Weathered rock interface identification and excavation parameter adaptive control method based on vibration spectrum analysis

Abstract

This invention relates to a method for identifying weathered rock strata interfaces and adaptively controlling excavation parameters based on vibration spectrum analysis, belonging to the field of intelligent excavator construction technology. It addresses the problems of relying on manual experience for soil-rock interface identification during weathered rock strata trench excavation, low efficiency of mechanical switching, and slope vibration risks. The method collects vibration acceleration signals using a triaxial accelerometer, extracts frequency domain feature vectors including the dominant frequency amplitude, high-frequency band energy proportion, spectral centroid, and kurtosis value, and inputs these vectors into a multi-class support vector machine model to output category labels. When the category label changes from "cohesive silt" to "strongly weathered gneiss," the input current to the proportional solenoid valve of the excavator's hydraulic main pump is reduced, and a breaker hammer warning signal is activated. When the category label is "moderately weathered gneiss with interbedded boulders" and the high-frequency band energy proportion and kurtosis value exceed the safe range, the pilot oil pressure of the breaker hammer is limited. This method enables real-time identification of the soil-rock interface and adaptive control of excavation parameters.

Description

Technical Field

[0001] This invention belongs to the field of intelligent construction technology of excavators, specifically involving a method for weathered rock layer interface identification and adaptive control of excavation parameters based on vibration spectrum analysis. Background Technology

[0002] Excavation of foundation trenches in weathered rock strata is a common procedure in geotechnical engineering construction. When the geological conditions of the excavation area are characterized by a soft upper layer and a hard lower layer (i.e., cohesive silt in the upper layer and weathered gneiss in the lower layer), construction faces the challenge of real-time identification of the rock-soil interface. In existing technologies, excavator operators mainly rely on experience, such as listening and vibration sensing, to judge the current working medium, which suffers from strong subjectivity and delayed response. When manual experience is insufficient for accurate judgment, over-excavation or under-excavation often occurs: over-excavation disturbs the undisturbed soil at the bottom of the trench, requiring additional backfilling; under-excavation necessitates secondary excavation, both increasing construction costs and time. When weathered rock strata contain boulders, the collision between the bucket teeth and the boulders generates a strong impact, accelerating not only the wear of the bucket teeth and fatigue of the mechanical structure, but also the high-frequency impact energy propagating through the rock mass, potentially inducing the expansion of weathering fissures and increasing the risk of slope instability.

[0003] To replace manual judgment, existing research has proposed technical solutions for identifying soil and rock categories based on vibration signals. For example, Chinese patent document CN113075120A discloses a real-time soil category identification method, which collects vibration signals by installing sensors on the excavation excavation mechanism, extracts time-domain and frequency-domain features, and compares them with a calibration database to determine the soil category. US patent application US20230152279A1 discloses a material identification technology based on vibration signals, which extracts features from vibration signals and inputs them into a machine learning model to predict the ground material type. However, the frequency-domain features used for identification in the above solutions refer to parameters such as amplitude and frequency range in general, and are not specifically designed for weathered rock excavation scenarios. In weathered rock excavation, both uniformly weathered rock layers and weathered rock layers containing boulders exhibit increased vibration amplitude and high-frequency components, but there is a fundamental difference in the suddenness of impact: in the case of boulders, the bucket teeth collide periodically with the boulders, and the vibration signal exhibits the characteristics of "stable background superimposed with instantaneous spikes". The characteristic parameters used in the existing solutions mentioned above cannot effectively characterize the suddenness of this impact, making it impossible to distinguish between "uniformly weathered rock layers" and "rock-bearing weathered rock layers," and the subsequent control strategies for these two working conditions are completely different.

[0004] In terms of hydraulic breaker control, existing technologies mainly focus on adjusting the impact frequency based on material stiffness or chisel rebound time to improve crushing efficiency. For example, patent document CN119897207A discloses an intelligent electro-hydraulic hydraulic breaker adaptive control system that analyzes material stiffness by emitting sound waves and receiving reflected signals, and adjusts the impact force and frequency accordingly. Patent document CN104532897A discloses an adaptive intelligent hydraulic breaker that uses a grating sensor to detect the chisel rebound time to determine the hardness of the impacted object, and then adjusts the frequency accordingly. However, the control of the above schemes relies on the detection of a single physical quantity (such as sound wave reflection time or chisel rebound time) by their own sensors, and cannot obtain information about the rock and soil type of the current working medium. When encountering conditions containing boulders, the above schemes can only perceive "increased hardness" but cannot recognize the semantic information of "boulder impact." Their response is the same as when encountering uniform hard rock, lacking dedicated safety control logic for boulder impact.

[0005] In summary, existing vibration signal-based soil and rock identification methods fail to consider the critical characteristic of sudden impact during weathered rock excavation, resulting in an inability to effectively distinguish between homogeneous hard rock and rock-bearing conditions. Furthermore, the hydraulic breaker control schemes are not linked to the soil and rock category identification results, making it impossible to adopt differentiated control strategies based on the type of working condition. In the scenario of weathered rock foundation trench excavation, there is a lack of an integrated method capable of identifying different working conditions through specific frequency domain feature combinations and adaptively adjusting excavation and breaking parameters based on the identification results. Summary of the Invention

[0006] One objective of this invention is to address the following problem: During the excavation of foundation trenches in weathered rock strata, the cutting resistance of the bucket teeth increases significantly when the working medium transitions from cohesive silt to weathered gneiss. If the excavation thrust is not reduced in time, the bucket teeth will bear excessive loads in the hard rock, accelerating wear and even causing breakage. Simultaneously, the operator cannot immediately detect the need to switch to the hydraulic breaker upon contact with the hard rock, resulting in a reaction delay of several seconds. When the weathered rock strata contain boulders, the collision between the bucket teeth and the boulders generates a sudden and intense impact. If the hydraulic breaker operates at its normal frequency, the high-frequency impact energy propagating through the rock mass may induce the expansion of weathering fissures, increasing the risk of slope instability.

[0007] To achieve the above objectives, this invention provides a method for weathered rock strata interface identification and adaptive control of excavation parameters based on vibration spectrum analysis, comprising the following steps: Step 1: Fix a triaxial accelerometer on the surface of the stick or rocker arm structure of the excavator working device, and continuously collect vibration acceleration signals during the excavation operation at a preset sampling frequency; Step 2: Perform a fast Fourier transform on the vibration acceleration signal within the acquisition window to extract the frequency domain feature vector. The frequency domain feature vector includes the main frequency amplitude, the energy proportion of the high-frequency band, the spectral centroid, and the kurtosis value. Step 3: Input the frequency domain feature vector into the pre-trained multi-class support vector machine model and output the category label of the current working medium. The category label should be divided into at least "cohesive silt", "strongly weathered gneiss" and "moderately weathered gneiss with boulders". Step 4: Generate the first and second control commands based on the category labels: When the category label changes from "cohesive silt" to "strongly weathered gneiss", the first control command is used to reduce the input current of the proportional solenoid valve of the excavator's hydraulic main pump to reduce the stick digging thrust, while the second control command is used to activate the breaker hammer operation prompt signal in the cab. When the category label is "moderately weathered gneiss with interbedded boulders", if the high-frequency band energy ratio and kurtosis value exceed the preset safety range, a third control command is generated. The third control command is used to limit the pilot oil circuit pressure of the hydraulic breaker to reduce the impact frequency.

[0008] Preferably, vibration acceleration signals during the excavation operation are continuously acquired at a preset sampling frequency, specifically including: The output signal of the triaxial accelerometer is acquired at the first sampling frequency, and the output signal is amplified by a programmable gain amplifier located between the triaxial accelerometer and the analog-to-digital converter. The gain value of the programmable gain amplifier is determined by the category label of the current working medium. When the category label is "viscous silt", the first gain value is set, and when the category label is "strongly weathered gneiss" or "moderately weathered gneiss with boulders", the second gain value is set to be greater than the first gain value. The signal amplified by the programmable gain amplifier is monitored in real time. When the signal amplitude exceeds the preset impact threshold within N consecutive sampling points, the current acquisition mode is switched from continuous acquisition mode to impact event triggered acquisition mode. In impact event triggered acquisition mode, the signal is acquired at a second sampling frequency higher than the first sampling frequency and amplified with a second gain value until the signal amplitude falls back below the preset impact threshold and continues for M sampling points, then it is restored to the continuous acquisition mode. The value of the second sampling frequency is determined based on the kurtosis value calculated in the previous sampling window. The larger the kurtosis value, the higher the second sampling frequency.

