Forest trunk borer audio frequency identification method and system based on edge calculation

Abstract

The invention discloses a forest trunk borer audio frequency identification method and system based on edge calculation, and relates to the technical field of audio frequency identification, and the method comprises the following steps: building a simultaneous model between a temperature and humidity coupling parameter and an electroacoustic response, collecting a mutation sample through an extreme environment condition, and generating an audio frequency fingerprint database of a self-oscillation risk; and distributing the audio frequency fingerprint database to an edge computing node, extracting field acoustic data through a dual-channel synchronous sampling mode of a reference microphone channel and a reference cavity channel, and computing the center frequency of narrow-band interference. According to the method, the audio frequency fingerprints are generated by constructing the temperature and humidity-electroacoustic simultaneous model, the interference frequency is accurately positioned by combining edge sampling and phase-locked tracking, the insect sound envelope is extracted through time-frequency decomposition, the adjustment instruction is generated through environment prediction, disturbance-induced frequency hopping is injected, residual interference is filtered out, the insect sound recognition closed loop is constructed, and the recognition stability and accuracy are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of audio recognition, specifically to a forest borer pest audio recognition method and system based on edge computing. BACKGROUND

[0002] The "forest borer pest audio recognition based on edge computing" is a method of identifying borer pest activity by real-time processing and analyzing audio signals in forest environments on edge devices. This technology uses acoustic sensors to collect unique audio data generated by pests feeding on wood in the forest, and performs efficient local data processing and analysis through an edge computing platform to quickly identify pest activity without transmitting large amounts of data to remote servers. This method can provide real-time monitoring and feedback in the forest, helping forest managers to detect pest invasions in a timely manner and take appropriate control measures, thereby effectively reducing the loss of forest resources.

[0003] The prior art has the following disadvantages: In the prior art, when the forest environment experiences a sudden drop in temperature accompanied by a significant increase in air humidity, the acoustic sensor is easily affected by temperature and humidity coupling and produces self-excited oscillation, forming narrowband false sound signals in a specific frequency band. Because this type of false sound signal covers or interferes with the pulse sound produced by borer pest larvae biting, the recognition system cannot accurately extract the target acoustic features, resulting in long-term false negatives. Once this false negative persists in the monitoring process, it may cause the spread of forest pests to adjacent forest plots, resulting in widespread damage and causing serious ecological and economic losses.

[0004] The above information disclosed in the background section is only used to enhance the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present application is to provide a forest borer pest audio recognition method and system based on edge computing to solve the problems in the background technology.

[0006] In order to achieve the above-mentioned purpose, the present application provides the following technical solution: a forest borer pest audio recognition method based on edge computing, comprising the following steps: S100, a simultaneous model between temperature and humidity coupling parameters and electro-acoustic response is established, extreme environmental condition samples are collected to generate a self-excited oscillation risk audio fingerprint library; S200, distribute the audio fingerprint library to the edge computing node, extract the on-site acoustic data through the dual-channel synchronous sampling method of the reference microphone channel and the reference cavity channel, and calculate the center frequency of the narrowband interference; S300 performs time-varying phase-locked tracking based on the center frequency, calculates the phase noise spectral density, and reverse-engineers the time-series curve of the intensity of the self-excited oscillation source changing with time. S400 performs sparse time-frequency decomposition based on time-series curves, suppresses interference frequency bands caused by resonance, and extracts pulsed acoustic envelopes related to larval biting. The S500 inputs the insect sound envelope into the edge sensor control module, combines it with real-time temperature and humidity parameters to perform predictive inference, and generates a set of adjustment parameters for control. The S600 uses dynamic control based on a set of adjustment parameters. By injecting random phase disturbances and superimposing nanosecond-level temperature and humidity steps, it guides the self-excited resonant frequency jump, eliminates residual narrowband interference, and forms a closed-loop identification process of modeling, sampling, tracking, decomposition, prediction and control.

[0007] Preferably, step S100 includes: An acoustic response testing device was built in a closed environment control chamber, and acoustic sampling operations were performed using temperature control components, humidity control components, heat transfer units and multiple sets of audio sensors. The temperature was adjusted from room temperature to low temperature and then raised to high temperature, while the humidity was adjusted from low humidity to high humidity. The above parameters were kept constant at different time periods, and the audio signal and environmental parameter curves under the sudden change environmental conditions were collected. The collected audio data is subjected to anomaly screening, interference classification and signal comparison operations. Invalid samples that are not repeatable or have no clear cause are removed, and interference signals with stable characteristics and unique causes are extracted. Construct audio fingerprint feature entries that include raw data of interference signals, spectrograms, response curves, temperature and humidity change trajectories, and energy density information, and complete the construction of the audio fingerprint feature library.

[0008] Preferably, step S200 includes: The audio fingerprint feature database contains multiple data units, including interference signals, spectral features, environmental factors, and frequency markers, which are compressed and stored in a structured manner and then synchronously transmitted to the front-end sampling equipment deployed in the forest area via wireless communication. A reference acoustic acquisition channel and a reference acoustic acquisition channel are established in the edge sampling device, and dual-channel synchronous sampling is achieved through a co-source clock triggering mechanism to acquire real acoustic signals from the forest area and non-interference reference acoustic signals respectively; Perform dual-channel difference analysis and spectrum calculation, extract the interference enhancement waveform and compare it with the pre-stored fingerprint sample, and output the current interference center frequency; Configure the frequency focusing parameters of the subsequent sampling window based on the center frequency, and generate local interference snapshots to provide the initial data basis for the frequency tracking stage.

[0009] Preferably, step S300 includes: The identified interference center frequency is used as the starting frequency input parameter for the current frequency band analysis. The sampling bandwidth is set within the range above and below the center frequency and the sampling rate is increased. Bandpass sampling is performed on the reference acoustic channel and the reference acoustic channel according to a fixed time window to obtain the target frequency band signal. Each sampled waveform is periodically expanded and the period start point, peak, trough and phase intersection point are extracted. The phase shift is measured with the center frequency as a reference and the time point is marked. The phase fluctuation trend and phase change rate are recorded. After obtaining the phase change sequence, a mapping relationship between the phase change and the energy output of the interference source is established. The waveform energy after dual-channel differential processing is used as a correction factor to calculate the intensity of the interference source and generate a time evolution curve. The time evolution curve of interference intensity is used to guide the target signal extraction and noise cancellation in the next stage, and the sampling bandwidth and passband centroid are adjusted accordingly to achieve dynamic frequency locking and identification window optimization.

