Nondestructive testing method for conopomorpha sinensis damage based on sound wave signal analysis

Abstract

The invention relates to the field of sound wave analysis and insect pest detection, and discloses a conopomorpha sinensis damage nondestructive testing method based on sound wave signal analysis, which comprises the following steps: firstly, synchronously collecting sound wave signals and environmental noise in fruits under a sound coupling condition, extracting potential sound events, and constructing a fruit initial sound wave data packet; carrying out adaptive noise reduction and feature purification processing on the data packet to obtain a harmful sound event feature data packet; then, training a special deep convolutional loop network model by using the feature data packet, and constructing a target model capable of intelligently identifying a hazard state; and finally, performing online real-time signal acquisition and synchronous analysis on the litchi fruits in combination with the model, and automatically generating a hazard detection report. According to the method, high-sensitivity and high-specificity lossless judgment on early-stage harm of conopomorpha sinensis is realized, and a key technical support is provided for accurate prevention and control of an orchard and intelligent sorting after harvesting.

Description

Technical Field

[0001] This invention relates to the fields of acoustic wave analysis and pest detection, and in particular to a non-destructive detection method for litchi fruit borer damage based on acoustic signal analysis. Background Technology

[0002] The litchi fruit borer is one of the most serious fruit-boring pests damaging litchi fruit. Its larvae bore into the pit and pulp during the mid-to-late stages of fruit development, causing fruit drop, insect excrement contamination, and directly resulting in loss of commercial value. Effective detection of this pest is a crucial step in litchi cultivation, post-harvest processing, and quality grading. Currently, detection of the litchi fruit borer mainly relies on traditional manual fruit dissection methods. Existing technologies either fail to meet the requirements of being non-destructive, rapid, and low-cost, or lack effective signal processing and feature recognition methods when faced with the unique acoustic detection scenarios created by the litchi fruit borer in live fruit. Current methods involve randomly selecting samples for destructive dissection and visually inspecting for pest infestation. This method has significant limitations, including: destructive nature leading to sample damage, making it unsuitable for comprehensive quality inspection of commercial fruit; low efficiency, relying on manual labor, slow speed, high labor intensity, and difficulty in meeting the testing needs of large-scale orchards or sorting lines; and a high false negative rate, as early or internal damage without obvious external symptoms cannot be identified by appearance, making it highly susceptible to missed detection. Therefore, there is an urgent practical need to develop a fast, accurate, and non-destructive testing technology.

[0003] To overcome the shortcomings of traditional methods, some non-destructive testing techniques have been explored for use in detecting internal defects and pests in agricultural products. Therefore, this invention proposes a non-destructive testing method for litchi stem borer damage based on acoustic signal analysis, which can overcome the above-mentioned problems and achieve early, accurate, and practical non-destructive testing. Summary of the Invention

[0004] This invention overcomes the shortcomings of the prior art and provides a non-destructive detection method for litchi stem borer damage based on acoustic signal analysis.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides a non-destructive detection method for litchi stem borer damage based on acoustic signal analysis, comprising the following steps: Acoustic signals from lychee fruits are collected, along with ambient noise. The short-time average energy of the fruit's acoustic signals is calculated to output the initial acoustic data packets for the fruit. Adaptive acoustic event association preprocessing is performed based on the initial acoustic data packet of the fruit to obtain a harmful acoustic event feature data packet. By combining harmful sound time feature data packets, a model capable of intelligently identifying the damage status of litchi fruit borer is constructed, namely, the target damage status identification model. By combining the target hazard status identification model, online real-time signal acquisition and synchronous analysis of litchi fruit are carried out, and a hazard detection report of litchi fruit is generated based on the analysis results.