[0009] Preferably, in the frequency domain feature vector: the dominant frequency amplitude is the amplitude corresponding to the point with the maximum amplitude in the spectrum after Fast Fourier Transform; the high-frequency band energy ratio is defined as: dividing the entire frequency band from 0 to half the sampling frequency into a low-frequency band and a high-frequency band, and taking the ratio of the sum of the power spectral values ​​corresponding to all frequency points in the high-frequency band to the sum of the power spectral values ​​corresponding to all frequency points in the entire frequency band; the spectral centroid is the first moment of the power spectrum, calculated by the following formula: Where C is the centroid of the spectrum, f is the frequency, and P(f) is the power spectral amplitude at frequency f. s is the sampling frequency; the kurtosis value is the fourth-order normalized statistical moment of the time-domain waveform of the vibration acceleration signal.

[0010] Preferably, the frequency domain feature vector is input into a pre-trained multi-class support vector machine model, specifically including: The multi-class support vector machine model is constructed using a one-to-one multi-class strategy based on a decision binary tree. Its structure includes three binary nodes: the first node uses the dominant frequency amplitude and the high-frequency band energy ratio as input features to determine whether the current working medium belongs to "silt" or "rock strata"; the second node uses the spectral centroid and the kurtosis value as input features to further subdivide the samples that the first node judges as "rock strata" into "strongly weathered gneiss" and "moderately weathered gneiss with boulders"; the third node uses the kurtosis value and the high-frequency band energy ratio as input features to perform a confirmatory secondary discrimination on the samples that the first node judges as "silt" to filter out samples that the first node misclassifies as "moderately weathered gneiss with boulders" as "silt". The training of the multi-class support vector machine model includes: optimizing the kernel function parameters and penalty factor of the support vector machine model with the goal of maximizing the weighted classification accuracy through cross-validation and grid search methods. The weight coefficient of each sample is set according to the construction risk level of its category label. The construction risk level is determined in the order of "moderately weathered gneiss with boulders" higher than "strongly weathered gneiss" higher than "cohesive silt".

[0011] Preferably, the category labels are generated using the following hierarchical generation mechanism: The multi-class support vector machine model outputs the posterior probability value corresponding to each class label. When the maximum posterior probability value is greater than or equal to the preset first confidence threshold, the class label corresponding to the maximum posterior probability value is directly output as the determined class label of the current working medium. When the maximum a posteriori probability value is less than the first confidence threshold but greater than or equal to the preset second confidence threshold, the category label corresponding to the maximum a posteriori probability value is output as the candidate category label of the current working medium, and a "transition state" flag is attached. The "transition state" flag triggers the first control instruction in step four to reduce the execution intensity by a preset scaling factor. When the maximum posterior probability value is less than the second confidence threshold, the category label output by the previous acquisition window remains unchanged, and a "low confidence" flag is added. The "low confidence" flag prohibits the control command update in step four. The first confidence threshold is determined based on the construction risk level of the category label corresponding to the maximum posterior probability value: the construction risk level is determined in the order of "moderately weathered gneiss with boulders" higher than "strongly weathered gneiss" higher than "cohesive silt". The higher the construction risk level, the lower the first confidence threshold.

[0012] Preferably, when the category label changes from "viscous silt" to "strongly weathered gneiss", the first control command controls the input current of the proportional solenoid valve to gradually decrease from the current operating current value to the preset target current value according to a preset first decay rate. The first decay rate is determined based on the maximum a posteriori probability value before the category label change: the higher the maximum a posteriori probability value, the slower the first decay rate, so as to prioritize the smoothness of the action when the identification confidence is high and prioritize the timeliness of the response when the identification confidence is low. The target current value is determined based on the preset percentage corresponding to the "strongly weathered gneiss" category label, with the preset percentage being 40% to 60% of the current operating current value; The second control command determines the flashing frequency or sound alert intensity of the warning signal in the cockpit based on the real-time change rate of the main frequency amplitude and the energy ratio of the high-frequency band. When the increase of the main frequency amplitude in K consecutive sampling points exceeds the preset increase threshold, or the real-time value of the energy ratio of the high-frequency band exceeds the preset energy threshold, the flashing frequency or sound alert intensity is increased to guide the operator to switch the breaker first. When the category label includes the "transitional state" indicator, the first control command reduces the attenuation of the input current of the proportional solenoid valve by a proportional factor, which ranges from 0.5 to 0.8.

[0013] Preferably, the preset safety range is jointly defined by the upper limit threshold of high frequency band energy proportion and the upper limit threshold of kurtosis value, wherein the upper limit threshold of high frequency band energy proportion and the upper limit threshold of kurtosis value are determined according to the geomechanical parameters of the slope of the construction area: the geomechanical parameters include the uniaxial compressive strength of weathered gneiss, the density of fracture development and the depth of groundwater level; When the high-frequency band energy ratio exceeds the upper limit threshold of the high-frequency band energy ratio and the kurtosis value exceeds the upper limit threshold of the kurtosis value, the third control command controls the input current of the electro-hydraulic proportional pressure reducing valve installed on the pilot oil circuit of the breaker to gradually decrease from the current working current value according to the preset second decay rate. The output pressure of the electro-hydraulic proportional pressure reducing valve decreases as the input current decreases, thereby continuously reducing the impact frequency of the breaker. The target value for reducing the input current of the electro-hydraulic proportional pressure reducing valve is determined according to the degree of kurtosis exceeding the limit: the greater the kurtosis exceeds the upper limit threshold, the lower the target value, and the lower the corresponding impact frequency of the hydraulic breaker. When the kurtosis value exceeds the preset low-frequency safety lockout threshold, the third control command controls the input current of the electro-hydraulic proportional pressure reducing valve to reduce to zero, so as to prevent the breaker from being activated. At the same time, the "low-frequency safety lockout" prompt signal in the cockpit is activated. The low-frequency safety lockout threshold is 1.5 to 2.0 times the upper limit of the kurtosis value.

[0014] Preferably, before extracting the frequency domain feature vector in step two, the following steps are also included: A rain sensor is installed on the top of the excavator cab or the upper slewing body, and a humidity sensor is installed on the mounting base surface of the triaxial accelerometer. When the rainfall intensity detected by the rain sensor exceeds a preset first intensity threshold, the vibration acceleration signal is subjected to noise reduction processing and high-frequency energy compensation processing in sequence. The noise reduction intensity parameter of the noise reduction processing is determined according to the rainfall intensity. The greater the rainfall intensity, the greater the noise reduction intensity, so as to suppress the impact noise generated by raindrops hitting the surface of the equipment. The high-frequency energy compensation processing applies a frequency-dependent compensation gain to the digital signal after analog-to-digital conversion that is higher than the preset boundary frequency. The compensation gain increases with the frequency, and the reference value of the compensation gain is determined according to the rainfall intensity level at which the rainfall intensity is located. The rainfall intensity is input into a pre-established rainfall intensity-feature correction mapping table. The extracted dominant frequency amplitude and high-frequency band energy proportion are multiplied by the corresponding correction coefficients to obtain the corrected dominant frequency amplitude and high-frequency band energy proportion. The rainfall intensity-feature correction mapping table is pre-established by conducting calibration excavation experiments on known soil and rock types under dry conditions and different rainfall intensities. After replacing the original values ​​with the corrected main frequency amplitude and the corrected high-frequency band energy ratio, they together with the spectral centroid and kurtosis value form the corrected frequency domain feature vector, which is then input into the multi-class support vector machine model. When the rainfall intensity detected by the rain sensor exceeds the preset second intensity threshold, the first confidence threshold used when generating the category label is multiplied by a preset scaling factor, with the scaling factor ranging from 0.7 to 0.9.

