Egg quality detection system and method based on multi-sensor fusion

Abstract

The invention discloses an egg quality detection system and method based on multi-sensor fusion, and relates to the technical field of multi-dimensional data process.The method comprises the steps that texture features are extracted from visible light images, and feature wavelengths are extracted from spectrums; pressure distribution and shell deformation in the pressing process are synchronously collected; recording a pressure-time curve and a deformation-time curve, extracting characteristic parameters, and constructing a three-dimensional dynamic model; performing wavelet transformation on the response curve of each sensor, extracting time-frequency domain characteristics, and generating a smell fingerprint spectrum; scanning the surface of the egg to excite resonance response, recording a sound pressure signal, and converting the sound pressure signal into frequency domain data through fast Fourier transform; and presetting a business rule, and generating a final decision by combining the model output with a rule matching result. The pressure feature extraction module, the smell feature extraction module and the crack degree quantification module are arranged to realize data depth association, and the quality inspection credibility is improved through an interpretable decision model.

Description

Technical Field

[0001] The present invention relates to the technical field of multidimensional data processing, and in particular to an egg quality detection system and method based on multi-sensor fusion. Background Art

[0002] As one of the most consumed sources of animal protein worldwide, egg quality is directly linked to food safety, consumer health, and the economic benefits of the industry. Traditional egg quality testing relies primarily on visual inspection for cracks and stains, as well as simple physical tests like weight and size. These methods suffer from low efficiency, high subjectivity, and limited testing parameters.

[0003] With the large-scale development of the food industry, the demand for automated, multi-dimensional, and non-destructive inspection technologies is becoming increasingly urgent. Currently, single-sensor inspection methods, such as spectroscopy or pressure, are unable to fully cover key indicators of eggs, such as external defects, cracks, stains, internal quality, freshness, albumen height, and chemical composition, including fat and protein content. Summary of the Invention

[0004] In order to solve the above technical problems, an egg quality detection system and method based on multi-sensor fusion are provided. This technical solution solves the problems raised in the above background technology.

[0005] In order to achieve the above objects, the technical solution adopted by the present invention is: The egg quality detection system based on multi-sensor fusion includes: A spectral data acquisition module, which is used to integrate a visible light camera and a near-infrared spectral sensor to synchronously collect egg surface images and internal component spectral data; A spectral feature extraction module is used to correct baseline drift by combining standard normal variable transformation, extract texture features from visible light images, extract characteristic wavelengths from spectra, and construct a "spectrum-image" joint feature vector after dimensionality reduction through principal component analysis; A pressure data acquisition module, which is used to customize a pressure sensor array and, in combination with a laser displacement sensor, synchronously acquires the pressure distribution and shell deformation during the pressing process; A pressure feature extraction module is used to record pressure-time curves and deformation-time curves, extract characteristic parameters, construct a three-dimensional dynamic model, and analyze time series characteristics through a long-short-term memory network. The characteristic parameters include maximum pressure, deformation recovery rate, and pressure fluctuation frequency; An odor feature extraction module, which uses metal oxide semiconductor, conductive polymer, and quantum tunneling composite sensors to cover the detection range of egg spoilage markers, performs wavelet transform on the response curve of each sensor, extracts time-frequency domain features, and generates an odor fingerprint; A crack severity quantification module is used to design an electromagnetic vibrator to scan the egg surface to stimulate a resonant response, record the sound pressure signal, convert it into frequency domain data through a fast Fourier transform, and quantify the crack severity by combining it with a crack depth model; The quality classification generation module is used to calculate the marginal contribution value of the input features to the model output for each test result, quantify its contribution to the classification decision, preset business rules, and combine the model output with the rule matching results to generate the final decision.

[0006] Preferably, the recording of pressure-time curves and deformation-time curves, extracting characteristic parameters, constructing a three-dimensional dynamic model, and analyzing time series characteristics through a long short-term memory network specifically include: Find the pressure peak from the pressure curve, record its value and occurrence time, observe the periodic fluctuations in the pressure curve, count the number of fluctuations per unit time, and calculate the time required for the pressure to rise from the initial value to the maximum value; Find the maximum deformation value from the deformation curve, record its value and the corresponding pressure value, calculate the percentage of the recovery amount to the maximum deformation value after unloading the pressure and output it as the recovery ratio, and record the time required for the deformation to recover to 10% after the pressure is unloaded to zero; Comparing the changing trends of the pressure and deformation curves and analyzing the phase relationship between the two, the phase relationship includes a synchronous relationship and a lagging relationship; Calculate the enclosed area of the pressure-deformation curve during loading and unloading to evaluate the energy loss caused by friction within the material; With time as the horizontal axis, pressure as the vertical axis, and deformation as the vertical axis, a three-dimensional coordinate system is constructed to map the synchronously collected pressure and deformation data into three-dimensional space to form a discrete data point set; The cubic spline interpolation method is used for discrete data points to generate a continuous pressure-deformation-time surface; Compare the model prediction results with the actual measurement data and calculate the deviation between the predicted and actual deformation variables; Adjust model parameters in the cubic spline interpolation method based on error analysis results; The pressure and deformation data are divided into segments of fixed length in chronological order, and the data are normalized and scaled to a value between 0 and 1. Construct a bidirectional long short-term memory network, which includes an input layer, a hidden layer, and an output layer; The input layer receives dual-channel data of pressure and deformation, the hidden layer captures temporal dependencies through memory units, and the output layer predicts the pressure and deformation values at the next moment; The dataset is divided into a training set and a test set. The training set is used to train the memory unit weights, and the test set is used to evaluate the model performance. Analyze the network's dependence on historical data, identify the key time points that affect the prediction results, and extract the dynamic relationship between pressure fluctuation frequency and deformation recovery rate through model output results.