[0010] Preferably, the phase shift measurement uses an ideal sine wave as a reference model, calculates the phase difference for each cycle, and records the phase change with millisecond-level resolution to determine the stability and burst characteristics of the interference signal.

[0011] Preferably, step S400 includes: The audio sampling data segments are divided according to the self-excited source intensity time-series curve, and high-intensity interference periods are identified. Short-time Fourier transform is performed on each segment to extract the interference frequency band with fixed frequency position and high energy density. For the identified interference frequency bands, a bandwidth attenuation filtering method is used to reduce the frequency band, dynamically adjust the filtering window, and control the interference energy to decrease to below the set threshold. Continuous time window analysis is performed on the waveform after interference suppression. Based on amplitude, rise time, waveform asymmetry and dominant frequency range, candidate signals of insect sound pulses are identified and the insect sound envelope segment is reconstructed. The frequency slope change and energy density characteristics of the insect sound envelope are structurally verified, and segments that do not conform to the insect sound characteristics are removed to generate insect sound recognition data with confidence for subsequent processing.

[0012] Preferably, step S500 includes: The insect sound envelope data is synchronized with real-time temperature and humidity data on a time axis to construct a joint data unit of insect sound characteristics and environmental conditions; Risk levels are determined based on the duration of insect sounds, the range of dominant frequencies, the number of pulses, and the ambient temperature and humidity levels. The evolution trend of insect sounds is analyzed to identify the active state of insect pests. Based on the active state, three types of adjustment parameters are generated: sampling sensitivity, identification threshold, and response delay, to adjust the signal acquisition and feedback response strategy. The adjustment parameters are dynamically updated based on the changing trends of insect sounds, and the confirmed adjustment parameters are distributed to each identification and execution channel in the form of instructions to complete the control behavior.

[0013] Preferably, the sampling sensitivity parameter in the parameter set is achieved by increasing the sampling bit depth and compressing the sampling frequency band, the identification threshold parameter is achieved by lowering the insect sound amplitude threshold, and the response delay parameter is achieved by shortening the time interval between identification and execution.

[0014] Preferably, step S600 includes: Injecting a perturbation signal with consistent frequency and controlled phase causes the sensing element to enter a slightly unstable state. Simultaneously applying a temperature rise and a humidity jump alters the resonance characteristics of the acoustic sensing components, inducing a frequency jump and breaking frequency lock. Based on the frequency offset, the inverted signal structure is invoked to perform interference removal processing such as amplitude preservation, time alignment, and phase reversal in real time; Record the disturbance response state, temperature and humidity jump amplitude, interference removal degree and identification confidence level to complete a closed-loop operation of sampling and identification strategy update.

[0015] The edge computing-based acoustic recognition system for forest borers includes a modeling fingerprint generation module, an interference identification and frequency localization module, a phase-locked tracking and time-series modeling module, a time-frequency decomposition and insect sound extraction module, a predictive reasoning and instruction generation module, and a dynamic control and interference suppression module. The modeling fingerprint generation module establishes a joint model between temperature and humidity coupling parameters and electroacoustic response, and generates an acoustic fingerprint database of self-excited oscillation risk by collecting mutation samples under extreme environmental conditions. The interference identification and frequency localization module distributes the audio fingerprint database to the edge computing nodes, extracts the on-site acoustic data through a dual-channel synchronous sampling method of the reference microphone channel and the reference cavity channel, and calculates the center frequency of the narrowband interference. Phase-locked tracking and timing modeling module 0 performs time-varying phase-locked tracking based on the center frequency, calculates the phase noise spectral density, and reverse-engineers the timing curve of the self-excited oscillation source intensity changing with time; The time-frequency decomposition and insect sound extraction module performs sparse time-frequency decomposition based on time-series curves, suppresses interference frequency bands caused by resonance, and extracts pulse-type acoustic envelopes related to larval biting. The predictive reasoning and instruction generation module inputs the insect sound envelope into the edge sensor control module, combines it with real-time temperature and humidity parameters to perform predictive reasoning, and generates a set of adjustment parameters for control. The dynamic control and interference suppression module performs dynamic control based on the set of adjustment parameters. By injecting random phase disturbances and superimposing nanosecond-level temperature and humidity steps, it guides the self-excited resonant frequency jump and eliminates residual narrowband interference.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention constructs a combined model of temperature, humidity, and electroacoustic response and generates an acoustic fingerprint database, enabling the system to pre-identify self-excitation risks. Furthermore, it utilizes dual-channel synchronous sampling and phase-locked tracking technology at edge nodes to accurately locate the interference center frequency. By combining time-varying phase analysis and sparse time-frequency decomposition, the weak insect sound pulse envelope corresponding to larval biting behavior can be recovered from the interfered original signal, significantly improving the accuracy of feature extraction. Based on this, dynamic adjustment commands are generated through the fusion analysis of insect sounds and environmental parameters, and phase perturbation and temperature / humidity jump strategies are introduced to interrupt frequency hopping and remove residual interference sources, giving the entire acoustic recognition process closed-loop and adaptive capabilities. Compared with existing technologies, this method not only improves the recognition stability and anti-interference ability under complex climatic conditions in forest areas but also significantly reduces the false negative and false positive rates, providing efficient and reliable technical support for early warning and precise control of forest pests. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of the method for audio recognition of forest borers based on edge computing according to the present invention.