[0015] The present invention has at least the following beneficial effects: First, a triaxial accelerometer is installed on the excavator's boom or arm to collect vibration signals. The signals are then extracted using a Fast Fourier Transform (FFT) to form a four-dimensional frequency domain feature vector, comprising the dominant frequency amplitude, high-frequency energy percentage, spectral centroid, and kurtosis. This vector is input into a multi-class support vector machine (SVM) model to output soil and rock category labels. Based on these labels, control commands are generated for the hydraulic main pump current and the breaker's pilot pressure. This method integrates identification and control into a closed loop. When the category label changes from cohesive silt to strongly weathered gneiss, the excavation thrust is automatically reduced, and a switch to the breaker is prompted. When the condition is identified as containing boulders and the characteristics exceed a threshold, the breaker's impact frequency is automatically limited. This avoids the lag and subjectivity of manual judgment, reduces mechanical damage caused by continuous high-load operation of the bucket teeth in hard rock, and lowers the risk of slope instability induced by high-frequency breaker impact.

[0016] 2. The gain value of the programmable gain amplifier is determined by the soil and rock category label. Low gain is used for silty sand layers to avoid excessive noise amplification, while high gain is used for rock layers to prevent saturation distortion of the impact signal. Simultaneously, the amplifier switches between continuous acquisition mode and impact event-triggered acquisition mode based on signal amplitude monitoring results. In impact mode, the sampling frequency is increased and dynamically determined by the kurtosis value. This scheme reduces the data processing burden by operating at a low sampling frequency during stable operation, and captures high-frequency details at a high sampling frequency during impact events, balancing signal fidelity and system power consumption.

[0017] 3. The dominant frequency amplitude reflects the magnitude of rock-breaking resistance, the energy proportion of the high-frequency band and the spectral centroid describe the degree of vibration energy migration to higher frequencies, and the kurtosis value characterizes the suddenness of the impact. The combination of these four characteristic parameters can distinguish between three working conditions: silt layer, uniformly weathered rock layer, and weathered rock layer containing boulders. In particular, the sensitivity of the kurtosis value to the impact of boulders allows the boulder-containing working condition to be detected separately, providing a basis for subsequent differentiated control.

[0018] 4. The multi-class support vector machine adopts a decision binary tree structure, with each node using the feature combination that has the strongest discriminative ability for different classification tasks, avoiding feature redundancy and classification overlap. During the training phase, the weighted classification accuracy is the target, and higher weights are assigned to high-risk categories. This ensures that the model prioritizes the recognition recall of working conditions containing boulders during parameter optimization, reducing the possibility of missing high-risk working conditions.

[0019] 5. Based on the comparison between the posterior probability value and the confidence threshold, class labels are output in a stratified manner, with the addition of transition state or low confidence indicators. In the transition zone of the soil-rock interface, the transition state indicator reduces the execution intensity of control commands by a scaling factor, avoiding frequent jumps in class labels and control oscillations; the low confidence indicator prohibits control command updates, preventing malfunctions caused by transient interference.

[0020] 6. The first control command controls the proportional solenoid valve current to gradually decrease to the target value at a decay rate. The decay rate is related to the posterior probability value before the category label change; the decay is slower when the confidence level is high to ensure stability, and faster when the confidence level is low to ensure timeliness. The target current value is set to 40% to 60% of the current value, preserving the basic digging force to complete the current action while protecting the bucket teeth. The second control command adjusts the intensity of the prompt signal based on the increase in the main frequency amplitude and the proportion of high-frequency band energy, enabling the operator to perceive the rock hardness and rationally arrange the switching timing.

[0021] 7. The safe zone is jointly defined by the upper limit threshold of high-frequency band energy proportion and the upper limit threshold of kurtosis value, and the threshold is related to the slope geomechanical parameters, making the frequency limiting control compatible with the construction site conditions. The third control command controls the current of the electro-hydraulic proportional pressure reducing valve to continuously reduce the target current value, and the target current value is determined according to the kurtosis value exceeding the limit, realizing the matching of frequency limiting intensity and impact intensity. The low-frequency safety lock threshold prohibits the use of the hydraulic breaker when the kurtosis value is extremely high, thus preventing the occurrence of high-risk impact events from the hardware level.

[0022] 8. Under rainy construction conditions, the absorption and attenuation of high-frequency energy by water is compensated, and the corrected frequency domain feature vector recovers to a level similar to that under dry conditions. The support vector machine model can still maintain basic soil and rock category recognition capabilities under rainy conditions. At the same time, the adaptive reduction of the confidence threshold makes the system more inclined to output higher-risk category labels when the uncertainty of recognition increases under rainy conditions, and to initiate thrust reduction or frequency limiting control in advance to adapt to the characteristics of reduced slope stability and reduced tolerance for misjudgment under rainy conditions.

[0023] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0024] Figure 1 This is a flowchart of the weathered rock layer interface identification and excavation parameter adaptive control method based on vibration spectrum analysis of the present invention. Figure 2 This is a flowchart of the rain signal processing and feature correction of the present invention. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to examples, so that those skilled in the art can implement it based on the description.

[0026] (1) such as Figure 1As shown, this invention discloses a method for identifying weathered rock layer interfaces and adaptive control of excavation parameters based on vibration spectrum analysis, applied to the operation scenario of excavators excavating foundation trenches in weathered gneiss sites. A triaxial accelerometer is bolted to the surface of the excavator's boom or arm structure. This sensor can be a piezoelectric accelerometer with a range of ±50g, such as the PCB 356A02 model, used to continuously collect vibration acceleration signals generated during excavation. An embedded controller is configured in the electrical cabinet in the excavator's cab. The controller can be an ARM-based microcontroller unit (MCU), such as the STM32H7 series, which is connected to the accelerometer via an analog input port and continuously receives vibration signals at a preset fixed sampling frequency (e.g., 10 kHz). The controller internally runs signal processing and recognition algorithms. When the excavator starts an excavation cycle, the controller captures the vibration acceleration signal with a sliding acquisition window of a preset time length (e.g., 0.2 seconds) and performs a fast Fourier transform on the time-domain signal to obtain the power spectrum. Subsequently, four characteristic parameters are extracted from the spectrum: the dominant frequency amplitude is the amplitude corresponding to the point of maximum amplitude in the spectrum; the high-frequency band energy proportion is the ratio of the sum of squares of the power spectrum amplitudes above 400 Hz to the sum of squares of the power spectrum amplitudes across the entire frequency band; the spectral centroid is calculated using the first-order moment formula; and the kurtosis is the fourth-order normalized statistical moment of the time-domain waveform in that segment. These four parameters together constitute a four-dimensional frequency domain feature vector.

[0027] The controller pre-stores a trained multi-class support vector machine (SVM) model. This model is trained offline using vibration data collected on-site and labeled by geological engineers before deployment. During the training phase, vibration samples with different postures and weathering degrees are collected for three media: cohesive silt, strongly weathered gneiss, and moderately weathered gneiss with interbedded boulders. Cross-validation and grid search methods are used to optimize the radial basis function parameters and penalty factor of the SVM. In this embodiment, the kernel parameter γ can be 0.1, and the penalty factor C can be 10. During real-time operation, the controller inputs the extracted four-dimensional frequency domain feature vector into the model, and the model outputs the category label of the current working medium. The model output is one of three categories: cohesive silt, strongly weathered gneiss, or moderately weathered gneiss with interbedded boulders. When the category label changes from cohesive silt to strongly weathered gneiss, the controller generates a first control command and a second control command. The first control command is connected to the proportional solenoid valve driver of the excavator's hydraulic main pump via the controller's pulse width modulation output port. This reduces the main pump's displacement by decreasing the input current, thereby reducing the stick's digging thrust. The target current value can be set to 50% of the current operating current. Simultaneously, the second control command is connected to an LED indicator or buzzer in the cab via the controller's digital output port to prompt the operator to switch to the hydraulic breaker. When the category label identifies it as moderately weathered gneiss with interbedded boulders, the controller further determines whether the currently calculated high-frequency band energy percentage exceeds a preset upper limit threshold (e.g., set to 40%) and whether the kurtosis value exceeds a preset upper limit threshold (e.g., set to 6.0). If both conditions are met, a third control command is generated. This command is connected to the electro-hydraulic proportional pressure reducing valve installed on the breaker's pilot oil circuit via the controller's analog output port. This reduces the pilot oil circuit pressure by decreasing the input current, thereby continuously reducing the breaker's impact frequency. The target current value can be set in stages according to the extent of kurtosis exceeding the limit. For example, when the kurtosis exceeds the upper limit threshold by 30%, the target current value is reduced to 40% of the original value.

