relative to the target model truck;

[0054] Step 4.2, take the reconstruction error as an anomaly score, and take the expected error As the threshold of the anomaly score, the threshold value of the anomaly score S is processed, wherein 0 represents that the threshold processing result is that the anomaly score is within the normal range, and 1 represents the anomaly:

[0055]

[0056] Step 4.3, input the sample of the detection result belonging to the target vehicle into the trained truck overload detection model for overload detection, obtain the reconstruction error, and perform threshold judgment, and the sound sample detected as abnormal is determined as the sound sample of the overloaded truck, otherwise it is considered as not overloaded.

[0057] Compared with the prior art, the present application has the following advantages:

[0058] First, compare static weighing and dynamic weighing, using sound detection overload simple and fast, low cost and high efficiency.

[0059] Then compare the overload detection based on machine vision, sound signal is not affected by light and obstacles, low cost and convenient installation.

[0060] Finally, the unsupervised autoencoder has strong feature learning and reconstruction ability, and the network structure is relatively simple, which is very suitable for sound detection task.

[0061] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be achieved and obtained by the structure specifically pointed out in the written description, claims, and accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0062] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application.

[0063] Figure 1 is a flowchart of the present application.

[0064] Figure 2 is a model training process of the present application taking autoencoder as an example. DETAILED DESCRIPTION

[0065] It is easy to understand that according to the technical scheme of the present application, a person skilled in the art can imagine various embodiments of the present application without changing the essential spirit of the present application. Therefore, the following specific embodiments and drawings are only exemplary description of the technical scheme of the present application, and should not be regarded as the whole or regarded as the limitation or limitation of the technical scheme of the present application. On the contrary, the purpose of providing these embodiments is to make the person skilled in the art understand the present application more thoroughly. The preferred embodiments of the present application are described in detail below in conjunction with the drawings, which constitute a part of this application, and are used to explain the innovative ideas of the present application together with the embodiments of the present application.

[0066] The present application is a two-stage truck driving sound detection to realize overload detection, the first stage is the driving sound detection of target model truck, the second stage is the overload detection of the model truck. The first stage constructs a model of the features of all truck driving sound samples belonging to the target model through the encoding and decoding process of the autoencoder, and uses it to detect the vehicle type of the test sample. The second stage constructs a model of the features of the normal load driving sound samples of the target model truck through the autoencoder, and uses it to detect the overload of the test sample.

[0067] As an embodiment, a truck overload detection method based on acoustic signals is provided. First, the driving sound of a target truck model is collected, and blind source separation is performed on the obtained audio signal data. A sparse constraint non-negative matrix factorization algorithm is used to separate and restore the source signal of the truck driving sound from the observation signal by adding a sparse constraint condition to the source signal and the mixing matrix.

[0068] Then, the features are extracted. Fast Fourier transform (FFT) is performed on each frame of signal to convert the time domain signal to the frequency domain, and the energy spectrum is calculated. A triangular bandpass filter is designed to perform Mel filtering on each frame of energy spectrum in the frequency domain, and the sum is taken to obtain the logarithm. The linear energy spectrum is converted to the logarithmic energy spectrum to obtain the feature: the logarithmic Mel energy.

[0069] Then, the extracted features are used as input to train the autoencoder network, and a two-stage truck driving sound detection autoencoder model is constructed. The first stage is the driving sound detection of the target truck model, and the second stage is the overload detection of the truck model.

[0070] The modeling process of the first stage is as follows: the feature matrix of the driving sound samples of the target truck model is input into the autoencoder, and the input features are reduced and reconstructed through the encoding and decoding process. The network parameters are adjusted using the back propagation (BP) algorithm and the optimization algorithm to minimize the reconstruction error, which represents that the model has learned the most effective features of the input data, and a driving sound model of the target truck model is constructed.

[0071] The modeling process of the second stage is as follows: the feature matrix of the normal load driving sound samples of the truck model in the first stage is input into the autoencoder, and the reconstruction error is minimized through the network calculation of the autoencoder to obtain the truck overload detection model.

