Intelligent detection method and detection device for empty containers
By using a support vector machine model to process container vibration signals, the problem of complex operating parameters in existing technologies is solved, enabling efficient and intelligent empty container detection that is adaptable to the detection of containers of different types and locations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENZHEN MAXVISION TECH
- Filing Date
- 2023-08-31
- Publication Date
- 2026-06-09
AI Technical Summary
Existing container inspection technologies involve complex determination of working parameters, which affects the inspection results, and are not effective for inspecting containers of different types and locations.
By employing a support vector machine (SVM) model, the vibration signal spectrum of the container in both empty and non-empty states is obtained. Frequency values with non-attenuating amplitudes are removed, and wavelet transform and short-time Fourier transform are performed to establish a decision function and achieve adaptive detection.
No need to predetermine working parameters, improving detection efficiency and effectiveness, adapting to different container types and locations, safe, radiation-free, non-invasive detection, saving labor and time costs.
Smart Images
Figure CN117312967B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of container inspection technology, and in particular to an intelligent method and device for detecting empty containers. Background Technology
[0002] Among existing container inspection technologies, such as the instrument and method for quickly detecting whether a container is empty disclosed in Chinese invention patent publication number CN107621653A (hereinafter referred to as Document 1), through steps one to four, the principle of acoustics is used to quickly identify whether the container is empty or not without opening the container.
[0003] However, Reference 1 has the following shortcomings: In step four, it conducts multiple experiments for different types of containers, different transmitting and receiving locations, and different sound source frequencies to determine the system's operating parameters. However, determining these operating parameters is quite complex and can easily affect the actual detection results. Summary of the Invention
[0004] The purpose of this invention is to provide an intelligent method and device for detecting empty containers, addressing the shortcomings of existing technologies. This method eliminates the need to predetermine operating parameters, making the detection method simpler and easier to operate, thereby improving the detection effect.
[0005] This invention achieves the above objective through the following technical solution: a method for intelligent detection of empty containers, comprising:
[0006] Initialization steps: Obtain the spectrum of vibration signals of the collected container in empty and non-empty states, remove the frequency values in the spectrum that do not decay in amplitude to obtain effective vibration data, perform wavelet transform on the effective vibration data to obtain a sample library of frequency decomposition amplitude curves, extract the differential features in the sample library and obtain the sample feature dataset.
[0007] Modeling steps: Establish a support vector machine (SVM) model with a kernel function, and use the SVM model to solve for the Lagrange multipliers and intercepts of the optimal classification hyperplane in the feature space of the sample feature dataset to obtain the decision function;
[0008] Acquisition steps: Acquire the actual time-domain signal of the composite attenuation generated by the vibration of the container to be detected;
[0009] Denoising step: Perform a short-time Fourier transform on the acquired actual time-domain signal to obtain a spectrum, remove the frequency values in the spectrum that do not decay in amplitude, and obtain the effective time-domain signal;
[0010] Frequency decomposition step: Perform wavelet transform on the effective time-domain signal to obtain amplitude curves of several decomposed frequencies;
[0011] Judgment steps: Extract the difference features of the amplitude curve and input them into the decision function, output the result of the decision function, and determine whether the container to be detected is empty or not based on the result.
[0012] As a further aspect of the present invention: in the modeling step, the kernel function and the decision function are respectively represented as follows:
[0013] Kernel function:
[0014] Where k(x,z) is the Gaussian kernel function, and x and z are vectors of two sample feature datasets, ||xz|| 2 Let x be the square of the distance between vectors x and z, and θ be the variance of the Gaussian function.
[0015] Decision function:
[0016] Here, sign, also called sgn, is the sign function, used to extract the sign of a number, K(x,x) i ) is the kernel function. For Lagrange multipliers, b * x is the intercept. i Let be the vector of the feature dataset of the i-th sample.
[0017] As a further aspect of the present invention: the sample feature dataset includes linearly separable features, the SVM model includes a linear SVM model corresponding to the linearly separable features, and the Lagrange multipliers of the linear SVM model are solved according to the following formula:
[0018]
[0019] st∑ i α i y i =0,α i ≥0,
[0020] In the formula, α i α is the Lagrange multiplier for the linear SVM model. j x is the Lagrange multiplier for the linear SVM model. i For the i-th instance, x j For the j-th instance, y i For x i The class marker, y j For x j The class tag.