[0019] Figure 2 This is a schematic diagram of the module of the forest borer audio recognition system based on edge computing of the present invention. Detailed Implementation

[0020] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0021] This invention provides, for example Figure 1 and Figure 2 The edge computing-based acoustic identification method for forest trunk borers, as shown, includes the following steps: S100 establishes a simultaneous mathematical model between temperature and humidity coupling parameters and electroacoustic response in the edge computing system. It collects mutation samples under artificially set extreme environmental conditions, generates an acoustic fingerprint feature library of self-excited oscillation risk, and constructs a preset reference dataset for interference discrimination. To address the acoustic interference caused by the combined effects of sudden temperature drops and increased humidity in forest environments, a method is proposed to establish an acoustic fingerprint feature database based on the relationship between temperature-humidity coupling and electroacoustic response. The method includes the following steps: An acoustic response testing device was constructed within a sealed environmental control chamber, comprising a temperature control component, a humidity control component, a heat transfer unit, and multiple sets of acoustic sensors. The temperature control component uses an electrothermal film combined with a miniature fan to controllably heat and cool the air within the chamber. The humidity control component uses a controllable atomizer and dehumidifier to jointly control humidity changes. The heat transfer unit employs a combination of convection and radiation to ensure stable and controllable temperature gradients in different spatial areas. For the acoustic sensors, each location is equipped with a set of standard sensors and a set of response comparison sensors, used to collect actual response data and reference data, respectively. In this device, the temperature is first lowered from room temperature to -10 degrees Celsius, then instantly raised to 40 degrees Celsius, while the humidity rapidly increases from 30% to over 95%, and maintained constant over different time periods, creating a typical temperature and humidity abrupt change scenario. Throughout this process, the acoustic sensors continuously collect sound wave output signals, and the environmental recording unit simultaneously records the current temperature and humidity parameter change curves. Each round of experiment lasts no less than 30 minutes, and the experimental cycle is no less than 72 hours, to ensure the capture of abnormal audio responses under various disturbance conditions.

[0022] All acquired audio data were uniformly numbered and serialized to create a multi-field data form including sound signal, sampling time, temperature change rate, humidity change amplitude, response amplitude, spectral image, and reference baseline. Abnormal segments in all sound signals were screened using manual review, with particular attention paid to phenomena such as concentrated spectral enhancement, abrupt changes in waveform periodicity, and sudden increases in background noise before and after abrupt changes in temperature and humidity rates. During the screening process, the inducing environmental conditions for each group of abnormal signals were clearly marked, including the time of the abrupt change, the direction of temperature and humidity change, the rate of change, and the duration. Non-interference segments were removed, retaining only signals that might exhibit self-oscillation characteristics for subsequent analysis. By comparing the signal similarity between different experimental periods and different sensors, individual signals with strong randomness, poor repeatability, and inability to be stably reproduced were removed. Finally, the screening results were compiled into a set of self-oscillation samples with clear environmental causes, stable reproducibility, and clear signal characteristics, serving as a preliminary candidate set of interference samples.

[0023] Based on the selected set of interfering acoustic waves, targeted classification and summarization were carried out. First, judgment criteria were established based on five dimensions: dominant frequency distribution range, amplitude fluctuation mode, harmonic component structure, initial abrupt change morphology, and spectral energy distribution density. Each interfering signal must possess stable characteristics in these five aspects to be classified into the same interference category. Then, each category of signal was matched one-to-one with its triggering conditions, extracting specific physical quantities such as triggering time point, temperature and humidity change rate, interference start and end times, total interference duration, maximum peak frequency, and corresponding noise energy density, forming a complete pairing data chain between signal and trigger. Each data chain contains at least ten sets of highly similar signal instances, and their consistency across different sensors and experimental batches was verified through comparative analysis. Throughout the process, mixed signals containing multiple uncertain disturbance sources were excluded to ensure that each category of interfering acoustic wave data has a single, clear physical trigger and traceable characteristic triggering conditions.

[0024] Based on the completed classification and summarization, a complete audio fingerprint feature database was constructed according to a standard structure. This database is presented in entry form, with each entry containing the original acoustic signal sequence, a three-dimensional spectrogram, a response waveform, a temperature and humidity change curve, a summary of the interference cause, the time of interference occurrence, the main interference frequency, the peak duration, the energy density distribution, and the offset value relative to the baseline response. To improve data comparison efficiency, the feature database is classified into three main categories: interference type, inducing environment, and response frequency band. Data under each index is accompanied by a manual verification label, indicating the number of times it was verified in different experimental rounds. All database content is stored using the same sampling rate and resolution format as the sensor sampling structure in actual forest deployments, requiring no format conversion for subsequent field comparisons.

[0025] S200 synchronously distributes the audio fingerprint feature library to the edge computing nodes, extracts on-site acoustic data through a dual-channel audio synchronous sampling method of the reference microphone channel and the reference cavity channel, and calculates the center frequency of narrowband interference within a preset frequency band to drive the initialization of the next tracking algorithm. Based on the completion of the acoustic fingerprint feature database, in order to achieve interference identification and initialization processing in the forest deployment environment, it is necessary to transmit the fingerprint data to the front-end device and extract the on-site acoustic signal using dual-channel synchronous sampling technology to accurately calculate the current interference center frequency. This process includes the following specific steps: All categorized interference acoustic signals in the acoustic fingerprint feature database are structured according to a predetermined format to ensure they can be directly read and accessed by front-end devices. To accommodate the limited storage and computing power of edge-deployed devices, each interference acoustic data is compressed into a minimum data unit consisting of an acoustic segment, spectrum analysis results, induced environmental parameters, and marker features. The acoustic segment is stored as a 16-bit integer at a fixed sampling rate; the spectrum analysis results record energy change trends in segmented envelope form; the induced environmental parameters, including temperature change rate, humidity jump amplitude, and trigger time, are stored as numerical pairs; and the marker features are the peak energy frequency and duration of each interference event in the frequency domain. All data units are pushed in batches by the central device to edge devices deployed at various observation points in the forest area during off-peak nighttime hours via wireless communication links, such as industrial-grade low-power wide-area communication protocols. A digital signature mechanism is included during the push process to ensure the consistency of the received data content during comparison. To ensure the integrity of the receiving end, the receiving device performs a line-by-line verification after each round of data import. The verification includes the length of the acoustic data, the position of the spectral peak, and the integrity of environmental parameters. If a deviation is detected, a data retransmission request is automatically initiated until the data is restored.