[0072] Finally, the overload detection is performed. The sample to be tested is preprocessed and the features are extracted, and the truck driving sound detection is performed in two stages to determine whether the sample belongs to the target truck model and whether it is overloaded.

[0073] In the first stage, the truck model detection is performed. The feature matrix is input into the trained driving sound model of the target truck model to obtain the reconstruction error score. Whether the sample to be tested belongs to the target truck model is determined by the threshold condition of the reconstruction error. If it belongs to the target truck model, the second stage is entered, otherwise the overload detection cannot be completed and the detection process is ended.

[0074] In the second stage, the overload detection is performed. The feature matrix of the target truck model is input into the trained truck overload detection model. Whether the sample to be tested belongs to the normal load sample is determined by the threshold condition of the reconstruction error. If it belongs to the normal load sample, it is not overloaded, otherwise it is determined as an overload sample.

[0075] As an embodiment of the present application, a two-stage truck driving sound detection is used to realize the overload detection method, and the specific steps are as follows:

[0076] Step 1: Blind source separation process, using a sparse constraint based non-negative matrix factorization method:

[0077] Step 1.1, in the non-negative matrix factorization, the original vehicle driving sound signal matrix V = WH has been obtained, wherein W and H are factor matrices, that is, the sound signal of the target truck to be separated and restored and the background noise signal.

[0078] Step 1.2, first update W iteratively, wherein t represents the number of iterations, i and j are the row number and column number of the matrix, and l and k are parameters for iteration by KL divergence deviation, V ij is the value of the observation matrix corresponding to the (i, j) subscript, W ik is the value of the source signal matrix W corresponding to the (i, k) subscript, H kj is the value of the noise matrix H corresponding to the (k, j) subscript, W il is the value of the source signal matrix W corresponding to the (i, l) subscript, H lj is the value of the noise matrix H corresponding to the (l, j) subscript.

[0079]

[0080] Step 1.3, normalize W by column according to the following formula, and set negative elements in W and H to zero, wherein W ik t+1 is the matrix value before normalization of the t+1 iteration, W ik ' t+1 is the value after normalization.

[0081]

[0082] Step 1.4, update H, wherein a is a weight:

[0083]

[0084] Step 1.5, substitute W and H to calculate the objective function,

[0085]

[0086] If the value of the objective function is greater than the specified separation index, return to step 1.2, otherwise end the process. At this time, W is the truck source signal without background noise.

[0087] Step 2: Audio signal feature extraction process, taking log mel energy as an example:

[0088] Step 2.1, FFT is performed on each frame of signal to convert time domain signal to frequency domain:

[0089]

[0090] where x(q) is the input signal, N is the number of FFT points, L is the frame length, n represents the number of signal frames, and p represents the serial number of the spectral line in the frequency domain;

[0091] Step 2.2, the energy of the spectral line is calculated for each frame of FFT data:

[0092] E n (p)=[X n (p)] 2 0≤p<L

[0093] Step 2.3, the energy spectrum is filtered through a Mel filter and summed to obtain the output energy:

[0094]

[0095] In the frequency domain, multiply each frame of energy spectrum with the frequency domain response of the Mel filter and add them up, where m represents the serial number of the Mel filter and M is the total number of Mel filters.

[0096] Step 2.4, take the logarithm of the energy of the Mel filter to convert the linear energy spectrum to the logarithmic energy spectrum, obtaining the feature to be extracted - the logarithmic Mel energy:

[0097] X(n, m) = log[S(n, n)] 0≤m<M

[0098] Step 3: Construct a two-stage truck driving sound model, including a target model truck driving sound model and an overload detection model, first construct the former:

[0099] Step 3.1, input the features X of the target model truck driving sound, pass through the encoder to obtain the latent features D:

[0100] D = g(GX + b e )

[0101] where G is the weight matrix of the encoding layer and the input layer, be is the node bias of the encoding layer, and the node activation function g(.) is selected as RELU.