[0007] Preferably, the metal oxide semiconductor, conductive polymer and quantum tunneling composite sensor is used to cover the detection range of egg spoilage markers, and the response curve of each sensor is subjected to wavelet transform to extract time-frequency domain features to generate the odor fingerprint, which specifically includes: Fresh eggs and eggs at different stages of corruption were selected and placed in a sealed container to allow volatile gases to be fully released. The corruption stages included mild, moderate, and severe. A metal oxide semiconductor, conductive polymer, and quantum tunneling composite sensor array is placed in a container, and gas is extracted to synchronously record changes in resistance and conductivity. Perform three-layer wavelet decomposition on the response curve of each sensor to generate low-frequency approximation coefficients and high-frequency detail coefficients; The energy and mean are extracted as low-frequency features reflecting the steady-state process of gas adsorption; Standard deviation and entropy are extracted as high-frequency features that capture transient fluctuations of gas molecule collisions; Arrange the time-frequency domain features of each sensor in order to form a feature vector, and merge the feature vectors of the three sensors to construct a 9-dimensional feature matrix; A heat map is generated with sensor type as the horizontal axis and time-frequency characteristics as the vertical axis, and the color depth represents the size of the characteristic value; The polar coordinate distribution of each feature was plotted based on the spoilage stage, and the differences between fresh and spoiled eggs were compared.

[0008] Preferably, the electromagnetic vibrator is designed to scan the egg surface to excite the resonance response, the sound pressure signal is recorded and converted into frequency domain data through fast Fourier transform, and the crack severity is quantified in combination with the crack depth model. Specifically, the following steps are performed: Neodymium iron boron permanent magnets are used as the main magnets to form a closed magnetic circuit with the silicon steel core. A copper coil is wound in the air gap of the core. The number of coil turns is determined based on the target exciting force and frequency range. The coil is driven by a function generator, a sine frequency sweep signal is input, and the displacement of the excitation head is monitored by a laser displacement sensor; Select an egg with at least one crack depth and spray a thin layer of graphite powder on the egg surface as an acoustic coupling medium to enhance the contact stiffness between the excitation head and the eggshell. The egg surface is divided into a 10×10 grid, and a three-axis mobile platform is used to perform vibration excitation and signal acquisition. At each grid point, the exciter outputs a sine sweep signal and simultaneously records the sound pressure signal on the eggshell surface. The displacement of the excitation head and the vibration velocity of the egg surface were recorded synchronously as auxiliary verification data; Perform fast Fourier transform analysis on each signal segment to generate a spectrum diagram with the horizontal axis being frequency and the vertical axis being sound pressure amplitude; Extract the amplitude peak value corresponding to each frequency point, construct a "frequency-amplitude" curve, scan the "frequency-amplitude" curve using a sliding window algorithm, and identify the frequency corresponding to the local maximum amplitude as the resonant frequency; Calculate the difference between the resonant frequency of the cracked egg and the resonant frequency of the fresh egg to reflect the reduction in stiffness caused by the crack; The damping ratio of the cracked area is calculated by the half-power bandwidth method, and the ratio of the sound pressure amplitude at the resonant frequency of the cracked area to the amplitude of the fresh area is calculated. Generate a crack distribution heat map, map the crack depth of each grid point to the egg surface, and use color depth to indicate the severity, with red indicating severe cracks and green indicating no cracks.

[0009] Furthermore, a method for detecting egg quality based on multi-sensor fusion is proposed, which is used to implement the above-mentioned egg quality detection system based on multi-sensor fusion, including: Integrate a visible light camera and a near-infrared spectral sensor to simultaneously collect egg surface images and internal component spectral data; The baseline drift is corrected by combining standard normal variable transformation, texture features are extracted from visible light images, characteristic wavelengths are extracted from spectra, and after dimensionality reduction through principal component analysis, a "spectrum-image" joint feature vector is constructed. A customized pressure sensor array, combined with a laser displacement sensor, synchronously collects pressure distribution and shell deformation during the pressing process; Record pressure-time curves and deformation-time curves, extract characteristic parameters, construct a three-dimensional dynamic model, and analyze the time series characteristics through a long short-term memory network. The characteristic parameters include maximum pressure, deformation recovery rate, and pressure fluctuation frequency; The system uses metal oxide semiconductor, conductive polymer and quantum tunneling composite sensors to cover the detection range of egg spoilage markers. Wavelet transform is performed on the response curve of each sensor to extract time-frequency domain features and generate odor fingerprints. An electromagnetic vibrator was designed to scan the egg surface to stimulate resonance response. The sound pressure signal was recorded and converted into frequency domain data through fast Fourier transform. Combined with the crack depth model, the crack severity was quantified. For each test result, calculate the marginal contribution value of the input feature to the model output, quantify its contribution to the classification decision, preset business rules, combine the model output with the rule matching results, and generate the final decision.