[0026] At the edge deployment site, to ensure the quality of the acquired audio signals, a reference audio acquisition channel and a reference acoustic channel were established, achieving synchronous acquisition through physical separation. The reference audio acquisition channel uses an omnidirectional electret condenser sensor, fixed at a tree trunk position 1.5 meters above the ground, to acquire acoustic signals from a real forest area. The reference acoustic channel is enclosed in a metal shielded housing filled with acoustically treated porous sound-absorbing material, creating a quasi-ideal sound field unaffected by external environmental disturbances. Both acquisition channels are synchronously sampled via a synchronous trigger circuit driven by the same clock source, with a sampling rate set at 96,000 points per second and a quantization depth of 24 bits to ensure the capture of subtle high-frequency interference fluctuations. To avoid the impact of terrain undulations or vegetation density on the signal transmission path, each channel is connected to a data buffer device via cables of the same length and shielding level, ensuring physical consistency of the signal path. During acquisition, each set of sampled data includes time-series waveforms from both the reference and reference channels, with their time base, length, and data structure remaining completely consistent for subsequent difference enhancement analysis.

[0027] Based on the acquired dual-channel audio data, frame-by-frame difference analysis is performed. By subtracting the differences between the reference channel and the baseline channel signals at the same time point by point, the significance of the interference signal is enhanced. After obtaining the difference waveform, the waveform is imported into the frequency distribution analysis process. The energy distribution of the signal in the range of 0 to 48 kHz is obtained through frequency domain reconstruction. Subsequently, standard samples of the interference type corresponding to the current time period are read from the audio fingerprint feature library previously stored in the device. The peak frequency, main bandwidth, peak energy density, and energy distribution trend in the current spectrum are compared item by item to confirm whether there are characteristics of self-excited interference in this time period. If the current difference spectrum and a certain interference sample reach a preset threshold similarity in all comparison dimensions, the frequency position corresponding to the sample is taken as the center frequency of the current interference event, with the unit accuracy controlled within 1 Hz. This center frequency value will serve as the starting frequency reference for the subsequent tracking stage, and the current identification time point and on-site temperature and humidity data will be marked simultaneously to provide initial parameters for modeling the interference evolution trend.

[0028] Based on the obtained center frequency, this frequency value is pushed to the front-end sampling controller to adjust the frequency focusing parameters of subsequent sampling windows. Specifically, a high-resolution sampling sub-band is set within the original sampling passband, extending 100 Hz upwards and downwards from the center frequency. This sub-band is used for detailed analysis of micro-frequency drift and interference jump phenomena in the next stage. Simultaneously, the sampling controller extracts and temporarily stores the interference-enhanced portions of the reference channel waveform and the baseline channel waveform, combining this with the current frequency information to generate a local interference snapshot. This snapshot contains a continuous data segment of at least 10 milliseconds before and after the interference occurs, used to initialize phase matching conditions in the next stage of phase-locked tracking. Through this processing flow, not only is real-time matching and interference identification of audio fingerprint data achieved in the natural forest environment, but the dual-channel structure also improves the interference detection rate and successfully provides key input parameters for subsequent frequency tracking, phase analysis, and source strength modeling.

[0029] S300 performs time-varying phase-locked tracking based on the extracted center frequency, calculates the phase noise spectral density function of the target frequency band, and inversely deduces the continuous functional relationship of the intensity change of the self-excited oscillation source over time based on the phase-locked output results, generating time-series data curves for dynamic feature analysis. To obtain the complete dynamic characteristics of the energy variation of self-excited disturbance sources in the forest environment over time, a time-domain phase-locked tracking process needs to be constructed based on the disturbance center frequency extracted in the previous step, and the source intensity evolution curve is inferred from the phase perturbation. The entire process consists of the following steps: The interference center frequency identified in the previous stage is used as the starting frequency input parameter for the current frequency band analysis, and the frequency focusing strategy of the sampling process is reset. The sampling controller limits the sampling bandwidth to a range of 100 Hz above and below this center frequency and increases the sampling rate to 192,000 points per second to ensure sufficient time domain resolution. During the sampling process, a sampling window of 10 milliseconds is used to acquire the signal within the target frequency band by performing bandpass sampling on both the reference acoustic channel and the standard acoustic channel. Within the entire sampling interval, a segment of high-fidelity waveform data is acquired for each time window for subsequent phase change extraction.

[0030] After acquiring the waveform data of the target frequency band, a precise periodic expansion operation is performed on each sampled waveform segment, and the start point, peaks, troughs, and phase crossover points of each complete cycle are extracted. Using the center frequency as a reference, the actual phase shift relative to the ideal sinusoidal reference model is measured within each cycle. Each 10-millisecond waveform segment is divided into multiple complete cycles, and a phase difference value is calculated for each cycle, with its time point marked at millisecond resolution. This process simultaneously records the instantaneous trend of phase fluctuations and the segment-by-segment phase change rate to obtain potential drift characteristics or sudden disturbances in the interference signal. Unlike existing processing methods that focus on amplitude intensity, this approach emphasizes extracting the stability and continuity of the phase disturbance itself, capturing the inherent fluctuation state of the interference signal during stable transmission.

[0031] After generating a high-temporal-resolution phase change sequence, a mapping relationship between this sequence and the energy output of the interference source is constructed. The phase change rate within each time interval is numerically integrated, and combined with the spectral energy density change in the current frequency band, the relative energy output level of the interference source within that time interval is calculated. To reduce the interference impact of environmental background fluctuations, the waveform energy processed by dual-channel differential processing is used as a correction factor in this stage to uniformly calibrate the calculated intensity values. Finally, a time evolution curve of the self-excited interference source intensity is generated, with the horizontal axis representing the sampling time point and the vertical axis representing the intensity value, in units of standardized relative ratios. The rising segment of the curve reflects the excitation process of the interference source, the plateau segment represents its stable output state, and the falling segment may correspond to the stage of self-excited source instability or suppression by external disturbances. The entire curve presents the complete interference energy lifecycle with millisecond-level accuracy.