[0006] Furthermore, in a preferred embodiment of the present invention, the step of collecting fruit acoustic signals in the lychee fruit, simultaneously collecting environmental noise, and calculating the short-time average energy of the fruit acoustic signals for outputting initial fruit acoustic data packets specifically involves: Acquire the equipment used for acoustic signal acquisition, calibrate it as the target acoustic acquisition equipment, and identify the litchi fruit samples that need to be tested for litchi fruit borer damage, calibrate them as the target litchi fruit samples. The target acoustic wave acquisition device contains a high-sensitivity acoustic sensor probe. Acoustic coupling gel is injected into the probe coupling cavity, and the target lychee fruit sample is placed into the probe coupling cavity so that the equatorial region of the target lychee fruit sample is completely covered and in contact with the acoustic coupling gel. The ambient noise within a preset area around the target acoustic wave acquisition device is collected and calibrated as the target ambient noise. Simultaneously, the acoustic wave signal of the target lychee fruit sample is collected in real time through the target acoustic wave acquisition device and calibrated as the target fruit acoustic wave signal. The target environmental noise and the target fruit sound wave signal will be amplified by fixed gain, and high-pass filtering and low-pass filtering will be performed simultaneously. The target environmental noise and the target fruit sound wave signal will be conditioned by analog signal and converted into digital signals by high-speed analog-to-digital converter to obtain digital signals of environmental noise and fruit sound wave. A digital signal processing device is used to preset a standard acquisition time for the silent baseline level of the fruit sound wave digital signal. The digital signal processing device is used to calculate the mean amplitude and standard deviation of the fruit sound wave digital signal within the standard acquisition time, which are used as the silent baseline level of the fruit sound wave digital signal. Based on the silent baseline level of the fruit acoustic digital signal, the initial energy detection threshold of the fruit acoustic digital signal is preset, and the fruit acoustic digital signal is traversed through the sliding time window mode to calculate the short-time average energy of the signal within each time window. If the short-time average energy of the signal within a time window is greater than the initial energy detection threshold of the fruit sound wave digital signal, then the corresponding time window is determined to be the starting point of a potential sound event. And when the short-time average energy of the signal within the time window is not greater than the initial energy detection threshold of the fruit sound wave digital signal, it is determined to be the node of a potential sound event. The original waveform data of all potential acoustic events in the fruit's digital acoustic signal are saved, and combined with the digital signal of ambient noise, to construct the initial acoustic data package of the fruit.

[0007] Furthermore, in a preferred embodiment of the present invention, the adaptive acoustic event association preprocessing based on the initial acoustic data packet of the fruit to obtain the harmful acoustic event feature data packet specifically includes: The initial noise-reduced fruit digital signal is obtained by adaptive noise cancellation processing of the fruit sound wave digital signal and the environmental noise digital signal in the initial sound wave data packet of the fruit through the least mean square adaptive filter. In the preliminary noise reduction of the fruit digital signal, based on different potential sound events, wavelet threshold noise reduction is performed on each event segment separately. That is, the preliminary noise reduction of the fruit digital signal is decomposed into multiple layers by a preset wavelet basis function until the preliminary noise reduction of the fruit digital signal corresponding to each potential sound event is completed with wavelet threshold noise reduction, and the target noise reduction of the fruit digital signal is obtained. Each potential sound event is labeled as a purified sound event. Multi-dimensional feature calculations are performed on the digital signals of the target noise-reducing fruit corresponding to all purification sound events, including time domain features, frequency domain features, and envelope features. At the same time, a historical feature network is obtained, and known harmless litchi stem interference source sound event features are retrieved from the historical feature network to realize the construction of a harmless interference feature database. By combining the harmless interference feature database, the Mahalanobis distance feature matching calculation is performed on each purification sound event. If the Mahalanobis distance between the multi-dimensional features of the target noise-reducing fruit digital signal corresponding to the purification sound event and the features of the harmless lychee stem interference source sound event is higher than the preset value, the purification sound event is determined to be a non-pest interference event. If it is not higher, it is determined to be a pest interference event. Feature vector analysis was performed on pest disturbance events, and feature data packages of harmful sound events were constructed based on the feature vector analysis results.

[0008] Furthermore, in a preferred embodiment of the present invention, the step of performing feature vector analysis in conjunction with pest disturbance events and constructing a harmful sound event feature data packet based on the feature vector analysis results specifically includes: The time-spectrum diagram of the target noise-reduced fruit digital signal corresponding to the pest interference event is plotted to obtain the time-spectrum diagram of the pest interference event. Based on the time-spectrum diagram of the pest interference event, the target noise-reduced fruit digital signal corresponding to the pest interference event is processed by a standardized and dimension-fixed feature vector extraction process to obtain the feature vector of the pest interference event. Based on the feature vectors of all pest disturbance events, and combined with the number of pest disturbance events and the time spectrum of pest disturbance events, a feature data package of harmful sound events is constructed.