[0102] Step 3.2, pass through the decoder to obtain the reconstruction of the input sample, where bd is the decoding layer bias:

[0103]

[0104] Step 3.3, the loss function is calculated by reconstructing the features and the mean square error function of the input features, and the loss function here is the reconstruction error of the sample through the model. Among them, x r is the input vector of the rth dimension, is the reconstructed vector of the rth dimension, and n is the dimension of the feature matrix:

[0105]

[0106] Step 3.4, the optimization algorithm uses gradient descent algorithm to update the weight and bias to reduce the loss function. Assuming that in the tth round of training, the weight G update formula is:

[0107]

[0108] The bias b update formula is:

[0109]

[0110] Among them, the learning rate η is set to 0.001.

[0111] Step 3.5, judge whether the reconstruction error is less than the expected error If not, repeat steps 3.2 to 3.4, otherwise end the process, and the first stage model is completed.

[0112] Step 3.6, build the second stage overload detection model, the input is the feature of the normal load sample of the truck, and the steps repeat 3.1 to 3.5, and finally get the truck overload detection model.

[0113] Step 4: two stage truck driving sound detection, including truck model detection and overload detection process:

[0114] Step 4.1, first, the truck model is detected, the sound features of the sample to be detected are input into the trained sound autoencoder model of the target truck model, and the reconstruction error of the sample relative to the target truck model is obtained.

[0115] Step 4.2, the reconstruction error is taken as the anomaly score, and the expected error is taken as the threshold of the anomaly score, and the threshold value of the anomaly score S is processed, where 0 represents that the threshold value processing result is that the anomaly score is within the normal range, and 1 represents the anomaly:

[0116]

[0117] Step 4.3, the detected abnormal sound sample has a large difference with the target model truck sound, and its nature may be generated by other models of trucks; otherwise, the detection result of this sample belongs to the target vehicle, which is regarded as the running sound of the target model truck, and the second stage overload detection is continued. The sample is continued as the input of the overload detection model, and after the reconstruction error is obtained, the threshold judgment of step 4.2 is also carried out. The detected abnormal sound sample has a large difference with the normal load running sound, and it is judged as the sound sample of the overloaded truck, otherwise it is regarded as not overloaded.

[0118] The above merely provides the preferred embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of the changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

[0119] It should be understood that, in order to simplify the present application and help those skilled in the art to understand various aspects of the present application, in the above description of the exemplary embodiments of the present application, various features of the present application are sometimes described in a single embodiment, or with reference to a single figure. However, the present application should not be interpreted as the features included in the exemplary embodiments are all necessary technical features of the patent claims of the present application.

[0120] It should be understood that the modules, units, components, etc. included in the device of one embodiment of the present application can be adaptively changed to be arranged in a device different from the embodiment. Different modules, units or components included in the device of the embodiment can be combined into one module, unit or component, or divided into multiple sub-modules, sub-units or sub-components.

Claims

1. A method for detecting overloading of a truck based on acoustic signals, characterized in that, The specific steps are: Collect the original truck driving sound of the target model, and pretreat the original truck driving sound audio signal data to extract independent truck driving sound signals; Feature extraction is performed on the truck driving sound signals; A truck driving sound model is constructed, which includes a target model truck driving sound model and an overload detection model, both of which use a self-encoder model; The extracted features are used as input to train the self-encoder network, and the truck driving sound signals to be detected are input into the trained self-encoder network for truck model detection and overload detection. The specific method is: Step 4.1, input the sound features to be detected into the trained target model truck driving sound model to obtain the reconstruction error relative to the target model truck; Step 4.2, reconfigure the error as an abnormal score, and the expected error Threshold processing is performed on the abnormal score S, where 0 indicates that the threshold processing result is that the abnormal score is within the normal range, and 1 indicates that the abnormal score is abnormal: Step 4.3, input the samples whose detection results belong to the target vehicle model into the trained truck overload detection model for overload detection to obtain the reconstruction error relative to the target model truck, and perform threshold judgment. The sound sample detected as abnormal is determined as an overloaded truck sound sample, otherwise it is considered as not overloaded.