[0032] The generated interference intensity evolution curve is used to guide the target signal extraction and noise cancellation strategy in the next stage. Specifically, when the curve is rising or plateauing, the corresponding time period is marked as the interference-dominant stage and is not used for insect sound feature extraction; when the curve enters a significant downward stage, this segment is used as a candidate window for insect sound signal exposure and is prioritized for target signal extraction. In addition, the sampling controller actively adjusts the sampling bandwidth and passband centroid for the next stage based on the peak position of the intensity curve, achieving dynamic frequency locking and recognition window optimization.

[0033] S400, based on the obtained self-excited source intensity time-series curve, applies a sparse time-frequency analysis algorithm to decompose the target signal, and extracts pulse-type acoustic envelope information related to the biting behavior of pest larvae by identifying and suppressing the interference frequency band caused by mechanical structure resonance. To accurately extract the biting sounds of insect larvae in complex forest acoustic backgrounds, it is necessary to combine the generated self-excited interference intensity variation curve, utilize the time-frequency sparsity characteristics, and gradually segment and suppress the interference frequency band of the target signal to finally extract the insect sound envelope with pulse characteristics. The entire processing flow includes the following steps: Based on the obtained self-excited source intensity time-series curve, the entire audio sampling data is divided into multiple independent segments according to time, each segment corresponding to an interference energy level. High-intensity interference periods are identified based on the rising and high plateau regions in the curve, and these periods are designated as the key areas for analysis due to significant resonance interference. The original sampling waveform is extracted for each time segment and subjected to a short-time Fourier transform to generate a spectrum covering a frequency range of 0 to 20000 Hz. By comparing the energy distribution in the spectrum of each time segment, interference frequency bands with highly concentrated peaks, fixed frequency positions, and continuous temporal existence are identified. These frequency bands typically appear between 800 Hz and 2500 Hz, exhibiting sharp energy focusing and narrow bandwidth characteristics. They maintain stable frequency positions and similar amplitude levels across multiple time segments and are identified as resonance signals generated by the sensor mounting structure or the mechanical response of the tree.

[0034] For identified interference frequency bands, a bandwidth-oriented reduction operation is performed on the corresponding original waveform. An adjustable bandwidth attenuation filtering technique is employed, extending the interference frequency center point by 50 Hz above and below it, setting this as the processing window. In each frame, the amplitude energy is gradually reduced point by point, controlling the attenuation depth to be 80% to 90% of the original amplitude. In frequency fluctuation regions, the bandwidth boundary is dynamically adjusted through cross-frame analysis to accommodate the slow drift of the interference frequency over time. After each frequency band processing, the energy distribution map of the entire waveform is recalculated and compared with the original spectrum to verify whether the interference peak has been effectively suppressed. If residual peaks are found, the filter order is increased or the window width is adjusted in the next round of processing until the interference energy drops below a set threshold. This method can minimize the interference of resonance interference on the target signal identification process without destroying the content of non-interference frequency bands.

[0035] After completing interference suppression operations across all frequency bands, the processed signal is subjected to insect sound pulse feature identification and extraction. First, each processed waveform is divided into consecutive time windows, each 2 milliseconds in length. Envelope extraction is performed on each window, and its instantaneous peak amplitude, waveform steepness, rise and fall rate, peak symmetry, and spacing are analyzed. A window is considered a candidate insect sound pulse signal if its waveform features meet the following conditions: its maximum amplitude is more than six times the average of its five neighboring windows; its rise time does not exceed 0.3 milliseconds; its waveform asymmetry is greater than 0.6; and its dominant frequency is concentrated between 500 Hz and 1500 Hz. Windows that meet all the above conditions are selected into the insect sound candidate set and reassembled in chronological order to construct a complete insect sound pulse envelope segment.

[0036] The extracted insect sound pulse envelopes were fully aligned along the entire time axis and structurally verified. Specific methods included: detecting the periodic repetition of non-biological signals between envelope segments; analyzing whether frequency slope changes were consistent with biological activity characteristics; and checking the similarity between the energy density of each envelope segment and known standard samples of insect sounds. After comparison and confirmation, segments inconsistent with insect sound characteristics were removed, and envelope boundary information was added to the remaining segments. The start time, end time, number of pulses, maximum amplitude, dominant frequency position, and envelope duration of each insect sound segment were calculated. Finally, a set of data segments for insect pest behavior identification and feedback decision-making was generated. These segments possessed clear signal characteristics, sufficient interference suppression, and high-confidence structural pulse envelopes, which can be directly used for subsequent modeling, prediction, or identification verification tasks, significantly improving the identification accuracy and environmental adaptability of insect sound monitoring.

[0037] S500 inputs the extracted insect sound envelope data results to the edge sensor array control module, combines real-time ambient temperature and humidity parameters, performs predictive model inference, and generates a set of adjustment parameters containing fine-tuning strategies to indicate the generation of edge control behavior; After extracting the insect sound pulse envelope, to achieve dynamic response adjustment of the front-end recognition device, it is necessary to combine real-time acquired forest area temperature and humidity parameters to deduce the insect sound evolution trend and generate a set of control parameters, including sensitivity, threshold, and response delay, as the input basis for control behavior execution. This processing flow includes the following specific steps: The extracted insect sound pulse envelope data is incorporated into the parameter generation stage. All insect sound data is organized in a structured manner, with each insect sound signal containing information such as time start, end, dominant frequency range, maximum amplitude, number of pulses, average pulse interval, and pulse duration. These fields are sorted along a time axis and synchronously bound to the corresponding temperature and humidity data within the current time period. Temperature is obtained by a resistance temperature detector (RTD) sampled every minute, and humidity is measured by a capacitive sensor and updated on average every minute; both data are accompanied by complete timestamps. During the data alignment stage, all insect sound signal fields are mapped to the corresponding temperature and humidity recording segments, forming a joint data unit with insect sound characteristics and environmental conditions. If the duration of the insect sound envelope is greater than 15 seconds, and the dominant frequency is concentrated between 700 and 1200 Hz, while the ambient temperature remains between 28 and 32 degrees Celsius and the humidity is higher than 85%, then the data unit is marked as a high-risk trigger source and proceeds to the next processing step.