[0009] Furthermore, in a preferred embodiment of the present invention, the step of constructing a model capable of intelligently identifying the damage status of the litchi fruit borer by combining harmful sound time feature data packets, i.e., a target damage status identification model, specifically involves: Based on historical feature networks, the damage level of litchi fruit borer was determined, and biological tags were constructed based on the damage level of litchi fruit borer. A deep convolutional recurrent network algorithm is introduced to extract the feature vectors of pest interference events and the time spectrum of pest interference events from all harmful sound time feature data packets. These are then combined with biological labels for association processing to construct the initial training dataset for the deep convolutional recurrent network algorithm. At the same time, Gaussian noise within a standard threshold range is added to the feature vectors of pest interference events during the association process. The initial training dataset contains samples of all harmful sound events. Based on the deep convolutional recurrent network algorithm, a basic model of a deep convolutional recurrent network with two inputs and one output is constructed and labeled as the target basic model. An initial training dataset is imported into the target basic model, wherein the feature vector of pest interference events and the time spectrum of pest interference events in the initial training dataset are respectively input into different branches. In the branch corresponding to the feature vector of pest disturbance event, the abstract relationship between different feature vectors is learned through the target basic model, and the abstract feature vector of pest disturbance event is output. In the branch corresponding to the spectrogram of the pest interference event, a long short-term memory network is connected for convolutional analysis to capture the temporal relationship between the digital signals of the target denoised fruit corresponding to the pest interference event and output the temporal feature vector of the pest interference event. The abstract feature vector of pest interference events and the temporal feature vector of pest interference events are concatenated and used for dimensionality reduction and integration in the fully connected layer of the target basic model to output a new dataset. Based on the new dataset, the target basic model is trained iteratively with mean square error until the number of iterations meets the preset number, thereby generating a model for real-time intelligent identification of the damage level of litchi stem borer, which is labeled as the target damage status identification model.

[0010] Furthermore, in a preferred embodiment of the present invention, the step of combining the target hazard status identification model to perform online real-time signal acquisition and synchronous analysis of litchi fruit, and generating a hazard detection report for litchi fruit based on the analysis results, specifically includes: Acquire a litchi fruit borer detector and deploy a target hazard status identification model in the litchi fruit borer detector. The litchi fruit borer detector includes a target acoustic wave acquisition device and a digital signal processing device. Based on the litchi fruit borer detector, batch detection is carried out on the fruits that need to be detected for litchi fruit borers. The batch detection involves generating a harmful sound event feature data packet of litchi fruit through a target sound wave acquisition device and a digital signal processing device, and combining it with a target hazard status identification model to identify the quantitative level of litchi fruit hazard caused by litchi fruit borers. Based on the quantitative level of damage caused by litchi stem borers on different litchi fruits, a visual detection report is generated within the litchi stem borer detector.

[0011] This invention addresses the technical deficiencies in the prior art and offers the following advantages: First, by synchronously acquiring internal acoustic signals and environmental noise under acoustic coupling conditions and extracting potential acoustic events, an initial acoustic data packet for the fruit is constructed. Then, adaptive noise reduction and feature purification processing are applied to the data packet to obtain a feature data packet of harmful acoustic events. Next, this feature data packet is used to train a dedicated deep convolutional recurrent network model, constructing a target model capable of intelligently identifying hazard states. Finally, this model is combined to perform online real-time signal acquisition and synchronous analysis of litchi fruits, automatically generating a hazard detection report. This invention achieves highly sensitive, highly specific, and non-destructive identification of early-stage damage from litchi fruit borers, providing key technical support for precise orchard control and intelligent post-harvest sorting. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0013] Figure 1 A flowchart of a non-destructive detection method for litchi stem borer damage based on acoustic signal analysis is shown. Figure 2 A flowchart illustrating the method for constructing a data packet containing harmful sound event characteristics is shown. Detailed Implementation

[0014] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0015] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0016] Figure 1 A flowchart of a non-destructive detection method for litchi fruit borer damage based on acoustic signal analysis is shown, including the following steps: S102: Collect fruit acoustic signals in lychee fruit, collect environmental noise at the same time, and calculate the short-time average energy of the fruit acoustic signals to output the initial fruit acoustic data packets. S104: Adaptive acoustic event association preprocessing is performed based on the initial acoustic data packet of the fruit to obtain a harmful acoustic event feature data packet; S106: Combining harmful sound time feature data packages, a model capable of intelligently identifying the damage status of litchi fruit borer is constructed, namely, the target damage status identification model. S108: Combining the target hazard status identification model, online real-time signal acquisition and synchronous analysis of litchi fruit are performed, and a hazard detection report of litchi fruit is generated based on the analysis results.