[0021] As a further aspect of the present invention: the sample feature dataset includes linearly inseparable features, and the SVM model includes a nonlinear SVM model corresponding to the linearly inseparable features. The Lagrange multipliers of the nonlinear SVM model are solved using the following formula:
[0022]
[0023] st∑ i α i y i =0, 0≤α i ≤C,
[0024] In the formula, K(x) i ,x j Let k(x,z) be the Gaussian kernel function, C be the selected parameters, and α be the α value. i α is the Lagrange multiplier for the nonlinear SVM model. j x is the Lagrange multiplier for the nonlinear SVM model. i For the i-th instance, x j For the j-th instance, y i For x i The class marker, y j For x j The class label; when the kernel function k(x,z) is a positive definite kernel function, the solution of this formula belongs to a convex quadratic programming problem, and its optimal solution is the Lagrange multiplier vector. Let represent the optimal solution for the i-th Lagrange multiplier.
[0025] As a further aspect of the present invention, the intercept is determined by the following steps:
[0026] From the optimal solution of the Lagrange multiplier vector Select a positive component
[0027] The y-th sample feature dataset corresponding to index i i Substitute into the formula:
[0028]
[0029] Solve for b * In the formula, b * The intercept is... For the optimal solution of the i-th Lagrange multiplier, y i This is the data of the feature dataset for the i-th sample.
[0030] As a further aspect of the present invention, the initialization step further includes: normalizing the sample feature dataset to obtain a training set;
[0031] The modeling steps also include: using the SVM model to solve for the Lagrange multipliers and intercepts of the optimal classification hyperplane in the feature space of the training set, thereby obtaining the decision function.
[0032] As a further aspect of the present invention, the initialization step further includes: normalizing the sample feature dataset and obtaining a training set from the linearly inseparable features using the following nonlinear SVM algorithm:
[0033] T = {(x1,y1),(x2,y2),…,(x n ,y n )}, where x i ∈X=Rn,y i ∈Y={-1,+1},
[0034] Where, x i Let x be the i-th instance, and if n > 1, meaning x is multi-dimensional and has multiple attribute features, then x i Let y be a vector. i For x i The class marker, when y i When x is +1, i For a positive example, when y i When x is -1, i For negative examples, T is the training set;
[0035] The modeling steps also include: using the SVM model to solve for the Lagrange multipliers and intercepts of the optimal classification hyperplane in the feature space of the training set, thereby obtaining the decision function.
[0036] As a further aspect of the present invention: the actual time-domain signal in the denoising step is subjected to a short-time Fourier transform according to the following formula:
[0037] STFT f(w,τ) =∫ R f(t)g' w,τ (t)dt=∫ R f(t)g(t-τ)e -jwt dt = <f(t)·g w,τ (t)>,
[0038] g w,τ (t)=g(t-τ)e jwt ,
[0039]
[0040] In the formula, f(t) is a non-stationary signal, ω is the frequency, and e is the frequency. -jwtIt is a complex exponential function, g(t-τ) is the analysis window function, g(t) is the window function. The short-time Fourier transform yields the Fourier transform of the function f(t) at time τ. By continuously changing τ, we can obtain the Fourier transform of the function f(t) at different times.
[0041] As a further aspect of the present invention: the effective time-domain signal in the frequency decomposition step is subjected to wavelet transform according to the following formula:
[0042]
[0043] In the formula, f(t) is the signal to be analyzed, t is the independent variable, the time-domain signal, and the input of the wavelet transform, and WT(α,τ) is the result of the wavelet transform. Let α be the wavelet basis function, α be the scaling factor, and τ be the translation.
[0044] As a further aspect of the present invention: the initialization step is as follows: the vibration signals of the collected containers in empty and non-empty states are filtered using a bandpass filter, and then a short-time Fourier transform is performed to obtain a spectrum. Frequency values with non-attenuated amplitudes in the spectrum are removed to obtain a sample library. The differential features in the sample library are extracted to obtain a sample feature dataset.
[0045] The vibration signal in the initialization step and the actual time-domain signal in the acquisition step are both generated by using a pendulum to strike the container using the hammering method.
[0046] The actual time-domain signal in the acquisition step is acquired using a laser rangefinder sensor located on the same side as the pendulum that strikes the container.
[0047] The noise reduction steps are as follows: the actual time-domain signal acquired is filtered using a bandpass filter, and then a short-time Fourier transform is performed to obtain a spectrum. Frequency values with non-attenuated amplitudes are removed from the spectrum to obtain the effective time-domain signal.
[0048] The initialization step is preceded by an excitation step: the container is struck by a hammer to generate a vibration signal and / or a real time-domain signal with composite attenuation.
[0049] The present invention also provides another technical solution: an intelligent detection device for empty containers, comprising:
[0050] The excitation module is used to strike the container using a hammering method.
[0051] The initialization module is used to perform short-time Fourier transform on the vibration signals generated by hammering containers in empty and non-empty states to obtain a spectrum. Frequency values with non-attenuated amplitudes are removed from the spectrum to obtain effective vibration data. Wavelet transform is performed on the effective vibration data to obtain a sample library of frequency decomposition amplitude curves. The differential features in the sample library are extracted to obtain a sample feature dataset.