[0038] For high-risk triggering units, evolutionary trend analysis is performed to determine whether the current insect sound behavior has entered an active state. The analysis includes whether the number of insect sound pulses has increased significantly, whether the pulse interval has shortened to less than 0.8 seconds, whether the envelope duration has continuously exceeded 30 seconds, and whether the dominant frequency range has become concentrated. If all the above indicators are met, combined with the high temperature and humidity environmental conditions, it is determined that the pest is in an active rising phase, and the control parameters of the front-end equipment need to be adjusted immediately. Based on this judgment, three sets of adjustment parameters are generated: the first set is the sampling sensitivity improvement range, used to improve the ability to distinguish weak insect sound signals, specifically increasing the sampling bit depth from 16 bits to 24 bits; the second set is the trigger threshold correction range, lowering the amplitude threshold for recognition and judgment by 15% to capture earlier insect sounds; the third set is the response delay adjustment time, shortening the response delay from the default 3 seconds to less than 1 second from the recognition behavior to the feedback action, thereby improving the response speed.

[0039] After generating the adjustment parameters, a backtracking mechanism for the effectiveness of these parameters is established based on real-time updated environmental data to ensure that the control strategy remains dynamically adaptive. Before each round of adjustment, the level of insect activity over the past two minutes is statistically analyzed. If the insect activity has decreased below a critical level and the temperature and humidity conditions are stabilizing, the application of the adjustment parameters is automatically terminated to prevent over-adjustment from wasting system resources or reducing recognition performance. If the insect activity continues to rise over several consecutive cycles and the environmental temperature and humidity fluctuations are within a preset threshold, the current parameter combination is maintained and recorded as the recommended configuration for this scenario for rapid loading in subsequent similar situations.

[0040] The finalized adjustment parameters are encapsulated in a control command format and distributed to each front-end execution channel, including the signal acquisition unit, recognition unit, and behavior execution structure. At the signal acquisition end, the sampling rate is increased to 1.5 times the original value, and the bandwidth is narrowed to 200 Hz above and below the main frequency concentration area, improving the focusing ability of insect sounds. In the recognition unit, new judgment thresholds are set and insect sound recognition conditions are redefined, covering three indicators: frequency distribution, waveform sharpness, and envelope stability. In the behavior execution structure, active control actions related to insect sound recognition are activated, such as starting external control equipment, illuminating warning lights, or sending remote alert signals. This process, driven by data and following behavior, achieves dynamic linkage throughout the entire process from recognition to control, and has the ability to adjust in real time under different environmental conditions.

[0041] S600 performs dynamic control operations on the edge sensor system based on the set of adjustment parameters. By injecting a random phase disturbance signal with limited amplitude into the system and superimposing a step change in temperature and humidity with a time step of nanosecond, it guides the frequency shift of the self-excited resonance response, causing it to lose its lock-in stability. It also uses an adaptive filter structure of the inverted signal to perform real-time removal processing on the residual narrowband interference signal, forming a continuous closed-loop identification process of self-modeling, self-sampling, self-tracking, self-decomposition, self-prediction and self-regulation. To break the frequency lock-in state of self-excited interference sources and eliminate the impact of interference signals on insect sound recognition in real time, it is necessary to perform four stages at the front end of the recognition device based on the generated adjustment parameters: disturbance injection, environmental abrupt change, frequency induction, and filtering removal. This achieves a closed-loop operation from recognition to control. The process specifically includes the following steps: Upon receiving the set of adjustment parameters derived from the fusion of insect sound characteristics and environmental conditions, phase perturbation injection is immediately performed. This operation is accomplished by applying a set of artificially synthesized weak AC signals to the sensor drive circuit. The frequency of the signals is strictly consistent with the identified interference center frequency, while the phase is driven by a stable noise source and randomly fine-tuned every ten milliseconds. The adjustment range is set within ±7 degrees, and the amplitude is controlled below 15% of the baseline waveform amplitude to ensure that it does not interfere with normal insect sound sampling, but introduces a slight perturbation into the resonant path, causing the sensing element to be in a slightly phase unstable state. During the entire signal injection period, signal modulation is completed by a dedicated drive structure, with an injection time not exceeding 100 milliseconds, to excite the imbalance point of the resonant state.

[0042] Following the injection of the disturbance signal, abrupt changes in environmental physical conditions further exacerbate the instability tendency of the resonant path. This process simultaneously activates two control devices: a micro-heater that raises the ambient air temperature from room temperature to over 35 degrees Celsius within 5 milliseconds; and an electronically controlled humidity shocker that uses a rapid spray device to release supersaturated water vapor into the target area, increasing the relative humidity to over 95%. The step effect of these two physical variables alters the tension, capacitance distribution, and vibration mode of the sensing diaphragm surface, causing a shift in its originally locked frequency. The rise in temperature and humidity is controlled by nanosecond step increments, with simultaneous jumps to their maximum values, ensuring consistency in timing and energy of the excitation effect. During this process, the dominant frequency of the sensor's output signal rapidly shifts upwards or downwards from its original fixed value, exhibiting spectral jumps or energy peak splitting characteristics, indicating that the resonant locking state is disrupted.

[0043] When a frequency shift or energy peak change reaches a set amplitude, the anti-phase interference removal process is immediately activated. This step relies on a previously saved original interference waveform template. The segment in the newly acquired signal that most closely resembles the template waveform is selected, and its inverse signal is synthesized. Based on the original waveform, the inverse signal undergoes amplitude preservation, time alignment, and a 180-degree phase reversal to construct a suppression signal with complete cancellation characteristics. In real-time signal processing, whenever residual interference is detected in the target frequency band, the corresponding inverse signal is immediately loaded and superimposed onto the current signal stream at the corresponding sampling time point. This inverse processing not only weakens the energy of the main peak frequency but also eliminates its subharmonic response, effectively smoothing the spectral curve and reducing periodic interference induced by resonance.