[0028] In a specific trench excavation cycle, the excavator bucket teeth begin cutting into the surface cohesive silt. At this point, the controller extracts a frequency domain feature vector with a dominant frequency amplitude of approximately 1.2g, a high-frequency energy proportion of approximately 12%, and a kurtosis value of approximately 3.1. The support vector machine model outputs a cohesive silt label with high confidence, and the system does not intervene in the operation. As the bucket teeth gradually contact the lower strongly weathered gneiss layer, the vibration signal's dominant frequency amplitude rises above 3.5g, the high-frequency energy proportion increases to 32%, and the spectral centroid shifts towards higher frequencies. The support vector machine model detects a change in the category label to strongly weathered gneiss, and the controller immediately reduces the input current of the proportional solenoid valve, decreasing the boom thrust by approximately 50%. Simultaneously, indicator lights in the cab begin flashing as a warning. The operator switches the work tool to a hydraulic breaker based on the warning. If the bucket teeth encounter embedded rocks during the breaking process, the vibration signal kurtosis value rises sharply to 8.2, and the high-frequency energy proportion rises to 55%, both exceeding the safe range. The controller immediately generates a third control command, reducing the current of the electro-hydraulic proportional pressure reducing valve to a preset low target value. This significantly reduces the impact frequency of the hydraulic breaker, suppressing rock mass vibration. If the kurtosis value further increases and exceeds the low-frequency safety lock-in threshold (e.g., 1.8 times the upper limit threshold of kurtosis, i.e., 10.8), the controller reduces the input current of the electro-hydraulic proportional pressure reducing valve to zero, prohibiting the hydraulic breaker from operating. It also activates another red indicator light in the cab to prompt the operator to stop breaking and switch to another stripping process. Through this identification and control closed loop, this implementation achieves real-time, non-manual judgment of weathered rock layer interfaces, as well as adaptive adjustment of excavation and breaking parameters. Compared to traditional operating methods that rely on the operator's experience through sound and sight, this method can identify lithological changes earlier and more accurately, and automatically execute thrust limiting and frequency suppression actions, thereby reducing mechanical damage caused by the breaker teeth biting hard rock and lowering the risk of slope spalling induced by high-frequency impact of the hydraulic breaker.

[0029] (2) In another embodiment, continuously acquiring vibration acceleration signals during excavation operations at a preset sampling frequency specifically includes: the output of a triaxial accelerometer (selectable as a PCB 356A02 piezoelectric accelerometer) mounted on the excavator's boom or arm is connected to a signal conditioning circuit via a shielded cable. This signal conditioning circuit includes a programmable gain amplifier, such as the Analog Devices AD8250 instrumentation amplifier, whose gain can be configured via a digital interface and set between the accelerometer and the analog-to-digital converter (ADC). The ADC can be a 12-bit or 16-bit successive approximation ADC built into the controller, such as the internal ADC module of the STM32H7 series microcontroller, with sampling triggering controlled by a timer. The controller internally sets a basic first sampling frequency, such as 10 kHz, for normal operation in continuous acquisition mode. Simultaneously, the controller internally sets a higher second sampling frequency default value, such as 50 kHz, for the initial sampling frequency in impact event triggered acquisition mode. Furthermore, the controller is connected to the gain selection pin of the programmable gain amplifier via a digital input / output port to achieve dynamic adjustment of the gain value. The preset impact threshold can be set to 5g (approximately 49 meters per second squared), the number of consecutive over-threshold points N is 5, and the number of recovery duration points M is 10.

[0030] During system operation, the controller adjusts the gain value of the programmable gain amplifier in real time based on the category label of the current working medium output in step three. When the category label is "viscous silt," the controller outputs a logic level to set the programmable gain amplifier to the first gain value, for example, a gain of 1 (0 dB). When the category label is "strongly weathered gneiss" or "moderately weathered gneiss with interbedded boulders," the controller switches the gain to the second gain value, for example, a gain of 10 (20 dB). The amplified signal is then sampled by the analog-to-digital converter (ADC). In continuous acquisition mode, the controller periodically triggers the ADC to perform conversion at the first sampling frequency (10 kHz). Simultaneously, the controller continuously monitors the real-time amplitude of the digital signal. When the absolute value of the amplitude at N consecutive (e.g., 5) sampling points exceeds the preset impact threshold (5g), the controller determines that a rock impact event has occurred and automatically switches the acquisition mode to impact event triggered acquisition mode. In this mode, the controller modifies the timer period, initially increasing the sampling frequency to the default value of the second sampling frequency (e.g., 50 kHz) to capture the waveform of the first impact event at the initial high sampling frequency, while maintaining the second gain value unchanged. During the impact event-triggered acquisition mode, the controller dynamically adjusts the second sampling frequency based on the kurtosis value calculated in the previous acquisition window. The controller maintains a lookup table internally; for example, when the kurtosis value is between 3.0 and 4.5, the second sampling frequency is adjusted to 30 kHz; when the kurtosis value is between 4.5 and 6.0, it is adjusted to 40 kHz; and when the kurtosis value is greater than 6.0, it is adjusted to 50 kHz. The larger the kurtosis value, the stronger the impact, and the higher the required sampling frequency. Subsequently, the controller continues to monitor the signal amplitude. Once the amplitude falls below the impact threshold and remains below it for M (e.g., 10) sampling points, the controller resumes the timer period, switches back to continuous acquisition mode, and restarts at the first sampling frequency.

[0031] In a complete excavation cycle, initially the bucket teeth are in a silty sand layer, categorized as cohesive silt. The programmable gain amplifier is set to 1x gain, and the ADC samples continuously at 10 kHz. The controller detects stable amplitude and no impact events are triggered. When the bucket teeth encounter a strongly weathered rock layer, the category label changes to strongly weathered gneiss, and the controller immediately switches the gain to 10x while maintaining the sampling frequency at 10 kHz. Due to the increased rock hardness, the vibration amplitude increases significantly, but no sharp impact has yet occurred, so the amplitude may remain between 3g and 4g, not exceeding the 5g threshold. Subsequently, the bucket teeth encounter a boulder, generating a momentary strong impact, with the amplitude exceeding 5g for five consecutive sampling points. Upon detecting this condition, the controller immediately switches the sampling frequency to the second sampling frequency default value of 50 kHz to capture the first impact waveform at a high sampling frequency. After calculating the first impact acquisition window and obtaining the kurtosis value, if the impact continues, the controller adjusts the second sampling frequency based on the kurtosis value; for example, if the kurtosis value is 6.5, it remains at 50 kHz. After the impact, the amplitude drops below 5g and remains below 5g for 10 sampling points. The controller then restores the sampling frequency to 10 kHz. If the operator subsequently starts the breaker for crushing operations, the system continues to adaptively adjust the sampling parameters according to the above mechanism. Through the adaptive adjustment of the programmable gain and sampling frequency, this embodiment operates with low gain and low sampling frequency during stable operation in silty sand layers, reducing data redundancy and processing burden. When a rock impact event occurs, the gain is automatically increased and the initial impact waveform is captured at a preset high sampling frequency. Subsequently, the sampling frequency is dynamically adjusted according to the actual impact intensity to ensure that the high-frequency impact characteristics are captured completely and without distortion, providing high-quality raw data for subsequent spectrum analysis and accurate classification. Compared with the traditional vibration acquisition scheme using a fixed sampling frequency and fixed gain, this embodiment effectively reduces the overall power consumption and computational load of the system while ensuring the fidelity of key signals.