[0052] The SVM modeling module is used to build a Support Vector Machine (SVM) model with a kernel function, and then uses this SVM model to solve for the Lagrange multipliers and intercept of the optimal classification hyperplane in the feature space of the sample feature dataset, thus obtaining the decision function:
[0053]
[0054] Here, sign, also called sgn, is the sign function, used to extract the sign of a number, K(x,x) i ) is the kernel function. For Lagrange multipliers, b * x is the intercept. i Let be the vector of the feature dataset of the i-th sample;
[0055] The signal acquisition module is used to acquire the actual time-domain signal of composite attenuation generated by the hammering method on the container to be tested;
[0056] The noise reduction module is used to perform a short-time Fourier transform on the acquired actual time-domain signal to obtain a spectrum, and remove frequency values with non-attenuating amplitude from the spectrum to obtain the effective time-domain signal.
[0057] A frequency decomposition module is used to perform wavelet transform on the effective time-domain signal to obtain amplitude curves of several decomposed frequencies; and
[0058] The judgment module is used to extract the difference features of the amplitude curve and input them into the decision function, output the result of the decision function, and judge whether the container to be detected is empty or not based on the result.
[0059] The beneficial effects of this invention are:
[0060] Compared with existing technologies, this solution does not require pre-determining working parameters in the entire detection method. Instead, it establishes an SVM model to achieve an adaptive detection algorithm, which is more intelligent and effectively improves detection efficiency and detection results.
[0061] By using a noise reduction step, the vibration signal or noise caused by the vehicle's vibration is denoised and reduced at a specific frequency, thus eliminating the need for special requirements on the starting or stopping state of the container vehicle and improving the speed of container customs clearance.
[0062] It saves labor and time costs, the detection process is harmless to the human body, and it is a safe, radiation-free, non-invasive detection method. It does not require additional equipment to determine the size and location of containers and is highly compatible with containers of different sizes. Attached Figure Description
[0063] Figure 1 This is a flowchart of the intelligent detection method for empty containers according to the present invention.
[0064] Figure 2 This is a flowchart illustrating the modeling steps in the intelligent empty container detection method described in this invention.
[0065] Figure 3 This is a schematic diagram of the structure for performing the excitation step and the acquisition step of the present invention.
[0066] Figure 4 This refers to the original time-domain signal of the container acquired in this invention, namely the vibration signal in the initialization step and / or the actual time-domain signal in the acquisition step.
[0067] Figure 5 for Figure 4 The time-domain signal after removing jitter noise from the acquisition device is obtained by using a bandpass filter.
[0068] Figure 6 for Figure 5 The time-domain signal after removing the vehicle's vibration signal through short-time Fourier transform. Detailed Implementation
[0069] The present invention will now be described in detail with reference to the accompanying drawings.
[0070] like Figures 1-6 As shown in the figure, this embodiment of the invention provides an intelligent detection method for empty containers, including an initialization step, a modeling step, a data acquisition step, a noise reduction step, a frequency decomposition step, and a judgment step. Wherein:
[0071] Initialization steps: such as Figure 4 In this process, a short-time Fourier transform is performed on the vibration signals of the collected containers in both empty and non-empty states to obtain a spectrum. Frequency values with non-attenuating amplitudes are then removed from this spectrum to obtain the following result: Figure 6 The effective vibration data is used to perform wavelet transform on the effective vibration data to obtain a sample library of frequency decomposition amplitude curves. The differential features in the sample library are extracted to obtain a sample feature dataset.
[0072] Modeling steps: Establish a support vector machine (SVM) model with a kernel function, and use the SVM model to solve for the Lagrange multipliers and intercepts of the optimal classification hyperplane in the feature space of the sample feature dataset to obtain the decision function;
[0073] Data Acquisition Steps: Acquire the actual time-domain signal of the composite attenuation generated by the vibration of the container to be detected, such as... Figure 4 As shown;
[0074] Denoising step: Perform a short-time Fourier transform on the acquired actual time-domain signal to obtain a spectrum. Remove frequency values with non-attenuating amplitudes from the spectrum to obtain the effective time-domain signal, such as... Figure 6 As shown;
[0075] Frequency decomposition step: Perform wavelet transform on the effective time-domain signal to obtain amplitude curves of several decomposed frequencies;
[0076] Judgment Steps: Extract the differential features of the amplitude curve and input them into the decision function. Output the result of the decision function and determine whether the container to be detected is empty or not based on the result. Specifically, the differential features in the initialization and judgment steps are determined by analyzing the differences between empty and non-empty containers across multiple dimensions such as curve energy, volatility, attenuation, locality, and globality.