[0044] After the interference frequency jump and residual signal filtering are completed, the sampling window configuration for the next cycle is reloaded, and combined with the current insect sound recognition accuracy evaluation index, a complete closed-loop recognition process update is completed. Before the new round of sampling, the disturbance response state, temperature and humidity jump amplitude, interference elimination degree, and recognition judgment confidence recorded in the previous cycle are packaged and stored as behavioral feedback records for subsequent adaptive recognition strategy calls. The entire closed-loop process runs through six stages: modeling, sampling, tracking, decomposition, prediction, and control. Among them, disturbances and environmental excitations constitute the active path to break the resonance-locked structure, while inverse filtering and feedback records form a self-correction channel for interference control results. This strategy breaks away from the traditional dependence on fixed filtering parameters and static recognition models, and achieves accurate interruption and active removal of structural noise sources in unstable forest environments, ensuring that insect sound recognition is always in a highly reliable and dynamically adaptive state.

[0045] This invention constructs a combined model of temperature, humidity, and electroacoustic response and generates an acoustic fingerprint database, enabling the system to pre-identify self-excitation risks. Furthermore, it utilizes dual-channel synchronous sampling and phase-locked tracking technology at edge nodes to accurately locate the interference center frequency. By combining time-varying phase analysis and sparse time-frequency decomposition, the weak insect sound pulse envelope corresponding to larval biting behavior can be recovered from the interfered original signal, significantly improving the accuracy of feature extraction. Based on this, dynamic adjustment commands are generated through the fusion analysis of insect sounds and environmental parameters, and phase perturbation and temperature / humidity jump strategies are introduced to interrupt frequency hopping and remove residual interference sources, giving the entire acoustic recognition process closed-loop and adaptive capabilities. Compared with existing technologies, this method not only improves the recognition stability and anti-interference ability under complex climatic conditions in forest areas but also significantly reduces the false negative and false positive rates, providing efficient and reliable technical support for early warning and precise control of forest pests.

[0046] This invention provides, for example Figure 1 and Figure 2 The edge computing-based acoustic recognition system for forest borers, as shown, includes a modeling fingerprint generation module, an interference identification and frequency localization module, a phase-locked tracking and time-series modeling module, a time-frequency decomposition and insect sound extraction module, a predictive inference and instruction generation module, and a dynamic control and interference suppression module. The modeling fingerprint generation module establishes a joint model between temperature and humidity coupling parameters and electroacoustic response, and generates an acoustic fingerprint database of self-excited oscillation risk by collecting mutation samples under extreme environmental conditions. The interference identification and frequency localization module distributes the audio fingerprint database to the edge computing nodes, extracts the on-site acoustic data through a dual-channel synchronous sampling method of the reference microphone channel and the reference cavity channel, and calculates the center frequency of the narrowband interference. Phase-locked tracking and timing modeling module 0 performs time-varying phase-locked tracking based on the center frequency, calculates the phase noise spectral density, and reverse-engineers the timing curve of the self-excited oscillation source intensity changing with time; The time-frequency decomposition and insect sound extraction module performs sparse time-frequency decomposition based on time-series curves, suppresses interference frequency bands caused by resonance, and extracts pulse-type acoustic envelopes related to larval biting. The predictive reasoning and instruction generation module inputs the insect sound envelope into the edge sensor control module, combines it with real-time temperature and humidity parameters to perform predictive reasoning, and generates a set of adjustment parameters for control. The dynamic control and interference suppression module performs dynamic control based on the set of adjustment parameters. By injecting random phase perturbations and superimposing nanosecond-level temperature and humidity steps, it guides the self-excited resonant frequency jump, eliminates residual narrowband interference, and forms a closed-loop identification process of modeling, sampling, tracking, decomposition, prediction and control.

[0047] The edge computing-based acoustic recognition method for forest borers provided in this invention is implemented through the aforementioned edge computing-based acoustic recognition system for forest borers. For details on the specific methods and processes of the edge computing-based acoustic recognition system for forest borers, please refer to the embodiments of the edge computing-based acoustic recognition method for forest borers, which will not be repeated here.

[0048] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for audio recognition of forest trunk borers based on edge computing, characterized in that, Includes the following steps: S100 establishes a joint model between temperature and humidity coupling parameters and electroacoustic response, and generates an audio fingerprint database of self-excited oscillation risk by collecting mutation samples under extreme environmental conditions. S200 distributes the audio fingerprint database to edge computing nodes, extracts on-site acoustic data through dual-channel synchronous sampling of the reference microphone channel and the reference cavity channel, and calculates the center frequency of narrowband interference; S300 performs time-varying phase-locked tracking based on the center frequency, calculates the phase noise spectral density, and reverse-engineers the time-series curve of the intensity of the self-excited oscillation source changing with time. S400 performs sparse time-frequency decomposition based on time-series curves, suppresses interference frequency bands caused by resonance, and extracts pulsed acoustic envelopes related to larval biting. The S500 inputs the insect sound envelope into the edge sensor control module, combines it with real-time temperature and humidity parameters to perform predictive inference, and generates a set of adjustment parameters for control. The S600 uses dynamic control based on a set of adjustment parameters. By injecting random phase disturbances and superimposing nanosecond-level temperature and humidity steps, it guides the self-excited resonant frequency jump, eliminating residual narrowband interference.

2. The method for audio recognition of forest borers based on edge computing according to claim 1, characterized in that, Step S100 includes: An acoustic response testing device was built in a closed environment control chamber, and acoustic sampling operations were performed using temperature control components, humidity control components, heat transfer units and multiple sets of audio sensors. The temperature was adjusted from room temperature to low temperature and then raised to high temperature, while the humidity was adjusted from low humidity to high humidity. The above parameters were kept constant at different time periods, and the audio signal and environmental parameter curves under the sudden change environmental conditions were collected. The collected audio data is subjected to anomaly screening, interference classification and signal comparison operations. Invalid samples that are not repeatable or have no clear cause are removed, and interference signals with stable characteristics and unique causes are extracted. Construct audio fingerprint feature entries and complete the construction of the audio fingerprint feature database.

3. The method for audio recognition of forest borers based on edge computing according to claim 1, characterized in that, Step S200 includes: The audio fingerprint feature database contains multiple data units, including interference signals, spectral features, environmental factors, and frequency markers, which are compressed and stored in a structured manner and then synchronously transmitted to the front-end sampling equipment deployed in the forest area via wireless communication. A reference acoustic acquisition channel and a reference acoustic acquisition channel are established in the edge sampling device, and dual-channel synchronous sampling is achieved through a co-source clock triggering mechanism to acquire real acoustic signals from the forest area and non-interference reference acoustic signals respectively; Perform dual-channel difference analysis and spectrum calculation, extract the interference enhancement waveform and compare it with the pre-stored fingerprint sample, and output the current interference center frequency; Configure the frequency focusing parameters of the subsequent sampling window based on the center frequency, and generate local interference snapshots to provide the initial data basis for the frequency tracking stage.