[0010] Compared with the prior art, the present invention has the following beneficial effects: By integrating four types of sensors: spectrum (internal composition), pressure (shell strength), odor (degree of corruption), and vibration (crack depth), full-chain detection is achieved from the surface to the inside of the egg, from static characteristics to dynamic behavior, avoiding misjudgment caused by missing information from a single sensor. Preset stratification rules and priority matching of safety-related rules reduce errors caused by human intervention, thus achieving accurate, efficient, and transparent egg quality detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a flow chart of the egg quality detection system based on multi-sensor fusion of the present invention; Figure 2 This is a flow chart of the spectrum feature extraction module of the present invention; Figure 3 This is a flow chart of the pressure data acquisition module of the present invention; Figure 4 This is a flow chart of the pressure feature extraction module of the present invention; Figure 5 This is a flow chart of the odor feature extraction module of the present invention; Figure 6 This is a flow chart of the crack degree quantification module of the present invention; Figure 7 A flow chart of the quality classification generation module of the present invention. DETAILED DESCRIPTION

[0012] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0013] Reference Figure 1 As shown in the figure, the egg quality detection system based on multi-sensor fusion includes: A spectral data acquisition module, which is used to integrate a visible light camera and a near-infrared spectral sensor to synchronously collect egg surface images and internal component spectral data; A spectral feature extraction module is used to correct baseline drift by combining standard normal variable transformation, extract texture features from visible light images, extract characteristic wavelengths from spectra, and construct a "spectrum-image" joint feature vector after dimensionality reduction through principal component analysis; A pressure data acquisition module, which is used to customize a pressure sensor array and, in combination with a laser displacement sensor, synchronously acquires the pressure distribution and shell deformation during the pressing process; A pressure feature extraction module is used to record pressure-time curves and deformation-time curves, extract characteristic parameters, construct a three-dimensional dynamic model, and analyze time series characteristics through a long-short-term memory network. The characteristic parameters include maximum pressure, deformation recovery rate, and pressure fluctuation frequency; An odor feature extraction module, which uses metal oxide semiconductor, conductive polymer, and quantum tunneling composite sensors to cover the detection range of egg spoilage markers, performs wavelet transform on the response curve of each sensor, extracts time-frequency domain features, and generates an odor fingerprint; A crack severity quantification module is used to design an electromagnetic vibrator to scan the egg surface to stimulate a resonant response, record the sound pressure signal, convert it into frequency domain data through a fast Fourier transform, and quantify the crack severity by combining it with a crack depth model; The quality classification generation module is used to calculate the marginal contribution value of the input features to the model output for each test result, quantify its contribution to the classification decision, preset business rules, and combine the model output with the rule matching results to generate the final decision.

[0014] Reference Figure 2 As shown in the figure, the baseline drift is corrected by combining the standard normal variable transformation, the texture features are extracted from the visible light image, the characteristic wavelength is extracted from the spectrum, and after the dimension is reduced by principal component analysis, the "spectrum-image" joint feature vector is constructed. Specifically, the following are the steps: Gaussian filtering is used to smooth the image and convert the RGB image into a grayscale image to reduce the computational complexity and retain the texture structure information; By normalizing the mean and variance of each spectrum, baseline drift caused by uneven illumination and sensor response differences is eliminated; The standard normal variable transformation result is obtained by using the difference and standard deviation between the pixel gray value and the mean of the local area to perform ratio processing; The image is divided into blocks and the standard normal variable transformation is applied to each block to ensure the consistency of local texture features; By calculating the texture features of contrast, correlation, energy and entropy, the output is summarized as the gray-level co-occurrence matrix; The spectral shape is adjusted by linear regression to reduce the influence of scattering, and the peak and valley features of the spectrum are enhanced by using the second-order derivative; The characteristic wavelength is determined by iteratively selecting the direction with the largest change in the projection space; Initialize the wavelength set, iteratively calculate the projection vectors of the remaining wavelengths, select the wavelength with the largest projection value and add it to the set until the preset maximum number is reached; The extracted texture features and spectral feature wavelengths are concatenated into a high-dimensional joint feature vector; The data is centralized and the mean of the joint eigenvector is subtracted so that the data distribution is centered on the origin; The original high-dimensional feature vector is projected into the principal component space to obtain the reduced-dimensional "spectrum-image" joint feature.