[0077] Compared with existing technologies, this solution does not require pre-determining working parameters in the entire detection method. Instead, it establishes an SVM model to achieve an adaptive detection algorithm, which is more intelligent and effectively improves detection efficiency and detection results.
[0078] By using a noise reduction step, the vibration signal or noise caused by the vehicle's vibration is denoised and reduced at a specific frequency, thus eliminating the need for special requirements on the starting or stopping state of the container vehicle and improving the speed of container customs clearance.
[0079] It saves labor and time costs, the detection process is harmless to the human body, and it is a safe, radiation-free, non-invasive detection method. It does not require additional equipment to determine the size and location of containers and is highly compatible with containers of different sizes.
[0080] In one embodiment, the kernel function and the decision function in the modeling step are respectively represented as follows:
[0081] Kernel function:
[0082] Where k(x,z) is the Gaussian kernel function, and x and z are vectors of two sample feature datasets, ||xz|| 2 Let x be the square of the distance between vectors x and z, and θ be the variance of the Gaussian function.
[0083] Decision function:
[0084] Where K(x,x) i ) is the kernel function. For Lagrange multipliers, b * x is the intercept. i Let K(x,z) be the vector of the feature dataset for the i-th sample. Specifically, in the decision function, sign, also called sgn, is the sign function used to extract the sign of a number. Here, it is used to extract 1 or -1, corresponding to positive and negative examples, i.e., empty and non-empty bins, respectively. When K(x,z) is a positive definite kernel function, it is a convex quadratic programming problem, and a solution exists. Furthermore, an adaptive detection algorithm is implemented, which is more intelligent and effectively improves detection efficiency and results.
[0085] In another embodiment, such as Figure 2 As shown, the sample feature dataset is... Figure 2 The prepared dataset includes linearly separable features and linearly inseparable features. The SVM model includes a linear SVM model corresponding to the linearly separable features. The Lagrange multipliers of the linear SVM model are solved according to the following formula:
[0086]
[0087] st∑ i α i y i =0,α i ≥0,
[0088] This formula is used to solve a quadratic programming problem. Based on the duality of the Lagrange function and through certain simplification methods, the original problem is transformed. In the formula, α... i α is the Lagrange multiplier for the linear SVM model. j x is the Lagrange multiplier for the linear SVM model. i For the i-th instance, x j For the j-th instance, y i For x i The class marker, y j For x j The class markers are then used to implement adaptive detection algorithms, which are more intelligent and effectively improve detection efficiency and results.
[0089] In yet another embodiment, such as Figure 2 As shown, the sample feature dataset is... Figure 2 The prepared dataset includes linearly separable features and linearly inseparable features. The SVM model includes a nonlinear SVM model corresponding to the linearly inseparable features. For the nonlinear SVM model, an appropriate kernel function k(x,z) and parameter C are selected to construct and solve the optimization problem. The Lagrange multipliers of the nonlinear SVM model are solved according to the following formula:
[0090]
[0091] st∑ i α i y i =0, 0≤α i ≤C,
[0092] Solving for:
[0093] In the formula, K(x) i ,x j Let k(x,z) be the Gaussian kernel function. i ,x j Gaussian kernel function can be selected. C is the selected parameter, α i α is the Lagrange multiplier for the nonlinear SVM model. j x is the Lagrange multiplier for the nonlinear SVM model. i For the i-th instance, x j For the j-th instance, y i For x i The class marker, y j For x j The class label; when the kernel function k(x,z) is a positive definite kernel function, the solution of this formula belongs to a convex quadratic programming problem, and its optimal solution is the Lagrange multiplier vector. Let represent the optimal solution for the i-th Lagrange multiplier. Furthermore, this enables an adaptive detection algorithm, which is more intelligent and effectively improves detection efficiency and results.
[0094] In yet another embodiment, such as Figure 2 As shown, the intercept is determined according to the following steps:
[0095] From the optimal solution of the Lagrange multiplier vector Select a positive component
[0096] The y-th sample feature dataset corresponding to index i i Substitute into the formula:
[0097]
[0098] Solve for b * In the formula, b * The intercept is... For the optimal solution of the i-th Lagrange multiplier, y i This represents the data in the feature dataset for the i-th sample. Furthermore, an adaptive detection algorithm is implemented, which is more intelligent and effectively improves detection efficiency and results.
[0099] In another embodiment, such as Figure 2As shown, the initialization step also includes: normalizing the sample feature dataset to obtain a training set;
[0100] The modeling steps also include: using the SVM model to solve for the Lagrange multipliers and intercept of the optimal classification hyperplane in the feature space of the training set, thereby obtaining the decision function. This leads to an adaptive detection algorithm with higher intelligence, effectively improving detection efficiency and performance.