4. The method for audio recognition of forest borers based on edge computing according to claim 1, characterized in that, Step S300 includes: The identified interference center frequency is used as the starting frequency input parameter for the current frequency band analysis. The sampling bandwidth is set within the range above and below the center frequency and the sampling rate is increased. Bandpass sampling is performed on the reference acoustic channel and the reference acoustic channel according to a fixed time window to obtain the target frequency band signal. Each sampled waveform is periodically expanded and the period start point, peak, trough and phase intersection point are extracted. The phase shift is measured with the center frequency as a reference and the time point is marked. The phase fluctuation trend and phase change rate are recorded. After obtaining the phase change sequence, a mapping relationship between the phase change and the energy output of the interference source is established. The waveform energy after dual-channel differential processing is used as a correction factor to calculate the intensity of the interference source and generate a time evolution curve. The time evolution curve of interference intensity is used to guide the target signal extraction and noise cancellation in the next stage, and the sampling bandwidth and passband centroid are adjusted accordingly to achieve dynamic frequency locking and identification window optimization.

5. The method for audio recognition of forest borers based on edge computing according to claim 4, characterized in that, Phase shift measurement uses an ideal sine wave as a reference model, calculates the phase difference for each cycle, and records the phase change with millisecond-level resolution to determine the stability and suddenness characteristics of interference signals.

6. The method for audio recognition of forest borers based on edge computing according to claim 1, characterized in that, Step S400 includes: The audio sampling data segments are divided according to the self-excited source intensity time-series curve, and high-intensity interference periods are identified. Short-time Fourier transform is performed on each segment to extract the interference frequency band with fixed frequency position and high energy density. For the identified interference frequency bands, a bandwidth attenuation filtering method is used to reduce the frequency band, dynamically adjust the filtering window, and control the interference energy to decrease to below the set threshold. Continuous time window analysis is performed on the waveform after interference suppression. Based on amplitude, rise time, waveform asymmetry and dominant frequency range, candidate signals of insect sound pulses are identified and the insect sound envelope segment is reconstructed. The frequency slope change and energy density characteristics of the insect sound envelope are structurally verified, and segments that do not conform to the insect sound characteristics are removed to generate insect sound recognition data with confidence for subsequent processing.

7. The method for audio recognition of forest trunk borers based on edge computing according to claim 1, characterized in that, Step S500 includes: The insect sound envelope data is synchronized with real-time temperature and humidity data on a time axis to construct a joint data unit of insect sound characteristics and environmental conditions; Risk levels are determined based on the duration of insect sounds, the range of dominant frequencies, the number of pulses, and the ambient temperature and humidity levels. The evolution trend of insect sounds is analyzed to identify the active state of insect pests. Based on the active state, three types of adjustment parameters are generated: sampling sensitivity, identification threshold, and response delay, to adjust the signal acquisition and feedback response strategy. The adjustment parameters are dynamically updated based on the changing trends of insect sounds, and the confirmed adjustment parameters are distributed to each identification and execution channel in the form of instructions to complete the control behavior.

8. The method for audio recognition of forest borers based on edge computing according to claim 7, characterized in that, The sampling sensitivity parameter in the parameter set is achieved by increasing the sampling bit depth and compressing the sampling frequency band; the identification threshold parameter is achieved by lowering the insect sound amplitude threshold; and the response delay parameter is achieved by shortening the time interval between identification and execution.

9. The method for audio recognition of forest trunk borers based on edge computing according to claim 1, characterized in that, Step S600 includes: Injecting a perturbation signal with consistent frequency and controlled phase causes the sensing element to enter a slightly unstable state. Simultaneously applying a temperature rise and a humidity jump alters the resonance characteristics of the acoustic sensing components, inducing a frequency jump and breaking frequency lock. Based on the frequency offset, the inverted signal structure is invoked to perform interference removal processing such as amplitude preservation, time alignment, and phase reversal in real time; Record the disturbance response state, temperature and humidity jump amplitude, interference removal degree and identification confidence level to complete a closed-loop operation of sampling and identification strategy update.

10. A forest trunk borer acoustic recognition system based on edge computing, used to implement the forest trunk borer acoustic recognition method based on edge computing as described in any one of claims 1-9, characterized in that, It includes a modeling fingerprint generation module, an interference identification and frequency localization module, a phase-locked tracking and timing modeling module, a time-frequency decomposition and insect sound extraction module, a predictive inference and instruction generation module, and a dynamic control and interference suppression module. The modeling fingerprint generation module establishes a joint model between temperature and humidity coupling parameters and electroacoustic response, and generates an acoustic fingerprint database of self-excited oscillation risk by collecting mutation samples under extreme environmental conditions. The interference identification and frequency localization module distributes the audio fingerprint database to the edge computing nodes, extracts the on-site acoustic data through a dual-channel synchronous sampling method of the reference microphone channel and the reference cavity channel, and calculates the center frequency of the narrowband interference. Phase-locked tracking and timing modeling module 0 performs time-varying phase-locked tracking based on the center frequency, calculates the phase noise spectral density, and reverse-engineers the timing curve of the self-excited oscillation source intensity changing with time; The time-frequency decomposition and insect sound extraction module performs sparse time-frequency decomposition based on time-series curves, suppresses interference frequency bands caused by resonance, and extracts pulse-type acoustic envelopes related to larval biting. The predictive reasoning and instruction generation module inputs the insect sound envelope into the edge sensor control module, combines it with real-time temperature and humidity parameters to perform predictive reasoning, and generates a set of adjustment parameters for control. The dynamic control and interference suppression module performs dynamic control based on the set of adjustment parameters. By injecting random phase disturbances and superimposing nanosecond-level temperature and humidity steps, it guides the self-excited resonant frequency jump and eliminates residual narrowband interference.

Images