[0032] (3) In another embodiment, the specific composition and calculation method of the frequency domain feature vector includes: after the controller completes the fast Fourier transform of the vibration acceleration signal within the acquisition window, it obtains a discrete power spectrum sequence. The frequency resolution is related to the number of sampling points. For example, for a 0.2-second acquisition window and a 10 kHz sampling frequency, the frequency resolution is 5 Hz. The main frequency amplitude is defined as the amplitude corresponding to the point with the largest amplitude in the transformed spectrum. Its physical meaning is the intensity of the vibration component with the most concentrated energy during rock breaking. When the bucket teeth enter the weathered rock layer from silt, this amplitude can rise from about 1.0g to more than 3.5g. In the calculation of the high-frequency band energy ratio, the entire frequency band from 0 to half of the sampling frequency is divided into a low-frequency band and a high-frequency band. The dividing frequency is set according to the experimental calibration of the rock breaking vibration of weathered gneiss. In this embodiment, it is preferably 400 Hz. The controller accumulates the sum of squares of the power spectral amplitude at each frequency point within the high-frequency band (400 Hz to 5000 Hz), divides this sum by the sum of squares of the power spectral amplitudes at all frequency points across the entire frequency band, and the resulting ratio represents the energy proportion of the high-frequency band. The spectral centroid is the first moment of the power spectrum, which the controller calculates numerically using the following formula: Where f is a discrete frequency point, and P(f) is the power spectral amplitude at that frequency point. s This represents the current actual sampling frequency. The kurtosis value is the fourth-order normalized statistical moment of the time-domain waveform of the vibration acceleration signal in this segment. The controller first calculates the mean μ and standard deviation σ of the time-domain sequence, and then calculates it using the following formula: Where L is the number of sampling points in this segment, x i Let A be the acceleration value at the i-th sampling point. Let the amplitude of the dominant frequency be denoted as A. m The energy percentage of the high-frequency band is denoted as E. h Let C be the centroid of the spectrum and K be the kurtosis value. Then the four-dimensional eigenvector is represented as [A]. m E h [, C, K], which are inputs to the subsequent support vector machine model.

[0033] The selection of the four characteristic parameters mentioned above is not arbitrary, but corresponds to three different physical responses in the excavation of weathered rock layers. The dominant frequency amplitude directly reflects the magnitude of rock-breaking resistance and is most sensitive to changes in hardness. The proportion of high-frequency band energy and the spectral centroid together describe the degree of vibration energy migration to higher frequencies, and are positively correlated with the degree of rock weathering and hardness. When excavating silt layers, the proportion of high-frequency band energy is usually below 15%, and the spectral centroid is located in the 200 Hz to 300 Hz range; after entering strongly weathered gneiss layers, the proportion of high-frequency band energy rises to 25% to 40%, and the spectral centroid shifts to the right to 500 Hz to 700 Hz. The kurtosis value is specifically used to detect the suddenness of the impact. The vibration signal of uniform rock layers approximates a Gaussian distribution, with a kurtosis value close to 3.0; when the bucket teeth encounter boulders and produce a sudden and strong impact, the time-domain waveform shows a spike, and the kurtosis value can rise sharply to above 6.0, or even exceed 10.0. The controller updates the above feature values ​​in real time after each sliding acquisition window, providing a continuous and comparable feature stream for the classifier in step three.

[0034] In a specific operational scenario, the controller performs a Fast Fourier Transform (FFT) on a 0.2-second acquisition window (containing 2048 sampling points). Assuming the transformed spectrum shows the maximum amplitude at 550 Hz with an amplitude of 3.8g, the dominant frequency amplitude is recorded as 3.8. Subsequently, the controller calculates the sum of squares of the power spectrum amplitudes at all frequency points between 400 Hz and 5000 Hz as 0.65 square g, and the sum of squares across the entire frequency band as 1.80 square g, indicating a high-frequency energy proportion of 36.1%. The spectral centroid is obtained as 620 Hz through numerical integration. Simultaneously, the fourth moment of the time-domain acceleration data within this window is calculated, yielding a kurtosis value of 3.4. This four-dimensional vector [3.8, 0.361, 620, 3.4] is input into a support vector machine, and the model determines that it conforms to the characteristic spatial distribution of strongly weathered gneiss. If the bucket teeth collide with boulders during subsequent operations, causing the kurtosis value to suddenly rise to 8.7, the high-frequency band energy proportion to rise to 52%, the dominant frequency amplitude to rise to 5.2g, and the spectral centroid to shift to the right to 890 Hz, the model will output a label for moderately weathered gneiss interbedded with boulders. Compared with existing technologies that rely solely on a single amplitude or simple frequency band energy, this implementation method, through a combination of four feature parameters with clear physical correspondences, can distinguish and describe three geomechanical events: "changes in hardness," "gradual changes in weathering," and "sudden impacts." This provides more discriminative information for subsequent classifiers, avoids confusion between uniform hard rock and boulder interbedded conditions caused by relying solely on vibration intensity, and improves the accuracy of identification and the targeted nature of control command triggering.

[0035] (4) In another implementation, the generation mechanism of the category labels output in step three includes: the multi-class support vector machine model stored internally by the controller is constructed using a one-to-one multi-class strategy based on a decision binary tree. This decision binary tree structure contains three binary nodes, each of which is an independent binary support vector machine classifier. Each classifier uses a radial basis function kernel function, mathematically denoted as K(x...). i ,x j )=exp(-γ·|x i -x j | 2 ), where x i and x j Given two input feature vectors, |x i -x j | represents the Euclidean distance between the two, and γ is the kernel parameter, which determines the range of influence of a single training sample on the classification decision boundary. Each classifier also includes a penalty factor to control the degree of punishment for misclassified samples, balancing model complexity and empirical error.

[0036] The first node uses the extracted main frequency amplitude value A m and the proportion of high-frequency band energy E h The first node uses silt as input features to determine whether the current working medium belongs to "silt" or "rock strata." Its decision boundary is determined by the distribution of silt samples and weathered rock samples in the two-dimensional feature space of the training samples. The second node uses the spectral centroid C and kurtosis value K as input features, and only further subdivides the samples that the first node judged as "rock strata," distinguishing between "strongly weathered gneiss" and "moderately weathered gneiss with interbedded boulders." Since the kurtosis value K increases significantly under the boulder condition, and the spectral centroid C also shifts towards higher frequencies, this node has good linear separability on the two-dimensional plane spanned by these two features. The third node uses kurtosis value K and the proportion of high-frequency band energy E. h As input features, a confirmatory secondary discrimination is performed on samples identified as 'silt' by the first node, mainly addressing the possibility that the first node might misclassify 'moderately weathered gneiss with interbedded boulders' as 'silt'. Near the boundary of weathered rock layers, when the bucket teeth come into contact with interbedded boulders, due to the dominant frequency amplitude A... m The kurtosis K of such samples may not yet be significantly elevated, and the first node might misclassify them as silt based solely on the low dominant frequency amplitude. However, the kurtosis K of these samples is significantly increased due to rock impact, and the high-frequency band energy proportion E... h It also remains at a high level. The third node is determined by the kurtosis value K and the proportion of high-frequency band energy E. h The combined features are reviewed, and if the kurtosis value K of the sample exceeds 4.0 and the high-frequency band energy proportion E h If the percentage exceeds 25%, it will be reclassified as a rock stratum and further subdivided by the second node.