[0101] In another embodiment, such as Figure 2 As shown, the initialization step further includes: normalizing the sample feature dataset and obtaining a training set from the linearly inseparable features using the following nonlinear SVM algorithm:
[0102] T = {(x1,y1),(x2,y2),…,(x n ,y n )}, where x i ∈X=Rn,y i ∈Y={-1,+1},
[0103] Where, x i Let x be the i-th instance, and if n > 1, meaning x is multi-dimensional and has multiple attribute features, then x i Let y be a vector. i For x i The class marker, when y i When x is +1, i For a positive example, when y i When x is -1, i For negative examples, T is the training set;
[0104] The modeling steps also include: using the SVM model to solve for the Lagrange multipliers and intercepts of the optimal classification hyperplane in the feature space of the training set, thus obtaining the decision function. Specifically, the differences in decomposition curve features and temporal features between empty and non-empty boxes are analyzed, and then it is determined whether a box is empty. The extracted multidimensional features are mapped to a higher-dimensional space, and different classes of samples are partitioned in the sample space to obtain the optimal hyperplane, i.e., the SVM classifier, achieving robust classification with fewer samples. Furthermore, SVM maps the original data to a high-dimensional space, allowing the data to be better separated in the high-dimensional space, thereby achieving the purpose of classification or regression. In SVM, "support vectors" refer to those vectors on the decision boundary that have the largest distance to their neighboring data points, and thus can be used to determine the location of the decision boundary. For linearly inseparable features, we need to use a nonlinear SVM; the nonlinear SVM algorithm flow is as described above for the given training set T. Furthermore, an adaptive detection algorithm is implemented, which is more intelligent and effectively improves detection efficiency and detection results.
[0105] In another embodiment, such as Figure 1 As shown, the actual time-domain signal in the denoising step undergoes a short-time Fourier transform according to the following formula:
[0106] STFT f(w,τ) =∫ R f(t)g' w,τ (t)dt=∫ R f(t)g(t-τ)e -jwt dt = <f(t)·g w,τ (t)>,
[0107] g w,τ (t)=g(tv)e jwt ,
[0108]
[0109] In the formula, f(t) is a non-stationary signal, ω is the frequency, and e is the frequency. -jwt It is a complex exponential function, g(t-τ) is the analysis window function, and g(t) is the window function (which should be a symmetric function). The short-time Fourier transform yields the Fourier transform of function f(t) at time τ. By continuously changing τ, i.e., continuously moving the center position of the window function g(t), we can obtain the Fourier transform of function f(t) at different times. Specifically, for example... Figure 4 The time-domain data acquired is processed using a specific algorithm, and a bandpass filter is used for filtering to remove jitter noise inherent in the acquisition equipment during operation, resulting in the following: Figure 5 The time-domain signal of the acquisition device jitter is removed, and then vehicle jitter interference is removed by short-time Fourier transform, resulting in the following: Figure 6 The process removes the time-domain signal after vehicle vibration, thereby effectively improving recognition stability. First, a Short-Time Fourier Transform (STFT) is performed on the time-domain signal to obtain a spectrum. Then, frequencies with non-attenuating amplitudes are identified and eliminated. STFT is essentially a windowed Fourier transform. The STFT process involves multiplying the signal by a time-finite window function g(t) before the Fourier transform, assuming that the non-stationary signal is stationary within a short time interval of the analysis window. By moving the window function g(t) along the time axis, a set of local "spectrums" of the signal is obtained through segment-by-segment analysis. The transform formula is: taking the time function g(t) as the window function, multiplying g(t-τ) by the function to be analyzed f(t), and then performing a Fourier transform. Sliding the window g(t) along the time axis yields the frequency distribution across the entire time axis. Furthermore, specific frequencies are used to denoise and reduce vibration signals or noise caused by vehicle vibration. Consequently, there are no special requirements for the starting or stopping status of container vehicles, thus improving the speed of container customs clearance.
[0110] In another embodiment, such as Figure 1 As shown, the effective time-domain signal in the frequency decomposition step undergoes wavelet transform according to the following formula:
[0111]
[0112] In the formula, f(t) is the signal to be analyzed, t is the independent variable, the time-domain signal, and the input of the wavelet transform, and WT(α,τ) is the result of the wavelet transform. Let α be the wavelet basis function, τ be the scale factor, and τ be the translation. Specifically, frequency decomposition is performed on the preprocessed effective time-domain data (i.e., the effective vibration data or actual time-domain signal after short-time Fourier transform); wavelet transform is then performed on the time-domain signal after removing the vehicle vibration frequency to obtain the amplitude curve of each decomposed frequency. Wavelet transform is a local transformation of space (time) and frequency. Through scaling and translation operations, it can perform multi-scale refinement analysis of functions or signals, transforming the wavelet basis into a time-frequency spectrum through scaling and translation. As can be seen from the above equation, wavelet transform has two variables: scale α and translation τ. Scale α controls the scaling of the wavelet function, and translation τ controls the translation of the wavelet function. The scale corresponds to frequency (inversely proportional), and the translation τ corresponds to time. Furthermore, differential features are extracted from the frequency decomposition curves. Specifically, features with explicit identifiability are obtained by analyzing the differences between empty and non-empty boxes in multiple dimensions such as curve energy, volatility, attenuation, locality, and globality.