[0017] Furthermore, in a preferred embodiment of the present invention, the step of collecting fruit acoustic signals in the lychee fruit, simultaneously collecting environmental noise, and calculating the short-time average energy of the fruit acoustic signals for outputting initial fruit acoustic data packets specifically involves: Acquire the equipment used for acoustic signal acquisition, calibrate it as the target acoustic acquisition equipment, and identify the litchi fruit samples that need to be tested for litchi fruit borer damage, calibrate them as the target litchi fruit samples. The target acoustic wave acquisition device contains a high-sensitivity acoustic sensor probe. Acoustic coupling gel is injected into the probe coupling cavity, and the target lychee fruit sample is placed into the probe coupling cavity so that the equatorial region of the target lychee fruit sample is completely covered and in contact with the acoustic coupling gel. The ambient noise within a preset area around the target acoustic wave acquisition device is collected and calibrated as the target ambient noise. Simultaneously, the acoustic wave signal of the target lychee fruit sample is collected in real time through the target acoustic wave acquisition device and calibrated as the target fruit acoustic wave signal. The target environmental noise and the target fruit sound wave signal will be amplified by fixed gain, and high-pass filtering and low-pass filtering will be performed simultaneously. The target environmental noise and the target fruit sound wave signal will be conditioned by analog signal and converted into digital signals by high-speed analog-to-digital converter to obtain digital signals of environmental noise and fruit sound wave. A digital signal processing device is used to preset a standard acquisition time for the silent baseline level of the fruit sound wave digital signal. The digital signal processing device is used to calculate the mean amplitude and standard deviation of the fruit sound wave digital signal within the standard acquisition time, which are used as the silent baseline level of the fruit sound wave digital signal. Based on the silent baseline level of the fruit acoustic digital signal, the initial energy detection threshold of the fruit acoustic digital signal is preset, and the fruit acoustic digital signal is traversed through the sliding time window mode to calculate the short-time average energy of the signal within each time window. If the short-time average energy of the signal within a time window is greater than the initial energy detection threshold of the fruit sound wave digital signal, then the corresponding time window is determined to be the starting point of a potential sound event. And when the short-time average energy of the signal within the time window is not greater than the initial energy detection threshold of the fruit sound wave digital signal, it is determined to be the node of a potential sound event. The original waveform data of all potential acoustic events in the fruit's digital acoustic signal are saved, and combined with the digital signal of ambient noise, to construct the initial acoustic data package of the fruit.

[0018] It should be noted that fruit samples were obtained to pre-construct a non-destructive detection method for litchi fruit borers based on acoustic signal analysis. Acoustic coupling gel was injected into the probe coupling cavity because litchi pulp attenuates sound waves, and the acoustic impedance difference between air and the peel causes most of the signal to be lost due to interface reflection. Injecting the acoustic coupling gel and achieving complete contact in the fruit's largest diameter region (equatorial region) created an acoustic impedance-matched conduction channel, improving the sensitivity of the acquired signal. Simultaneous acquisition of environmental noise and fruit acoustic signals allowed for the algorithmic differentiation and removal of interference components related to background noise, which were then converted into digital signals, providing a foundation for subsequent acoustic wave analysis. A baseline was set at the beginning of the detection, including the mean and standard deviation. Subsequently, a sliding time window mode was used to traverse the fruit's digital acoustic signals, calculating the short-time average energy of the signal within each time window to capture sudden insect-related acoustic events. Its advantage is that it can automatically and objectively locate all short-time segments that may contain the target signal from a long-term sequence, achieving preliminary data compression and focusing, eliminating long periods of signal-free stationary periods, and greatly improving subsequent processing efficiency. This invention enables targeted acoustic signal acquisition of litchi fruit borers in complex field environments.