[0037] The training of this multi-class support vector machine model was completed offline before system deployment. The training data was collected as follows: representative sections of silt, strongly weathered gneiss, and moderately weathered gneiss containing interbedded boulders were selected in the construction area. Skilled operators drove excavators to perform normal excavation operations, simultaneously recording vibration acceleration signals and corresponding geological condition labels. For each condition, at least 500 vibration samples from valid acquisition windows were collected, and the extracted four-dimensional feature vector [A] was used to... m E h The training set is formed after [C, K]. Cross-validation and grid search are used to optimize model parameters during training. Specifically, the training set is randomly divided into 5 equal parts, with 4 parts used for training and 1 part for validation each time, repeated 5 times. The average classification accuracy is used as the objective function of the grid search. The parameter space of the grid search includes the kernel parameter γ and the penalty factor. The search range of the kernel parameter γ is set to logarithmically increasing from 0.01 to 10, and the search range of the penalty factor is set to logarithmically increasing from 0.1 to 100. In this embodiment, the kernel parameter γ is preferably set to 0.1, and the penalty factor to 10. Unlike conventional classification tasks, this training process aims to maximize the weighted classification accuracy. The weight coefficients of each training sample are set according to the construction risk level of its category label: the weight of cohesive silt samples is set to 1.0, the weight of strongly weathered gneiss samples is set to 2.0, and the weight of moderately weathered gneiss samples with interbedded boulders is set to 5.0. The higher the construction risk level of a category, the greater the cost of misclassification. Therefore, it is given a higher weight during training so that the model prioritizes the correct identification rate of high-risk categories when optimizing parameters.

[0038] In a complete recognition process, the controller extracts a four-dimensional feature vector [A] from the current acquisition window. m E h [,C, K], for example [3.8, 0.36, 620, 3.4]. This vector is first input to the first node, which is based on the main frequency amplitude A. m =3.8g and high-frequency band energy percentage E h =36% was identified as rock strata. Subsequently, the vector entered the second node, which, based on the spectral centroid C=620 Hz and kurtosis K=3.4, identified it as strongly weathered gneiss. If the feature vector of a certain acquisition window is [2.2, 0.28, 450, 5.5], the first node, based on the lower dominant frequency amplitude A... m =2.2g might be misclassified as silt, but the third node detected a kurtosis value K=5.5 and a high-frequency band energy percentage E. h=28%, triggering a correction mechanism, the sample is rerouted to the second node. The second node outputs the label of moderately weathered gneiss interbedded with boulders based on the spectral centroid C=450 Hz and kurtosis K=5.5. Compared with the existing technology that uses a single-plane multi-class support vector machine or simple threshold comparison, this implementation constructs a decision binary tree structure, which utilizes the four-dimensional feature vector step by step according to a hierarchical strategy of "coarse classification first, then fine classification". Each node uses the feature combination with the strongest discriminative ability for the classification task of that layer, avoiding feature redundancy and classification overlap. At the same time, a risk-sensitive sample weighting mechanism is introduced during the training phase, so that the model prioritizes the recognition recall rate of high-risk conditions, i.e., boulders, when optimizing parameters, thereby reducing the possibility of missing high-risk conditions from the model level. Actual operation tests on the embedded controller show that the average classification of this decision binary tree structure only requires the calculation of one or two support vector machine nodes, compared with the traditional one-to-one strategy which requires the calculation of all three nodes before voting, reducing the amount of computation by about 30%, making it more suitable for real-time operation on embedded platforms with limited computing power.

[0039] (5) After obtaining the original decision values ​​of the support vector machine model, the controller uses the Platt scaling method to map the decision values ​​to posterior probability values. Specifically, for each category label, the controller converts the output distance of the support vector machine into a posterior probability value between 0 and 1 using a pre-fitted Sigmoid function. The fitting parameters are pre-determined through cross-validation on the training set and stored in the controller. The controller internally presets two confidence thresholds: a first confidence threshold and a second confidence threshold. The second confidence threshold is fixed at 0.5 in this embodiment. The first confidence threshold is not a fixed value; its value is dynamically determined according to the construction risk level of the category label corresponding to the maximum posterior probability value. Specifically, when the candidate category corresponding to the maximum posterior probability is "cohesive silt", the first confidence threshold is 0.85; when the candidate category is "strongly weathered gneiss", the first confidence threshold is 0.75; when the candidate category is "moderately weathered gneiss with interbedded boulders", the first confidence threshold is 0.65. The higher the construction risk level, the lower the threshold setting, meaning a more lenient standard for judging high-risk categories. It tends to output a high-risk label even when confidence is low to ensure safety. Furthermore, the scaling factor is preset to 0.7 to reduce the intensity of control commands during the transition state.

[0040] After completing feature extraction and classification calculation in each acquisition window, the controller obtains the posterior probability values ​​corresponding to the three category labels, such as 0.72 for cohesive silt, 0.20 for strongly weathered gneiss, and 0.08 for moderately weathered gneiss with interbedded boulders. The controller identifies the maximum posterior probability value of 0.72 and its corresponding category, cohesive silt. Subsequently, the controller determines the relationship between the maximum posterior probability value of 0.72 and the current first confidence threshold (0.85 since the candidate category is cohesive silt). Because 0.72 is less than 0.85 but greater than the second confidence threshold of 0.5, the controller outputs cohesive silt as a candidate category label and adds a "transitional state" flag. This flag triggers a reduction in the execution intensity of the first control command in subsequent control flows. If the posterior probability values ​​calculated for the next acquisition window are 0.91 for cohesive silt, 0.06 for strongly weathered gneiss, and 0.03 for interbedded rocks, and the maximum posterior probability of 0.91 is greater than the first confidence threshold of 0.85, the controller will directly output cohesive silt as the definitive category label without any additional identifier. If, in a certain acquisition window, due to transient signal interference, the three posterior probability values ​​are 0.40 for cohesive silt, 0.35 for strongly weathered gneiss, and 0.25 for interbedded rocks, with a maximum value of 0.40, which is less than the second confidence threshold of 0.5, the controller will maintain the category label output from the previous acquisition window and add a "low confidence" identifier. This identifier is interpreted in subsequent control logic as prohibiting the updating of control commands; that is, the identification result of this acquisition window will not be used to trigger actions such as thrust adjustment or breaker frequency limiting.

[0041] In a specific soil-rock interface transition scenario, excavator bucket teeth are gradually cutting into a weathered rock layer from a pure silt layer. Before cutting in, the system continuously outputs a label for cohesive silt, with a posterior probability remaining above 0.9. When the bucket teeth just touch the top surface of the weathered rock layer, high-frequency components begin to appear in the vibration signal, with a slight increase in the dominant frequency amplitude, but the characteristics are not yet typical. At this point, the posterior probability output by the support vector machine may become 0.68 for cohesive silt, 0.30 for strongly weathered gneiss, and 0.02 for interbedded boulders. The first confidence threshold for cohesive silt, corresponding to the maximum posterior probability of 0.68, is 0.85. Since 0.68 is less than 0.85 but greater than 0.5, the system outputs cohesive silt with an additional transition state label. The transition state label causes the thrust reduction of the first control command to be scaled by a factor of 0.7, meaning the thrust is reduced by only 70% of the normal range, achieving a smooth transition in the boundary region rather than a drastic switch. As the bucket teeth continue to penetrate deeper, the characteristics of the strongly weathered rock strata gradually intensify, and the posterior probability values ​​change to 0.25 for cohesive silt, 0.72 for strongly weathered gneiss, and 0.03 for interbedded boulders. At this point, the strongly weathered gneiss corresponding to the maximum posterior probability of 0.72 has a first confidence threshold of 0.75. Since 0.72 is still less than 0.75, the system outputs "strongly weathered gneiss" with a transitional state label, and the thrust continues to decrease proportionally. When the bucket teeth have fully entered the rock strata, the posterior probabilities become 0.05 for cohesive silt, 0.92 for strongly weathered gneiss, and 0.03 for interbedded boulders. Since 0.92 is greater than the threshold of 0.75, the system outputs the category label "strongly weathered gneiss," the first control command is executed in full, and the thrust is reduced to the target value. Through the above-mentioned layered generation mechanism, this implementation avoids frequent jumps in category labels and drastic fluctuations in control commands when identifying boundary transition zones with high uncertainty. At the same time, by setting a lower confidence threshold for high-risk categories, it achieves a safety-first classification decision tendency. Compared with existing technologies that only output hard classification labels without providing confidence information, this implementation enables the controller to adopt differentiated control strategies based on the reliability of the recognition results, thereby improving the robustness and safety of the entire system under complex operating conditions.