[0113] In another embodiment, the initialization step is: for example... Figure 4 The vibration signals of the container collected in both empty and non-empty states are filtered using a bandpass filter to obtain the following results: Figure 5 The time-domain signal is removed to eliminate jitter from the acquisition device. Furthermore, jitter noise inherent in the acquisition device during operation is filtered out. A short-time Fourier transform is then performed to obtain a spectrum, and frequencies with non-attenuating amplitudes are removed from this spectrum to obtain the following result: Figure 6 The time-domain signal after removing vehicle vibration is used to obtain a sample library. Then, the vibration signal or noise caused by vehicle vibration is denoised and de-vibrated at specific frequencies. This eliminates the need for special requirements regarding the starting or stopping status of container vehicles, improving container customs clearance speed. Finally, the differential features in this sample library are extracted to obtain a sample feature dataset.
[0114] The vibration signal in the initialization step and the actual time-domain signal in the acquisition step are both generated by striking the container with a pendulum using a hammer-impact method. This hammer-impact method enables the intelligent empty container detection method to generate signals based on vibration, thereby resulting in a higher signal-to-noise ratio, more stable signals, and better robustness of the detection results. The pendulum uses a 250g iron hammer with a rubber surface. The pendulum swings to a certain height and then falls, thus maintaining a fixed hammering force. The striking location can be the convex, concave, or inclined surface of the container.
[0115] The actual time-domain signal acquired in the data acquisition step is collected using a laser rangefinder sensor positioned on the same side as the pendulum striking the container. This same-side acquisition and excitation facilitates installation and wiring design, enabling non-contact acquisition without scanning excitation; only a single strike of the container by the pendulum is required. Laser vibrometers offer high accuracy; they are currently the best measurement method for obtaining displacement and velocity resolution, achieving picometer-level amplitude resolution, high linearity, and ensuring amplitude consistency even at extremely high frequencies. Furthermore, the non-contact nature of the sensor enhances safety.
[0116] The denoising step is as follows: The acquired actual time-domain signal is filtered using a bandpass filter. This filters out jitter noise inherent in the acquisition equipment during operation. A short-time Fourier transform is then performed to obtain a spectrum, and frequencies with non-attenuating amplitudes are removed to obtain the effective time-domain signal. Furthermore, specific frequencies are used to denoise and reduce jitter caused by vehicle vibration, thus eliminating special requirements for container vehicle start-up or shutdown states and improving container customs clearance speed.
[0117] The initialization step is preceded by an excitation step: the container is struck by a hammer to generate a vibration signal and / or a real time-domain signal with composite attenuation.
[0118] A smart empty container detection device includes an excitation module, an initialization module, an SVM modeling module, a signal acquisition module, a noise reduction module, a frequency decomposition module, and a judgment module. Wherein:
[0119] The excitation module is used to strike the container using a hammering method.
[0120] The initialization module is used to perform short-time Fourier transform on the vibration signals generated by hammering containers in empty and non-empty states to obtain a spectrum. Frequency values with non-attenuated amplitudes are removed from the spectrum to obtain effective vibration data. Wavelet transform is performed on the effective vibration data to obtain a sample library of frequency decomposition amplitude curves. The differential features in the sample library are extracted to obtain a sample feature dataset.
[0121] The SVM modeling module is used to build a Support Vector Machine (SVM) model with a kernel function, and then uses this SVM model to solve for the Lagrange multipliers and intercept of the optimal classification hyperplane in the feature space of the sample feature dataset, thus obtaining the decision function:
[0122]
[0123] Here, sign, also called sgn, is the sign function, used to extract the sign of a number, K(x,x) i ) is the kernel function. For Lagrange multipliers, b * x is the intercept. i Let be the vector of the feature dataset of the i-th sample;
[0124] The signal acquisition module is used to acquire the actual time-domain signal of composite attenuation generated by the hammering method on the container to be tested;
[0125] The noise reduction module is used to perform a short-time Fourier transform on the acquired actual time-domain signal to obtain a spectrum, and remove frequency values with non-attenuating amplitude from the spectrum to obtain the effective time-domain signal.