[0015] Through statistical methods, the direction with the most significant change in the joint feature vector, namely the principal component, is found. The covariance relationship between all sample vectors is calculated to generate a covariance matrix. The matrix is decomposed to obtain the feature directions, which are sorted from large to small according to the amount of information. The top 5 main directions are retained, and the original high-dimensional vector is projected onto the selected main direction to generate a low-dimensional joint feature. This step greatly reduces the amount of calculation while retaining key information, thereby improving the efficiency of subsequent classification or prediction models.

[0016] Reference Figure 3 As shown, a customized pressure sensor array is combined with a laser displacement sensor to synchronously collect the pressure distribution and shell deformation during the pressing process. Specifically, Select a piezoresistive thin film pressure sensor and design an M×N grid based on the housing size, leaving blank areas at the edges of the array for subsequent temperature compensation and zero point calibration. Performing a zero point calibration process, collecting the output of at least one sensor in a non-pressed state, and recording the zero point offset; Use standard weights to load the array point by point and establish a pressure-voltage calibration curve; Match pressure and deformation data based on timestamps, correct for acquisition delays, and use bilinear interpolation on the pressure sensor array data to generate a high-resolution pressure distribution map; Draw the spatiotemporal evolution curves of pressure distribution and deformation field, analyze the deformation hysteresis effect at the pressing center, calculate the transfer function of deformation and pressure, and evaluate the shell structure stiffness.

[0017] With pressing time as the horizontal axis, the average pressure of the 4×4 sensor subarray in the central area and the corresponding laser-measured deformation curve are plotted to observe the deformation hysteresis effect. If the deformation peak is delayed by 50ms compared with the pressure peak, it is marked as a hysteresis response of the viscoelastic material properties, such as eggshell membrane.

[0018] Reference Figure 4 As shown, the pressure-time curve and deformation-time curve are recorded, characteristic parameters are extracted, and a three-dimensional dynamic model is constructed. The timing characteristics are analyzed through the long short-term memory network, including: Find the pressure peak from the pressure curve, record its value and occurrence time, observe the periodic fluctuations in the pressure curve, count the number of fluctuations per unit time, and calculate the time required for the pressure to rise from the initial value to the maximum value; Find the maximum deformation value from the deformation curve, record its value and the corresponding pressure value, calculate the percentage of the recovery amount to the maximum deformation value after unloading the pressure and output it as the recovery ratio, and record the time required for the deformation to recover to 10% after the pressure is unloaded to zero; Comparing the changing trends of the pressure and deformation curves and analyzing the phase relationship between the two, the phase relationship includes a synchronous relationship and a lagging relationship; Calculate the enclosed area of the pressure-deformation curve during loading and unloading to evaluate the energy loss caused by friction within the material; With time as the horizontal axis, pressure as the vertical axis, and deformation as the vertical axis, a three-dimensional coordinate system is constructed to map the synchronously collected pressure and deformation data into three-dimensional space to form a discrete data point set; The cubic spline interpolation method is used for discrete data points to generate a continuous pressure-deformation-time surface; Compare the model prediction results with the actual measurement data and calculate the deviation between the predicted and actual deformation variables; Adjust model parameters in the cubic spline interpolation method based on error analysis results; The pressure and deformation data are divided into segments of fixed length in chronological order, and the data are normalized and scaled to a value between 0 and 1. Construct a bidirectional long short-term memory network, which includes an input layer, a hidden layer, and an output layer; The input layer receives dual-channel data of pressure and deformation, the hidden layer captures temporal dependencies through memory units, and the output layer predicts the pressure and deformation values at the next moment; The dataset is divided into a training set and a test set. The training set is used to train the memory unit weights, and the test set is used to evaluate the model performance. Analyze the network's dependence on historical data, identify the key time points that affect the prediction results, and extract the dynamic relationship between pressure fluctuation frequency and deformation recovery rate through model output results.

[0019] The cubic spline interpolation method sorts the discrete point set in chronological order and uses the cubic spline interpolation method to fit the continuous surface. The interpolation polynomial is calculated piecewise to ensure the continuity of the first-order and second-order derivatives of the surface in adjacent intervals. The interpolation parameters such as node density are adjusted to balance smoothness and fitting accuracy. The deviation of the interpolated surface is compared with the original data points. If the error exceeds a threshold of 5%, the node density is increased and the interpolation is repeated.