[0019] Furthermore, in a preferred embodiment of the present invention, the step of constructing a model capable of intelligently identifying the damage status of the litchi fruit borer by combining harmful sound time feature data packets, i.e., a target damage status identification model, specifically involves: Based on historical feature networks, the damage level of litchi fruit borer was determined, and biological tags were constructed based on the damage level of litchi fruit borer. A deep convolutional recurrent network algorithm is introduced to extract the feature vectors of pest interference events and the time spectrum of pest interference events from all harmful sound time feature data packets. These are then combined with biological labels for association processing to construct the initial training dataset for the deep convolutional recurrent network algorithm. At the same time, Gaussian noise within a standard threshold range is added to the feature vectors of pest interference events during the association process. The initial training dataset contains samples of all harmful sound events. Based on the deep convolutional recurrent network algorithm, a basic model of a deep convolutional recurrent network with two inputs and one output is constructed and labeled as the target basic model. An initial training dataset is imported into the target basic model, wherein the feature vector of pest interference events and the time spectrum of pest interference events in the initial training dataset are respectively input into different branches. In the branch corresponding to the feature vector of pest disturbance event, the abstract relationship between different feature vectors is learned through the target basic model, and the abstract feature vector of pest disturbance event is output. In the branch corresponding to the spectrogram of the pest interference event, a long short-term memory network is connected for convolutional analysis to capture the temporal relationship between the digital signals of the target denoised fruit corresponding to the pest interference event and output the temporal feature vector of the pest interference event. The abstract feature vector of pest interference events and the temporal feature vector of pest interference events are concatenated and used for dimensionality reduction and integration in the fully connected layer of the target basic model to output a new dataset. Based on the new dataset, the target basic model is trained iteratively with mean square error until the number of iterations meets the preset number, thereby generating a model for real-time intelligent identification of the damage level of litchi stem borer, which is labeled as the target damage status identification model.

[0020] It should be noted that biological labels are constructed based on the quantitative damage level of the litchi fruit borer, including healthy, mildly damaged / young larvae, and severely damaged / old larvae, to facilitate damage classification after model detection. Feature vectors and time-spectrum maps of pest disturbance events are extracted and associated with the biological labels, allowing the model to understand the same event from different perspectives and improving recognition accuracy. During the association process, Gaussian noise within a standard threshold range is added to the feature vectors of pest disturbance events to slightly perturb the training data, simulating subtle variations in real-world data collection, improving robustness in real-world, variable environments and preventing overfitting. In the model, the feature vector branch primarily learns the abstract combination and high-level meaning of handcrafted features, while the time-spectrum map branch automatically learns local spatiotemporal patterns (such as the shape and texture of energy patches) in the time-spectrum map through a convolutional neural network. This dual-branch architecture synergistically utilizes precise quantitative features (feature vectors) and rich texture morphology information (time-spectrum map), fully leveraging the advantages of different data formats to provide a more comprehensive information foundation for making more accurate judgments. In subsequent steps, a Long Short-Term Memory (LSTM) network, a recurrent neural network adept at processing sequential data, is integrated into the temporal spectrum branch to capture the energy evolution over time in acoustic events. Finally, the abstract features and temporal features learned from the two branches are fused, and a fully connected layer is used for comprehensive decision-making to improve the model's discrimination accuracy and specificity. Finally, mean squared error iterative training is performed until the preset number of iterations is met, achieving model performance stabilization and yielding the target hazard state identification model.

[0021] Furthermore, in a preferred embodiment of the present invention, the step of combining the target hazard status identification model to perform online real-time signal acquisition and synchronous analysis of litchi fruit, and generating a hazard detection report for litchi fruit based on the analysis results, specifically includes: Acquire a litchi fruit borer detector and deploy a target hazard status identification model in the litchi fruit borer detector. The litchi fruit borer detector includes a target acoustic wave acquisition device and a digital signal processing device. Based on the litchi fruit borer detector, batch detection is carried out on the fruits that need to be detected for litchi fruit borers. The batch detection involves generating a harmful sound event feature data packet of litchi fruit through a target sound wave acquisition device and a digital signal processing device, and combining it with a target hazard status identification model to identify the quantitative level of litchi fruit hazard caused by litchi fruit borers. Based on the quantitative level of damage caused by litchi stem borers on different litchi fruits, a visual detection report is generated within the litchi stem borer detector.

[0022] It should be noted that the detector integrates target acoustic wave acquisition equipment and digital signal processing equipment. It is used to acquire acoustic waves from litchi fruit and perform digital signal processing. The resulting signal data packets are then imported into a model for analysis to quantitatively identify the level of damage caused by the litchi fruit stem borer. A visual detection report is subsequently generated, including the final damage status, the spectrum of the pest interference events, the total number of detected harmful acoustic events, the average energy, and the model confidence level. The purpose of this report is to support operator decision-making.