[0126] The frequency decomposition module is used to perform wavelet transform on the effective time-domain signal to obtain amplitude curves of several decomposed frequencies.
[0127] The judgment module is used to extract the difference features of the amplitude curve and input them into the decision function, output the result of the decision function, and judge whether the container to be detected is empty or not based on the result.
[0128] Working principle: For a vibration signal collected from a container, the signal includes the signal generated by the hammer impacting the container itself, the spatial reflection signal inside the container, the vehicle vibration signal when the vehicle is running, and the vibration noise generated by the acquisition equipment during operation. For such a complex composite vibration signal, after data preprocessing to remove interference, the composite wave energy in the response of empty and non-empty containers is decomposed in the frequency domain. By finding reference parameters, for the same container, different impact forces and different slots can be normalized to a certain range. By analyzing the differences between empty and non-empty containers in various dimensions, features with explicit recognition are obtained. The extracted multidimensional features are mapped to a higher-dimensional space, and different types of samples (i.e., the sample feature dataset obtained after extracting differential features) are partitioned in the sample space to obtain the optimal classification hyperplane.
[0129] In summary, the present invention possesses the excellent characteristics described above, which enhances its effectiveness in use compared to previous technologies, making it a highly practical product.
[0130] The above description is only a preferred embodiment of the present invention. For those skilled in the art, there will be changes in the specific implementation and application scope based on the ideas of the present invention. The content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for intelligent detection of empty containers, characterized in that, include: Initialization steps: Obtain the spectrum of vibration signals of the collected container in empty and non-empty states, remove the frequency values in the spectrum that do not decay in amplitude to obtain effective vibration data, perform wavelet transform on the effective vibration data to obtain a sample library of frequency decomposition amplitude curves, extract the differential features in the sample library and obtain the sample feature dataset. Modeling steps: Establish a support vector machine (SVM) model with a kernel function, and use the SVM model to solve for the Lagrange multipliers and intercepts of the optimal classification hyperplane in the feature space of the sample feature dataset to obtain the decision function; Acquisition steps: Acquire the actual time-domain signal of the composite attenuation generated by the vibration of the container to be detected; Denoising step: Perform a short-time Fourier transform on the acquired actual time-domain signal to obtain a spectrum, remove the frequency values in the spectrum that do not decay in amplitude, and obtain the effective time-domain signal; Frequency decomposition step: Perform wavelet transform on the effective time-domain signal to obtain amplitude curves of several decomposed frequencies; Judgment steps: Extract the difference features of the amplitude curve and input them into the decision function, output the result of the decision function, and determine whether the container to be detected is empty or not based on the result; Specifically, the vibration signal in the initialization step and the actual time-domain signal in the acquisition step are both generated by striking the container with a pendulum using the hammer impact method.
2. The intelligent detection method for empty containers according to claim 1, characterized in that, In the modeling step, the kernel function and the decision function are represented as follows: Kernel function: , Where k(x,z) is the Gaussian kernel function, and x and z are vectors of two sample feature datasets, respectively. Let x be the square of the distance between vectors x and z, and θ be the variance of the Gaussian function. Decision function: , Here, `sign`, also called `sgn`, is the sign function, used to extract the sign of a number. For kernel function, For Lagrange multipliers, The intercept is... Let i be the vector of the feature dataset of the i-th sample. for The class tag.
3. The intelligent detection method for empty containers according to claim 2, characterized in that, The sample feature dataset includes linearly separable features, and the SVM model includes a linear SVM model corresponding to the linearly separable features. The Lagrange multipliers of the linear SVM model are solved using the following formula: ， ， In the formula, For the Lagrange multipliers of the linear SVM model, For the Lagrange multipliers of the linear SVM model, Let i be the vector of the feature dataset of the i-th sample. Let j be the vector of the feature dataset of the j-th sample. for Class marker, for The class tag.
4. The intelligent detection method for empty containers according to claim 2, characterized in that, The sample feature dataset includes linearly inseparable features, and the SVM model includes a nonlinear SVM model corresponding to the linearly inseparable features. The Lagrange multipliers of the nonlinear SVM model are solved according to the following formula: ， ， In the formula, C is the selected parameter. For the Lagrange multipliers of the nonlinear SVM model, For the Lagrange multipliers of the nonlinear SVM model, Let i be the vector of the feature dataset of the i-th sample. Let j be the vector of the feature dataset of the j-th sample. for Class marker, for The class label; when the kernel function k(x,z) is a positive definite kernel function, the solution of this formula belongs to a convex quadratic programming problem, and its optimal solution is a Lagrange multiplication. subvector , Let represent the optimal solution for the i-th Lagrange multiplier.