[0020] Reference Figure 5As shown in the figure, a metal oxide semiconductor, conductive polymer and quantum tunneling composite sensor is used to cover the detection range of egg spoilage markers. The response curve of each sensor is subjected to wavelet transform to extract time-frequency domain features and generate an odor fingerprint. Specifically, the following are the steps: Fresh eggs and eggs at different stages of corruption were selected and placed in a sealed container to allow volatile gases to be fully released. The corruption stages included mild, moderate, and severe. A metal oxide semiconductor, conductive polymer, and quantum tunneling composite sensor array is placed in a container, and gas is extracted to synchronously record changes in resistance and conductivity. Perform three-layer wavelet decomposition on the response curve of each sensor to generate low-frequency approximation coefficients and high-frequency detail coefficients; The energy and mean are extracted as low-frequency features reflecting the steady-state process of gas adsorption; Standard deviation and entropy are extracted as high-frequency features that capture transient fluctuations of gas molecule collisions; Arrange the time-frequency domain features of each sensor in order to form a feature vector, and merge the feature vectors of the three sensors to construct a 9-dimensional feature matrix; A heat map is generated with sensor type as the horizontal axis and time-frequency characteristics as the vertical axis, and the color depth represents the size of the characteristic value; The polar coordinate distribution of each feature was plotted based on the spoilage stage, and the differences between fresh and spoiled eggs were compared.

[0021] Metal oxide semiconductors can be used to be sensitive to ammonia and hydrogen sulfide, reflecting gas adsorption through resistance changes. Conductive polymers can be used to detect volatile organic acids, responding through conductivity changes. Quantum tunneling composite sensors can be used to capture quantum tunneling current fluctuations caused by collisions of gas molecules.

[0022] Reference Figure 6 As shown in the figure, an electromagnetic vibrator is designed to scan the egg surface to stimulate the resonance response, record the sound pressure signal and convert it into frequency domain data through fast Fourier transform. Combined with the crack depth model, the severity of the crack is quantified. Specifically, the following are the steps: Neodymium iron boron permanent magnets are used as the main magnets to form a closed magnetic circuit with the silicon steel core. A copper coil is wound in the air gap of the core. The number of coil turns is determined based on the target exciting force and frequency range. The coil is driven by a function generator, a sine frequency sweep signal is input, and the displacement of the excitation head is monitored by a laser displacement sensor; Select an egg with at least one crack depth and spray a thin layer of graphite powder on the egg surface as an acoustic coupling medium to enhance the contact stiffness between the excitation head and the eggshell. The egg surface is divided into a 10×10 grid, and a three-axis mobile platform is used to perform vibration excitation and signal acquisition. At each grid point, the exciter outputs a sine sweep signal and simultaneously records the sound pressure signal on the eggshell surface. The displacement of the excitation head and the vibration velocity of the egg surface were recorded synchronously as auxiliary verification data; Perform fast Fourier transform analysis on each signal to generate a spectrum diagram with the horizontal axis being frequency and the vertical axis being sound pressure amplitude; Extract the amplitude peak value corresponding to each frequency point, construct a "frequency-amplitude" curve, scan the "frequency-amplitude" curve using a sliding window algorithm, and identify the frequency corresponding to the local maximum amplitude as the resonant frequency; Calculate the difference between the resonant frequency of the cracked egg and the resonant frequency of the fresh egg to reflect the reduction in stiffness caused by the crack; The damping ratio of the cracked area is calculated by the half-power bandwidth method, and the ratio of the sound pressure amplitude at the resonant frequency of the cracked area to the amplitude of the fresh area is calculated. Generate a crack distribution heat map, map the crack depth of each grid point to the egg surface, and use color depth to indicate the severity, with red indicating severe cracks and green indicating no cracks.

[0023] A function generator was used to output a sinusoidal sweep signal (100-1000 Hz, 10 Hz step size, and 10-second sweep time). The coil was driven by a power amplifier. A laser displacement sensor was installed on the excitation head to monitor the displacement in real time and feed it back to the control system to ensure that the excitation head displacement was stable within ±0.1 mm. A thin layer of graphite powder with a thickness of about 0.05 mm was evenly sprayed on the surface of the egg to enhance the acoustic coupling efficiency between the excitation head and the eggshell and reduce energy loss.

[0024] Reference Figure 7 As shown in the figure, for each test result, the marginal contribution value of the input feature to the model output is calculated to quantify its contribution to the classification decision. The business rules are preset and the final decision is generated by combining the model output with the rule matching results. Specifically, the following are included: The resonance frequency variation, vibration damping and sound pressure attenuation ratio of each egg are obtained from the electromagnetic exciter scanning system; For the trained crack classification model, analyze the influence of each feature on the final classification result one by one, and simulate the change in the model prediction accuracy after removing each feature; Count the fluctuation range of the model output results when the feature value changes in all test samples, and arrange the three features from high to low according to their contribution; Determine whether the change in resonant frequency exceeds 10 Hz and the vibration damping is 20% higher than that of a healthy egg. If so, mark it as highly cracked and destroy it. If not, do not output it. Determine whether the change in resonance frequency is between 5-10 Hz or the vibration damping is 10%-20% higher than that of healthy eggs. If so, mark it as moderate cracks and conduct manual re-inspection. If not, mark it as qualified.

[0025] For the trained crack classification model, each feature is virtually removed separately, and the feature value is replaced with the mean of the training set during prediction, while other features remain unchanged. The decrease in the accuracy of the model on the test set after removal is calculated. For all test samples, the fluctuation of the model output probability when a single feature value changes is analyzed. When the change in resonant frequency increases from 5Hz to 15Hz, the probability of the model predicting "crack" increases from 60% to 95%. The fluctuation range of this feature on the decision is 35 percentage points.