[0042] (6) In another embodiment, the specific execution method of the first and second control commands generated when the category label changes from "cohesive silt" to "strongly weathered gneiss" is as follows. The interface between the controller and the excavator's hydraulic system is connected to the proportional solenoid valve drive amplifier of the hydraulic main pump through a pulse width modulation output port. This drive amplifier can be a Rexroth VT-2000 series proportional amplifier, whose input is a 0 to 10 volt analog voltage signal or an equivalent pulse width modulation signal output by the controller, and whose output is the current driving the proportional solenoid valve. The controller internally presets a first decay rate curve, which is a set of discrete current decrease step sequence with a step interval period of 20 milliseconds. The first decay rate is not a fixed value, but is dynamically determined according to the maximum a posteriori probability value mentioned before the category label change. Specifically, the controller incorporates a lookup table: when the maximum a posteriori probability is greater than 0.90, the first decay rate is 3% of the current value per step; when the maximum a posteriori probability is between 0.75 and 0.90, it is 5% per step; and when the maximum a posteriori probability is less than 0.75, it is 8% per step. A higher a posteriori probability indicates greater confidence in the classification result, prioritizing smooth operation and thus resulting in slower decay. Conversely, a lower a posteriori probability suggests the possibility of a transition zone or precursors to inclusions, necessitating timely response and thus faster decay. The target current value is determined based on a preset percentage corresponding to the strongly weathered gneiss category label; in this embodiment, this percentage is set to 50% of the current operating current value. The current operating current value is obtained by the controller through sampling the current detection resistor in the proportional solenoid valve drive circuit via an analog-to-digital converter at the moment the category label changes, or read from the feedback signal of the drive amplifier.

[0043] The second control command is used to activate the breaker operation prompt signal in the cab. The prompt signal device consists of a dual-color LED and a small buzzer, mounted on the right-side control panel of the cab. The controller determines the intensity of the prompt signal based on the real-time change rate of the main frequency amplitude and the energy percentage of the high-frequency band. The amplification rate of the main frequency amplitude is calculated by dividing the difference between the main frequency amplitude of the current acquisition window and the previous acquisition window by the acquisition window duration, expressed in grams per second (g / s). The preset amplification threshold is set to 2.0 g / s. The preset energy threshold for the energy percentage of the high-frequency band is set to 35%. When the amplification rate of the main frequency amplitude exceeds 2.0 g / s for K consecutive sampling points (K is 3 in this embodiment), or when the real-time value of the energy percentage of the high-frequency band exceeds 35%, the controller increases the flashing frequency or sound intensity of the prompt signal. The base value of the flashing frequency is 1 Hz, and the increased value is 3 Hz; the base volume of the buzzer is 60 dB, and the increased volume is 80 dB. This tiered prompting allows the operator to intuitively perceive the hardness and impact intensity of the current rock stratum based on the prompt intensity, thereby rationally scheduling the timing of switching the hydraulic breaker. When the category label includes the aforementioned "transitional state" indicator, the first control command reduces the attenuation of the input current of the proportional solenoid valve by a proportional factor. Depending on the actual working conditions, the proportional factor can also be selected within the range of 0.5 to 0.8. For example, when the slope stability is good or the rock transition zone is thin, a smaller value of 0.5 can be used to accelerate thrust adjustment; when the transition zone is thick or the geological conditions are complex, a larger value of 0.8 can be used to make the thrust change more gradual. In this embodiment, the proportional factor is preset to 0.6, that is, in the transition state, the current attenuation amount per step is multiplied by 0.6, making the thrust reduction process more gradual, in order to adapt to the characteristics of incompletely determined lithology in the boundary area.

[0044] In a complete control scenario, the excavator operates normally in a cohesive silt layer, with a proportional solenoid valve input current of 600 mA and ample stick thrust. The controller continuously outputs a cohesive silt label, maintaining a posterior probability of approximately 0.92. When the bucket teeth touch the top surface of the strongly weathered gneiss layer, the dominant frequency amplitude increases from 1.5g to 3.2g within 0.2 seconds, an increase rate of 8.5g per second, exceeding the 2.0g per second threshold; simultaneously, the high-frequency band energy percentage increases from 12% to 31%. The controller detects a change in the category label from cohesive silt to strongly weathered gneiss, and the maximum posterior probability before the label change is 0.92. According to the lookup table, the first decay rate is set to 3% per step. Starting from the current operating current of 600 mA, the controller updates the target current value to 97% of the previous value every 20 milliseconds. That is, after the first step, it decreases to 582 mA; after the second step, it decreases to 564 mA, and so on, until it reaches the target value of 300 mA (50% of 600 mA). The entire attenuation process takes approximately 0.7 seconds. Simultaneously, because the main frequency amplitude increase has exceeded the threshold, the controller increases the LED flashing frequency to 3 Hz and the buzzer volume to 80 dB to guide the operator to switch the breaker in time. If a transition status indicator is attached when this change occurs, the attenuation amount per step is multiplied by a scaling factor of 0.6, meaning the actual attenuation per step is 1.8%, resulting in a smoother current decrease. Compared to existing technologies that immediately reduce the current to the target value upon identifying changes in rock strata or rely solely on the operator's subjective judgment to switch tools, this implementation avoids machine shaking and hydraulic shock caused by a sudden drop in thrust through a confidence-correlated attenuation rate design. It also reduces the cognitive delay from alert to execution by using tiered warning signals to provide the operator with a warning intensity matched to the rock strata hardness. Furthermore, it maintains appropriate digging force in the boundary area through proportional factor attenuation during transitional states to ensure successful cutting through the transition zone. These measures collectively achieve a smooth and safe transition from identification to control.

[0045] (7) In another embodiment, the specific execution method of generating the third control instruction in step four when the category label is "moderately weathered gneiss with interbedded boulders" is as follows. The controller internally presets a safety range defined by the upper limit threshold of high frequency band energy ratio and the upper limit threshold of kurtosis. The values ​​of these two upper limit thresholds are not fixed constants, but are determined according to the geomechanical parameters of the slope of the construction area. Before the system is deployed, the construction unit conducts supplementary geological surveys of the excavation area of ​​the foundation trench to obtain three parameters: uniaxial compressive strength, fissure development density, and groundwater depth of the weathered gneiss. For example, the uniaxial compressive strength of strongly weathered to moderately weathered gneiss is between 15 MPa and 28 MPa as measured by point load tests; the fissure development density is 5 to 12 per meter as obtained through structural surface statistical mapping; and the groundwater depth is 6 to 8 meters below the surface. Based on the above parameters, the threshold calculation module in the controller determines the upper limit threshold according to the pre-established mapping relationship: the lower the uniaxial compressive strength, the higher the fissure density, and the shallower the groundwater level, the more stringent the threshold setting. In this embodiment, for the working conditions, the upper limit threshold for the proportion of high-frequency band energy is set to 40%, and the upper limit threshold for kurtosis value is set to 6.0.

[0046] An electro-hydraulic proportional pressure reducing valve is installed in the pilot oil circuit of the hydraulic breaker. The Eaton Vickers KBFDG4V series or similar models can be selected. Its pilot pressure control port is connected to the analog output port of the controller via a proportional electromagnet. The controller's analog output is a 0-10 volt voltage signal or a 4-20 mA current signal, corresponding to the output pressure adjustment range of the electro-hydraulic proportional pressure reducing valve (e.g., 0-35 bar). The impact frequency of the hydraulic breaker is positively correlated with the pilot oil circuit pressure. Reducing the pilot pressure can decrease the main valve stroke and prolong the switching time, thereby reducing the piston reciprocating frequency. The controller has a preset second decay rate, which in this embodiment is set to reduce the current value by 5% every 50 milliseconds to continuously and smoothly adjust the pilot pressure. The target current value is determined according to the kurtosis value exceeding the limit. The controller divides the kurtosis value exceeding the limit into three intervals: when the exceeding limit is less than 20%, the target current value is 70% of the current value; when the exceeding limit is between 20% and 50%, the target current value is 40% of the current value; and when the exceeding limit is greater than 50%, the target current value is 15% of the current value. In addition, the controller has a built-in low-frequency safety lockout threshold, which is 1.8 times the upper limit threshold of the kurtosis value, i.e., 10.8.