5. The intelligent detection method for empty containers according to claim 4, characterized in that, The intercept is determined according to the following steps: From the optimal solution of the Lagrange multiplier vector Select a positive component ; The feature dataset of the i-th sample corresponding to index i. Substitute into the formula: ， Solve , in the formula, The intercept is... For the optimal solution of the i-th Lagrange multiplier, For the data of the i-th sample feature dataset, for The class tag.
6. The intelligent detection method for empty containers according to claim 1, characterized in that, The initialization step also includes: normalizing the sample feature dataset to obtain a training set; The modeling steps also include: using the SVM model to solve for the Lagrange multipliers and intercepts of the optimal classification hyperplane in the feature space of the training set, thereby obtaining the decision function.
7. The intelligent detection method for empty containers according to claim 4, characterized in that, The initialization step further includes: normalizing the sample feature dataset and obtaining a training set from the linearly inseparable features using the following nonlinear SVM algorithm: T={( , ),( , ),…,( , )},in ∈X= , ∈Y={1,+1}, in, For the i-th vector, i.e. For vectors, for The class tag, when When it is +1, As a positive example, when When it is -1, For negative examples, T is the training set; The modeling steps also include: using the SVM model to solve for the Lagrange multipliers and intercepts of the optimal classification hyperplane in the feature space of the training set, thereby obtaining the decision function.
8. The intelligent detection method for empty containers according to claim 1, characterized in that, The actual time-domain signal in the denoising step is subjected to a short-time Fourier transform according to the following formula: ， = ， = ， In the formula, f(t) is a non-stationary signal, and ω is the frequency. It is a complex exponential function, g(t—τ) is the analysis window function, g(t) is the window function, and the short-time Fourier transform yields the Fourier transform of the function f(t) at time τ. By continuously changing τ, we can obtain the Fourier transform of the function f(t) at different times.
9. The intelligent detection method for empty containers according to claim 1, characterized in that, The effective time-domain signal in the frequency decomposition step is subjected to wavelet transform according to the following formula: ， In the formula, f(t) is the signal to be analyzed, t is the independent variable, the time-domain signal and the input of the wavelet transform, WT(α,τ) is the result of the wavelet transform, φ is the wavelet basis function, α is the scale factor and τ is the translation amount.
10. The intelligent detection method for empty containers according to claim 1, characterized in that, The initialization steps are as follows: the vibration signals of the collected containers in empty and non-empty states are filtered using a bandpass filter, and then a short-time Fourier transform is performed to obtain a spectrum. Frequency values with non-attenuated amplitudes in the spectrum are removed to obtain a sample library. The differential features in the sample library are extracted to obtain a sample feature dataset. The actual time-domain signal in the acquisition step is acquired using a laser rangefinder sensor located on the same side as the pendulum that strikes the container. The noise reduction steps are as follows: the actual time-domain signal acquired is filtered using a bandpass filter, and then a short-time Fourier transform is performed to obtain a spectrum. Frequency values with non-attenuated amplitudes are removed from the spectrum to obtain the effective time-domain signal. The initialization step is preceded by an excitation step: the container is struck by a hammer to generate a vibration signal and / or a real time-domain signal with composite attenuation.
11. A smart detection device for empty containers, characterized in that, include: The excitation module is used to strike the container using a hammering method. The initialization module is used to perform short-time Fourier transform on the vibration signals generated by hammering containers in empty and non-empty states to obtain a spectrum. Frequency values with non-attenuated amplitudes are removed from the spectrum to obtain effective vibration data. Wavelet transform is performed on the effective vibration data to obtain a sample library of frequency decomposition amplitude curves. The differential features in the sample library are extracted to obtain a sample feature dataset. The SVM modeling module is used to build a Support Vector Machine (SVM) model with a kernel function, and then uses this SVM model to solve for the Lagrange multipliers and intercept of the optimal classification hyperplane in the feature space of the sample feature dataset, thus obtaining the decision function: ， Here, `sign`, also called `sgn`, is the sign function, used to extract the sign of a number. For kernel function, For Lagrange multipliers, The intercept is... Let i be the vector of the feature dataset of the i-th sample. for Class marker; The signal acquisition module is used to acquire the actual time-domain signal of composite attenuation generated by the hammering method on the container to be tested; The noise reduction module is used to perform a short-time Fourier transform on the acquired actual time-domain signal to obtain a spectrum, and remove frequency values with non-attenuating amplitude from the spectrum to obtain the effective time-domain signal. A frequency decomposition module is used to perform wavelet transform on the effective time-domain signal to obtain amplitude curves of several decomposed frequencies; and The judgment module is used to extract the difference features of the amplitude curve and input them into the decision function, output the result of the decision function, and judge whether the container to be detected is empty or not based on the result.