[0026] Furthermore, based on the same inventive concept as the above-mentioned egg quality detection system based on multi-sensor fusion, this solution also proposes an egg quality detection method based on multi-sensor fusion, including: Integrate a visible light camera and a near-infrared spectral sensor to simultaneously collect egg surface images and internal component spectral data; The baseline drift is corrected by combining standard normal variable transformation, texture features are extracted from visible light images, characteristic wavelengths are extracted from spectra, and after dimensionality reduction through principal component analysis, a "spectrum-image" joint feature vector is constructed. A customized pressure sensor array, combined with a laser displacement sensor, synchronously collects pressure distribution and shell deformation during the pressing process; Record pressure-time curves and deformation-time curves, extract characteristic parameters, construct a three-dimensional dynamic model, and analyze the time series characteristics through a long short-term memory network. The characteristic parameters include maximum pressure, deformation recovery rate, and pressure fluctuation frequency; The system uses metal oxide semiconductor, conductive polymer and quantum tunneling composite sensors to cover the detection range of egg spoilage markers. Wavelet transform is performed on the response curve of each sensor to extract time-frequency domain features and generate odor fingerprints. An electromagnetic vibrator was designed to scan the egg surface to stimulate resonance response. The sound pressure signal was recorded and converted into frequency domain data through fast Fourier transform. Combined with the crack depth model, the crack severity was quantified. For each test result, calculate the marginal contribution value of the input feature to the model output, quantify its contribution to the classification decision, preset business rules, combine the model output with the rule matching results, and generate the final decision.

[0027] Furthermore, the present solution also proposes a computer-readable storage medium on which a computer-readable program is stored. When the computer-readable program is called, the above-mentioned egg quality detection system based on multi-sensor fusion is executed.

[0028] It is understandable that the storage medium may be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape; an optical medium, such as a DVD; or a semiconductor medium, such as a solid state disk (SSD).

[0029] In summary, the advantages of the present invention are: by integrating four types of sensors, namely spectrum (internal composition), pressure (shell strength), odor (degree of corruption), and vibration (crack depth), full-chain detection from the surface to the inside of the egg, from static characteristics to dynamic behavior, can be achieved, avoiding misjudgment caused by missing information from a single sensor, pre-setting stratification rules, and giving priority to matching safety-related rules, reducing human intervention errors, and achieving precise, efficient, and transparent egg quality detection.

[0030] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. The egg quality detection system based on multi-sensor fusion is characterized by: include: A spectral data acquisition module, which is used to integrate a visible light camera and a near-infrared spectral sensor to synchronously collect egg surface images and internal component spectral data; A spectral feature extraction module is used to correct baseline drift by combining standard normal variable transformation, extract texture features from visible light images, extract characteristic wavelengths from spectra, and construct a "spectrum-image" joint feature vector after dimensionality reduction through principal component analysis; A pressure data acquisition module, which is used to customize a pressure sensor array and, in combination with a laser displacement sensor, synchronously acquires the pressure distribution and shell deformation during the pressing process; A pressure feature extraction module is used to record pressure-time curves and deformation-time curves, extract characteristic parameters, construct a three-dimensional dynamic model, and analyze time series characteristics through a long-short-term memory network. The characteristic parameters include maximum pressure, deformation recovery rate, and pressure fluctuation frequency; An odor feature extraction module, which uses metal oxide semiconductor, conductive polymer, and quantum tunneling composite sensors to cover the detection range of egg spoilage markers, performs wavelet transform on the response curve of each sensor, extracts time-frequency domain features, and generates an odor fingerprint; A crack severity quantification module is used to design an electromagnetic vibrator to scan the egg surface to stimulate a resonant response, record the sound pressure signal, convert it into frequency domain data through a fast Fourier transform, and quantify the crack severity by combining it with a crack depth model; The quality classification generation module is used to calculate the marginal contribution value of the input features to the model output for each test result, quantify its contribution to the classification decision, preset business rules, and combine the model output with the rule matching results to generate the final decision.

2. The egg quality detection system based on multi-sensor fusion according to claim 1, characterized in that: The method of correcting baseline drift by combining standard normal variable transformation, extracting texture features from visible light images, extracting characteristic wavelengths from spectra, and constructing a "spectrum-image" joint feature vector after dimensionality reduction by principal component analysis specifically includes: Gaussian filtering is used to smooth the image and convert the RGB image into a grayscale image to reduce the computational complexity and retain the texture structure information; By normalizing the mean and variance of each spectrum, baseline drift caused by uneven illumination and sensor response differences is eliminated; The standard normal variable transformation result is obtained by using the difference and standard deviation between the pixel gray value and the mean of the local area to perform ratio processing; The image is divided into blocks and the standard normal variable transformation is applied to each block to ensure the consistency of local texture features; By calculating the texture features of contrast, correlation, energy and entropy, the output is summarized as the gray-level co-occurrence matrix; The spectral shape is adjusted by linear regression to reduce the influence of scattering, and the peak and valley features of the spectrum are enhanced by using the second-order derivative; The characteristic wavelength is determined by iteratively selecting the direction with the largest change in the projection space; Initialize the wavelength set, iteratively calculate the projection vectors of the remaining wavelengths, select the wavelength with the largest projection value and add it to the set until the preset maximum number is reached; The extracted texture features and spectral feature wavelengths are concatenated into a high-dimensional joint feature vector; The data is centralized and the mean of the joint eigenvector is subtracted so that the data distribution is centered on the origin; The original high-dimensional feature vector is projected into the principal component space to obtain the reduced-dimensional "spectrum-image" joint feature.