[0023] Figure 2 A flowchart illustrating a method for constructing a data packet containing harmful sound event characteristics is shown, including the following steps: S202: Adaptive acoustic event association preprocessing is performed based on the initial acoustic data packet of the fruit to obtain a harmful acoustic event feature data packet; S204: Combine pest disturbance events with feature vector analysis, and construct a feature data package of harmful sound events based on the feature vector analysis results.

[0024] Furthermore, in a preferred embodiment of the present invention, the adaptive acoustic event association preprocessing based on the initial acoustic data packet of the fruit to obtain the harmful acoustic event feature data packet specifically includes: The initial noise-reduced fruit digital signal is obtained by adaptive noise cancellation processing of the fruit sound wave digital signal and the environmental noise digital signal in the initial sound wave data packet of the fruit through the least mean square adaptive filter. In the preliminary noise reduction of the fruit digital signal, based on different potential sound events, wavelet threshold noise reduction is performed on each event segment separately. That is, the preliminary noise reduction of the fruit digital signal is decomposed into multiple layers by a preset wavelet basis function until the preliminary noise reduction of the fruit digital signal corresponding to each potential sound event is completed with wavelet threshold noise reduction, and the target noise reduction of the fruit digital signal is obtained. Each potential sound event is labeled as a purified sound event. Multi-dimensional feature calculations are performed on the digital signals of the target noise-reducing fruit corresponding to all purification sound events, including time domain features, frequency domain features, and envelope features. At the same time, a historical feature network is obtained, and known harmless litchi stem interference source sound event features are retrieved from the historical feature network to realize the construction of a harmless interference feature database. By combining the harmless interference feature database, the Mahalanobis distance feature matching calculation is performed on each purification sound event. If the Mahalanobis distance between the multi-dimensional features of the target noise-reducing fruit digital signal corresponding to the purification sound event and the features of the harmless lychee stem interference source sound event is higher than the preset value, the purification sound event is determined to be a non-pest interference event. If it is not higher, it is determined to be a pest interference event. Feature vector analysis was performed on pest disturbance events, and feature data packages of harmful sound events were constructed based on the feature vector analysis results.

[0025] It should be noted that the least mean square adaptive filter can dynamically learn how the environmental noise reference signal "leaks" or "couples" into the main signal channel of the fruit, and generate a corresponding noise estimate in real time, which is then subtracted from the main signal. Simultaneously, wavelet thresholding denoising is performed on the latent sound time because random noise and impulse interference unrelated to the reference noise still remain in the signal. Wavelet transform has excellent time-frequency localization analysis capabilities. This step performs multi-level wavelet decomposition on each independent "latest sound event" segment, expanding the signal at different scales (frequency bands), then thresholding (setting to zero or reducing) the wavelet coefficients representing noise, and finally reconstructing the signal to obtain the purified sound event. Multi-dimensional feature calculations are performed on the purified sound events to build a database, and Mahalanobis distance is used to analyze whether the event is a pest interference event. Mahalanobis distance is a distance metric that considers the correlation between feature dimensions.

[0026] Furthermore, in a preferred embodiment of the present invention, the step of performing feature vector analysis in conjunction with pest disturbance events and constructing a harmful sound event feature data packet based on the feature vector analysis results specifically includes: The time-spectrum diagram of the target noise-reduced fruit digital signal corresponding to the pest interference event is plotted to obtain the time-spectrum diagram of the pest interference event. Based on the time-spectrum diagram of the pest interference event, the target noise-reduced fruit digital signal corresponding to the pest interference event is processed by a standardized and dimension-fixed feature vector extraction process to obtain the feature vector of the pest interference event. Based on the feature vectors of all pest disturbance events, and combined with the number of pest disturbance events and the time spectrum of pest disturbance events, a feature data package of harmful sound events is constructed.

[0027] It should be noted that the spectrograms and feature vectors of pest disturbance events can be used to construct harmful sound event feature data packages. The advantage is that this achieves modular output of the data processing workflow. The "harmful sound event feature data package" is a standardized interface that contains all rigorously refined and feature-processed information, which can be directly and efficiently input into subsequent deep learning models.