2. The acoustic signal based truck overload detection method of claim 1, wherein, The obtained audio signal data is pretreated using a sparse constraint non-negative matrix factorization method, and the specific process is: Step 1.1, construct the obtained original vehicle driving sound signal matrix V = WH, where W and H are the truck driving sound signal factor matrix and the background noise signal factor matrix of the target truck respectively; Step 1.2, update the truck driving sound signal factor matrix W of the target truck by KL divergence bias, and the specific formula is: where t denotes the iteration number, l and k are parameters for iteration by KL divergence bias, V ij is the value of the observation matrix corresponding to the (i, j) indices, W ik is the value of the source signal matrix W corresponding to the (i, k) indices, H kj is the value of the noise matrix H corresponding to the (k, j) indices, W il is the value of the source signal matrix W corresponding to the (i, l) indices, H lj is the value of the noise matrix H corresponding to the (l, j) indices; Step 1.3, normalize the columns of W according to the following formula, and set the negative elements in W and H to zero: where W ik t+1 Wt+1is the matrix value before normalization for the t+1 iteration ik ′ t+1 is the normalized value; Step 1.4, update H: In the formula, a is the weight; Step 1.5: calculate the objective function. If the value of the objective function is greater than the specified separation index, return to step 1.2, otherwise proceed to the next step. The objective function is specifically:

3. The acoustic signal based truck overload detection method of claim 1, wherein, The specific method for feature extraction of the truck driving sound signal is: perform fast Fourier transform on each frame of signal to convert the time domain signal to the frequency domain and calculate the energy spectrum; perform Mel filtering on each frame of energy spectrum in the frequency domain, and convert the linear energy spectrum to logarithmic energy spectrum to obtain the features.

4. The acoustic signal based truck overload detection method of claim 1 or 3, wherein, The specific method for feature extraction of the truck driving sound signal is: Step 2.1, perform FFT on each frame of signal to convert the time domain signal to the frequency domain: Where x(q) is the input signal, N is the number of FFT points, L is the frame length, n represents the number of signal frames, and p represents the serial number of the spectral line in the frequency domain; Step 2.2, calculate the energy of each spectral line after FFT: E n (p) = [X n (p)] 2 0≤p<L Step 2.3, sum the energy spectrum through the Mel filter to obtain the output energy: In the frequency domain, multiply each frame of energy spectrum by the frequency domain response of the Mel filter and add them together. m represents the serial number of the Mel filter, and M is the total number of Mel filters; Step 2.4, take the logarithm of the energy of the Mel filter to convert the linear energy spectrum to the logarithmic energy spectrum to obtain the features to be extracted - logarithmic Mel energy: X(n, m) = log[S(n, m)] 0 ≤ m < M.

5. The acoustic signal based truck overload detection method of claim 1, wherein, The specific steps for constructing the target model truck driving sound model are: Step 3.1, input the characteristics X of the target model truck driving sound, pass through the encoder, and get the latent features D: D = g(GX + b e ) where G is the weight matrix of the encoding layer and input layer, b e is the encoding layer node bias, and g(.) is the node activation function. Step 3.

2. Obtain the reconstruction of the input samples from the decoder, where b d is the decoded layer bias: Step 3.3, calculate the loss function by the mean square error function of the reconstructed features and the input features: where x r is an input vector of dimension r, is a reconstruction vector of dimension r, and n1is the dimension of the feature matrix. Step 3.4, the optimization algorithm adopts the gradient descent algorithm to update the weight and bias to reduce the loss function. Assuming that in the tth round of training, the weight G update formula is: The bias b update formula is: Where η is the learning rate; Step 3.

5. Determine if the reconstruction error is less than the desired error If not, repeat steps 3.2 through 3.4, otherwise end the process.

6. The acoustic signal based truck overload detection method of claim 1, wherein, The construction process of the overload detection model is the same as that of the target model truck driving sound model. The input is the characteristics of the normal load sample of the truck.

Images