3. The egg quality detection system based on multi-sensor fusion according to claim 2, characterized in that: The customized pressure sensor array, combined with the laser displacement sensor, synchronously collects the pressure distribution and shell deformation during the pressing process, specifically including: Select a piezoresistive thin film pressure sensor and design an M×N grid based on the housing size, leaving blank areas at the edges of the array for subsequent temperature compensation and zero point calibration. Performing a zero point calibration process, collecting the output of at least one sensor in a non-pressed state, and recording the zero point offset; Use standard weights to load the array point by point and establish a pressure-voltage calibration curve; Match pressure and deformation data based on timestamps, correct for acquisition delays, and use bilinear interpolation on the pressure sensor array data to generate a high-resolution pressure distribution map; Draw the spatiotemporal evolution curves of pressure distribution and deformation field, analyze the deformation hysteresis effect at the pressing center, calculate the transfer function of deformation and pressure, and evaluate the shell structure stiffness.

4. The egg quality detection system based on multi-sensor fusion according to claim 3, characterized in that: The recording of pressure-time curves and deformation-time curves, extraction of characteristic parameters, construction of a three-dimensional dynamic model, and analysis of temporal characteristics through a long short-term memory network specifically include: Find the pressure peak from the pressure curve, record its value and occurrence time, observe the periodic fluctuations in the pressure curve, count the number of fluctuations per unit time, and calculate the time required for the pressure to rise from the initial value to the maximum value; Find the maximum deformation value from the deformation curve, record its value and the corresponding pressure value, calculate the percentage of the recovery amount to the maximum deformation value after unloading the pressure and output it as the recovery ratio, and record the time required for the deformation to recover to 10% after the pressure is unloaded to zero; Comparing the changing trends of the pressure and deformation curves and analyzing the phase relationship between the two, the phase relationship includes a synchronous relationship and a lagging relationship; Calculate the enclosed area of the pressure-deformation curve during loading and unloading to evaluate the energy loss caused by friction within the material; With time as the horizontal axis, pressure as the vertical axis, and deformation as the vertical axis, a three-dimensional coordinate system is constructed to map the synchronously collected pressure and deformation data into three-dimensional space to form a discrete data point set; The cubic spline interpolation method is used for discrete data points to generate a continuous pressure-deformation-time surface; Compare the model prediction results with the actual measurement data and calculate the deviation between the predicted and actual deformation variables; Adjust model parameters in the cubic spline interpolation method based on error analysis results; The pressure and deformation data are divided into segments of fixed length in chronological order, and the data are normalized and scaled to a value between 0 and 1. Construct a bidirectional long short-term memory network, which includes an input layer, a hidden layer, and an output layer; The input layer receives dual-channel data of pressure and deformation, the hidden layer captures temporal dependencies through memory units, and the output layer predicts the pressure and deformation values at the next moment; The dataset is divided into a training set and a test set. The training set is used to train the memory unit weights, and the test set is used to evaluate the model performance. Analyze the network's dependence on historical data, identify the key time points that affect the prediction results, and extract the dynamic relationship between pressure fluctuation frequency and deformation recovery rate through model output results.

5. The egg quality detection system based on multi-sensor fusion according to claim 4, characterized in that: The method uses a metal oxide semiconductor, a conductive polymer, and a quantum tunneling composite sensor to cover the detection range of egg spoilage markers, performs wavelet transform on the response curve of each sensor, extracts time-frequency domain features, and generates an odor fingerprint, specifically including: Fresh eggs and eggs at different stages of corruption were selected and placed in a sealed container to allow volatile gases to be fully released. The corruption stages included mild, moderate, and severe. A metal oxide semiconductor, conductive polymer, and quantum tunneling composite sensor array is placed in a container, and gas is extracted to synchronously record changes in resistance and conductivity. Perform three-layer wavelet decomposition on the response curve of each sensor to generate low-frequency approximation coefficients and high-frequency detail coefficients; The energy and mean are extracted as low-frequency features reflecting the steady-state process of gas adsorption; Standard deviation and entropy are extracted as high-frequency features that capture transient fluctuations of gas molecule collisions; Arrange the time-frequency domain features of each sensor in order to form a feature vector, and merge the feature vectors of the three sensors to construct a 9-dimensional feature matrix; A heat map is generated with sensor type as the horizontal axis and time-frequency characteristics as the vertical axis, and the color depth represents the size of the characteristic value; The polar coordinate distribution of each feature was plotted based on the spoilage stage, and the differences between fresh and spoiled eggs were compared.