[0028] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A non-destructive detection method for litchi stem borer damage based on acoustic signal analysis, characterized in that, Includes the following steps: Acoustic signals from lychee fruits are collected, along with ambient noise. The short-time average energy of the fruit's acoustic signals is calculated to output the initial acoustic data packets for the fruit. Adaptive acoustic event association preprocessing is performed based on the initial acoustic data packet of the fruit to obtain a harmful acoustic event feature data packet. By combining harmful sound time feature data packets, a model capable of intelligently identifying the damage status of litchi fruit borer is constructed, namely, the target damage status identification model. By combining the target hazard status identification model, online real-time signal acquisition and synchronous analysis of litchi fruit are carried out, and a hazard detection report of litchi fruit is generated based on the analysis results.

2. The non-destructive detection method for litchi stem borer damage based on acoustic signal analysis as described in claim 1, characterized in that, The process involves collecting acoustic signals from lychee fruits, simultaneously collecting ambient noise, and calculating the short-time average energy of the fruit's acoustic signals to output initial acoustic data packets for the fruit. Specifically: Acquire the equipment used for acoustic signal acquisition, calibrate it as the target acoustic acquisition equipment, and identify the litchi fruit samples that need to be tested for litchi fruit borer damage, calibrate them as the target litchi fruit samples. The target acoustic wave acquisition device contains a high-sensitivity acoustic sensor probe. Acoustic coupling gel is injected into the probe coupling cavity, and the target lychee fruit sample is placed into the probe coupling cavity so that the equatorial region of the target lychee fruit sample is completely covered and in contact with the acoustic coupling gel. The ambient noise within a preset area around the target acoustic wave acquisition device is collected and calibrated as the target ambient noise. Simultaneously, the acoustic wave signal of the target lychee fruit sample is collected in real time through the target acoustic wave acquisition device and calibrated as the target fruit acoustic wave signal. The target environmental noise and the target fruit sound wave signal will be amplified by fixed gain, and high-pass filtering and low-pass filtering will be performed simultaneously. The target environmental noise and the target fruit sound wave signal will be conditioned by analog signal and converted into digital signals by high-speed analog-to-digital converter to obtain digital signals of environmental noise and fruit sound wave. A digital signal processing device is used to preset a standard acquisition time for the silent baseline level of the fruit sound wave digital signal. The digital signal processing device is used to calculate the mean amplitude and standard deviation of the fruit sound wave digital signal within the standard acquisition time, which are used as the silent baseline level of the fruit sound wave digital signal. Based on the silent baseline level of the fruit acoustic digital signal, the initial energy detection threshold of the fruit acoustic digital signal is preset, and the fruit acoustic digital signal is traversed through the sliding time window mode to calculate the short-time average energy of the signal within each time window. If the short-time average energy of the signal within a time window is greater than the initial energy detection threshold of the fruit sound wave digital signal, then the corresponding time window is determined to be the starting point of a potential sound event. And when the short-time average energy of the signal within the time window is not greater than the initial energy detection threshold of the fruit sound wave digital signal, it is determined to be the node of a potential sound event. The original waveform data of all potential acoustic events in the fruit's digital acoustic signal are saved, and combined with the digital signal of ambient noise, to construct the initial acoustic data package of the fruit.

3. The non-destructive detection method for litchi stem borer damage based on acoustic signal analysis as described in claim 1, characterized in that, The adaptive acoustic event association preprocessing based on the initial acoustic data packet of the fruit, to obtain the harmful acoustic event feature data packet, specifically involves: The initial noise-reduced fruit digital signal is obtained by adaptive noise cancellation processing of the fruit sound wave digital signal and the environmental noise digital signal in the initial sound wave data packet of the fruit through the least mean square adaptive filter. In the preliminary noise reduction of the fruit digital signal, based on different potential sound events, wavelet threshold noise reduction is performed on each event segment separately. That is, the preliminary noise reduction of the fruit digital signal is decomposed into multiple layers by a preset wavelet basis function until the preliminary noise reduction of the fruit digital signal corresponding to each potential sound event is completed with wavelet threshold noise reduction, and the target noise reduction of the fruit digital signal is obtained. Each potential sound event is labeled as a purified sound event. Multi-dimensional feature calculations are performed on the digital signals of the target noise-reducing fruit corresponding to all purification sound events, including time domain features, frequency domain features, and envelope features. At the same time, a historical feature network is obtained, and known harmless litchi stem interference source sound event features are retrieved from the historical feature network to realize the construction of a harmless interference feature database. By combining the harmless interference feature database, the Mahalanobis distance feature matching calculation is performed on each purification sound event. If the Mahalanobis distance between the multi-dimensional features of the target noise-reducing fruit digital signal corresponding to the purification sound event and the features of the harmless lychee stem interference source sound event is higher than the preset value, the purification sound event is determined to be a non-pest interference event. If it is not higher, it is determined to be a pest interference event. Feature vector analysis was performed on pest disturbance events, and feature data packages of harmful sound events were constructed based on the feature vector analysis results.