6. The egg quality detection system based on multi-sensor fusion according to claim 5, characterized in that: The electromagnetic vibrator is designed to scan the egg surface to stimulate resonance response, record the sound pressure signal and convert it into frequency domain data through fast Fourier transform. Combined with the crack depth model, the severity of the crack is quantified. Specifically, the following steps are involved: Neodymium iron boron permanent magnets are used as the main magnets to form a closed magnetic circuit with the silicon steel core. A copper coil is wound in the air gap of the core. The number of coil turns is determined based on the target exciting force and frequency range. The coil is driven by a function generator, a sine frequency sweep signal is input, and the displacement of the excitation head is monitored by a laser displacement sensor; Select an egg with at least one crack depth and spray a thin layer of graphite powder on the egg surface as an acoustic coupling medium to enhance the contact stiffness between the excitation head and the eggshell. The egg surface is divided into a 10×10 grid, and a three-axis mobile platform is used to perform vibration excitation and signal acquisition. At each grid point, the exciter outputs a sine sweep signal and simultaneously records the sound pressure signal on the eggshell surface. The displacement of the excitation head and the vibration velocity of the egg surface were recorded synchronously as auxiliary verification data; Perform fast Fourier transform analysis on each signal segment to generate a spectrum diagram with the horizontal axis being frequency and the vertical axis being sound pressure amplitude; Extract the amplitude peak value corresponding to each frequency point, construct a "frequency-amplitude" curve, scan the "frequency-amplitude" curve using a sliding window algorithm, and identify the frequency corresponding to the local maximum amplitude as the resonant frequency; Calculate the difference between the resonant frequency of the cracked egg and the resonant frequency of the fresh egg to reflect the reduction in stiffness caused by the crack; The damping ratio of the cracked area is calculated by the half-power bandwidth method, and the ratio of the sound pressure amplitude at the resonant frequency of the cracked area to the amplitude of the fresh area is calculated. Generate a crack distribution heat map, map the crack depth of each grid point to the egg surface, and use color depth to indicate the severity, with red indicating severe cracks and green indicating no cracks.

7. The egg quality detection system based on multi-sensor fusion according to claim 6, characterized in that: For each test result, the marginal contribution value of the input feature to the model output is calculated to quantify its contribution to the classification decision. The business rules are preset and the final decision is generated by combining the model output with the rule matching results. Specifically, the following steps are performed: The resonance frequency variation, vibration damping and sound pressure attenuation ratio of each egg are obtained from the electromagnetic exciter scanning system; For the trained crack classification model, analyze the influence of each feature on the final classification result one by one, and simulate the change in the model prediction accuracy after removing each feature; Count the fluctuation range of the model output results when the feature value changes in all test samples, and arrange the three features from high to low according to their contribution; Determine whether the change in resonant frequency exceeds 10 Hz and the vibration damping is 20% higher than that of a healthy egg. If so, mark it as highly cracked and destroy it. If not, do not output it. Determine whether the change in resonance frequency is between 5-10 Hz or the vibration damping is 10%-20% higher than that of healthy eggs. If so, mark it as moderate cracks and conduct manual re-inspection. If not, mark it as qualified.

8. An egg quality detection method based on multi-sensor fusion, used to implement the egg quality detection system based on multi-sensor fusion according to any one of claims 1 to 7, characterized in that: include: Integrate a visible light camera and a near-infrared spectral sensor to simultaneously collect egg surface images and internal component spectral data; The baseline drift is corrected by combining standard normal variable transformation, texture features are extracted from visible light images, characteristic wavelengths are extracted from spectra, and after dimensionality reduction through principal component analysis, a "spectrum-image" joint feature vector is constructed. A customized pressure sensor array, combined with a laser displacement sensor, synchronously collects pressure distribution and shell deformation during the pressing process; Record pressure-time curves and deformation-time curves, extract characteristic parameters, construct a three-dimensional dynamic model, and analyze the time series characteristics through a long short-term memory network. The characteristic parameters include maximum pressure, deformation recovery rate, and pressure fluctuation frequency; The system uses metal oxide semiconductor, conductive polymer and quantum tunneling composite sensors to cover the detection range of egg spoilage markers. Wavelet transform is performed on the response curve of each sensor to extract time-frequency domain features and generate odor fingerprints. An electromagnetic vibrator was designed to scan the egg surface to stimulate resonance response. The sound pressure signal was recorded and converted into frequency domain data through fast Fourier transform. Combined with the crack depth model, the crack severity was quantified. For each test result, calculate the marginal contribution value of the input feature to the model output, quantify its contribution to the classification decision, preset business rules, combine the model output with the rule matching results, and generate the final decision.

Images