4. The non-destructive detection method for litchi stem borer damage based on acoustic signal analysis as described in claim 3, characterized in that, The process involves combining pest disturbance events with feature vector analysis, and constructing a harmful sound event feature data package based on the feature vector analysis results. Specifically: The time-spectrum diagram of the target noise-reduced fruit digital signal corresponding to the pest interference event is plotted to obtain the time-spectrum diagram of the pest interference event. Based on the time-spectrum diagram of the pest interference event, the target noise-reduced fruit digital signal corresponding to the pest interference event is processed by a standardized and dimension-fixed feature vector extraction process to obtain the feature vector of the pest interference event. Based on the feature vectors of all pest disturbance events, and combined with the number of pest disturbance events and the time spectrum of pest disturbance events, a feature data package of harmful sound events is constructed.

5. The non-destructive detection method for litchi stem borer damage based on acoustic signal analysis as described in claim 1, characterized in that, The model, which combines harmful sound time feature data packets to construct a model for intelligent identification of the damage status of litchi fruit borers, namely the target damage status identification model, is as follows: Based on historical feature networks, the damage level of litchi fruit borer was determined, and biological tags were constructed based on the damage level of litchi fruit borer. A deep convolutional recurrent network algorithm is introduced to extract the feature vectors of pest interference events and the time spectrum of pest interference events from all harmful sound time feature data packets. These are then combined with biological labels for association processing to construct the initial training dataset for the deep convolutional recurrent network algorithm. At the same time, Gaussian noise within a standard threshold range is added to the feature vectors of pest interference events during the association process. The initial training dataset contains samples of all harmful sound events. Based on the deep convolutional recurrent network algorithm, a basic model of a deep convolutional recurrent network with two inputs and one output is constructed and labeled as the target basic model. An initial training dataset is imported into the target basic model, wherein the feature vector of pest interference events and the time spectrum of pest interference events in the initial training dataset are respectively input into different branches. In the branch corresponding to the feature vector of pest disturbance event, the abstract relationship between different feature vectors is learned through the target basic model, and the abstract feature vector of pest disturbance event is output. In the branch corresponding to the spectrogram of the pest interference event, a long short-term memory network is connected for convolutional analysis to capture the temporal relationship between the digital signals of the target denoised fruit corresponding to the pest interference event and output the temporal feature vector of the pest interference event. The abstract feature vector of pest interference events and the temporal feature vector of pest interference events are concatenated and used for dimensionality reduction and integration in the fully connected layer of the target basic model to output a new dataset. Based on the new dataset, the target basic model is trained iteratively with mean square error until the number of iterations meets the preset number, thereby generating a model for real-time intelligent identification of the damage level of litchi stem borer, which is labeled as the target damage status identification model.

6. The non-destructive detection method for litchi stem borer damage based on acoustic signal analysis as described in claim 1, characterized in that, The method involves combining a target hazard status identification model to perform online real-time signal acquisition and synchronous analysis of litchi fruits, and generating a hazard detection report for litchi fruits based on the analysis results. Specifically: Acquire a litchi fruit borer detector and deploy a target hazard status identification model in the litchi fruit borer detector. The litchi fruit borer detector includes a target acoustic wave acquisition device and a digital signal processing device. Based on the litchi fruit borer detector, batch detection is carried out on the fruits that need to be detected for litchi fruit borers. The batch detection involves generating a harmful sound event feature data packet of litchi fruit through a target sound wave acquisition device and a digital signal processing device, and combining it with a target hazard status identification model to identify the quantitative level of litchi fruit hazard caused by litchi fruit borers. Based on the quantitative level of damage caused by litchi stem borers on different litchi fruits, a visual detection report is generated within the litchi stem borer detector.

Images