[0047] During operation, when the category label output in step three is "moderately weathered gneiss with interbedded boulders," the controller reads the high-frequency band energy percentage and kurtosis value from the frequency domain feature vector of the current acquisition window. Assuming the current high-frequency band energy percentage is 52% and the kurtosis value is 8.2, the controller determines that both parameters exceed their respective upper limit thresholds (40% and 6.0), and then generates a third control command. The controller first determines the kurtosis value exceeding the limit: 8.2 minus 6.0 equals 2.2, exceeding the limit by approximately 37%, falling into the second grade interval. Therefore, the target current value is determined to be 40% of the current operating current value. Subsequently, the controller, according to a preset second decay rate, reduces the analog voltage output to the electro-hydraulic proportional pressure reducing valve by 5% of the current value every 50 milliseconds until the target value is reached. As the input current continuously decreases, the output pressure of the electro-hydraulic proportional pressure reducing valve gradually decreases, and the impact frequency of the hydraulic breaker gradually decreases from the normal operating frequency (e.g., 600 times per minute) to approximately 240 times per minute. If the kurtosis value further increases to 12.0 during subsequent operations due to the increased volume of the rock, exceeding the low-frequency safety lock-in threshold of 10.8, the controller will execute the low-frequency safety lock-in mechanism: directly reducing the input current of the electro-hydraulic proportional pressure reducing valve to zero milliamperes, completely closing the pilot oil circuit, and stopping all impact actions of the breaker. Simultaneously, the controller will connect a red "low-frequency safety lock-in" indicator light in the cab via its digital output port, alerting the operator that the current operating conditions have exceeded the safe operating range of the breaker and that other stripping techniques should be used. For example, the bucket teeth could be used to strip away the weathered rock surrounding the rock to expose it before prying it from the natural fracture surface, or a static expansion agent could be used.

[0048] Unlike existing adaptive control technologies for hydraulic breakers that rely solely on the rod rebound time to determine hardness or prioritize efficiency, this implementation method, for the first time, actively limits the impact frequency of the hydraulic breaker from a slope safety perspective by combining the high-frequency energy ratio and kurtosis value as safety criteria. The high-frequency energy ratio reflects the absolute intensity of high-frequency vibration during rock breaking, while the kurtosis value is highly sensitive to the suddenness of rock impacts. The combination of these two factors accurately distinguishes between "uniform hard rock" and "rock-bundled" conditions. The correlation design between threshold values ​​and geomechanical parameters allows the same control logic to automatically adapt to different geological conditions at different work sites by adjusting the threshold, eliminating the need for repeated trial and error on-site. The mechanism of determining the target current value based on the degree of kurtosis exceeding the limit achieves a positive correlation between frequency limiting intensity and impact intensity, avoiding the construction efficiency loss caused by a "one-size-fits-all" approach to frequency limiting. The low-frequency safety lockout mechanism prohibits the use of the hydraulic breaker at the hardware level when the kurtosis value is extremely high, eliminating the risk of slope crack expansion induced by high-energy impacts at the source. The above measures together form a complete protection loop from identification to frequency limiting to security locking, filling the gap in existing technology for effective safety guarantees under high-risk rock-clamping conditions.

[0049] (8) such as Figure 2As shown, in another embodiment, the process of rain environment perception and signal compensation correction before extracting the frequency domain feature vector in step two is as follows: A piezoelectric rain sensor is installed on the unobstructed area of ​​the top of the excavator cab or the upper rotating body via a bracket. A Vaisala RAINCAP type or similar piezoelectric raindrop sensing module can be selected. Its output is an analog voltage signal, linearly related to the rainfall intensity, with a measurement range of 0 to 20 mm / hour. A microelectromechanical system (MEMS) humidity sensor is attached to the mounting base surface of the triaxial accelerometer (i.e., the edge of the contact surface between the sensor and the metal structure of the boom or rocker arm). A Sensirion SHT30 type can be selected. It is connected to the controller via an I2C digital interface and is used to detect the local humidity state at the sensor-structure interface. The controller internally presets two rainfall intensity thresholds: the first intensity threshold is set to 2.5 mm / hour, corresponding to the boundary between light and moderate rain; the second intensity threshold is set to 8.0 mm / hour, corresponding to the boundary between moderate and heavy rain. In addition, the controller has a high-frequency band dividing frequency, which is set to 400 Hz in this embodiment, consistent with the dividing frequency of the high and low frequency bands mentioned above.

[0050] When the real-time rainfall intensity detected by the rain sensor exceeds the first intensity threshold of 2.5 mm per hour, the controller activates the rain signal processing mode. First, noise reduction processing is performed on the analog-to-digital converted vibration acceleration digital signal. The noise reduction algorithm uses adaptive spectral subtraction. The controller first acquires a segment of background vibration signal from a non-excavation period as a noise template, performs a Fast Fourier Transform on the noise template to obtain an estimate of the power spectrum of raindrop impact noise, and then subtracts this noise power spectrum from the power spectrum of the mixed signal to obtain the denoised signal spectrum. The noise reduction intensity parameter is determined based on the rainfall intensity: the greater the rainfall intensity, the higher the update frequency of the noise template, and the larger the over-subtraction factor of the spectral subtraction. For example, the over-subtraction factor is 1.5 for light rain, 2.0 for moderate rain, and 2.5 for heavy rain. After noise reduction processing, the controller performs high-frequency energy compensation processing on the signal. The compensation method is as follows: For each frequency point above 400 Hz after analog-to-digital conversion, a frequency-dependent compensation gain G(f) = G0 × (1 + α·(f-400)) is applied, where G0 is the reference gain, f is the frequency, and α is the frequency dependence coefficient. The reference gain G0 is determined according to the rainfall intensity level: 1.2 for light rain, 1.5 for moderate rain, 2.0 for heavy rain, and 3.0 for torrential rain. The frequency dependence coefficient α is taken as 0.0005 in this embodiment. The compensation gain increases linearly with increasing frequency to inversely compensate for the differential absorption effect of water bodies on high-frequency vibration energy.

[0051] The controller internally stores a pre-built rainfall intensity-feature correction mapping table, an example of which is shown below: Table 1 Rainfall Intensity-Feature Correction Mapping Table The mapping table was established as follows: A section of uniformly lithological gneiss, known to be strongly weathered, was selected in the construction area. Calibration excavation experiments were conducted under dry conditions, light rain conditions (rainfall intensity approximately 1.5 mm / h), moderate rain conditions (rainfall intensity approximately 5 mm / h), and heavy rain conditions (rainfall intensity approximately 12 mm / h). Vibration signals were collected from at least 20 acquisition windows under each rainfall condition. The ratios of the two characteristic parameters under each rainfall condition relative to the dry condition were calculated based on the extracted dominant frequency amplitude and high-frequency band energy proportion. Statistical results showed that the average correction coefficient for dominant frequency amplitude and light rain conditions was 1.12, 1.25 under moderate rain conditions, and 1.45 under heavy rain conditions; the average correction coefficient for high-frequency band energy proportion was 1.18 under light rain conditions, 1.40 under moderate rain conditions, and 1.80 under heavy rain conditions. These average values ​​were stored in the mapping table as correction coefficients for each rainfall intensity level. In actual operation, the controller obtains the corresponding correction coefficient based on real-time rainfall intensity interpolation or table lookup. The main frequency amplitude and high-frequency band energy proportion extracted from the current acquisition window are multiplied by the correction coefficient to obtain the corrected values. The corrected main frequency amplitude and high-frequency band energy proportion replace the original values ​​and, together with the uncorrected spectral centroid and kurtosis values, constitute the corrected four-dimensional frequency domain feature vector, which is then input into the multi-class support vector machine model in step three. When the rainfall intensity detected by the rain sensor exceeds the second intensity threshold of 8.0 mm / hour, the controller multiplies the first confidence threshold by a preset scaling factor in the category label generation process. In this embodiment, the scaling factor is 0.8, meaning the first confidence threshold for strongly weathered gneiss, originally set to 0.75, is reduced to 0.60, and the threshold for interbedded boulders, originally set to 0.65, is reduced to 0.52. Lowering the confidence threshold makes the system more inclined to output higher-risk category labels when the uncertainty of identification increases in rainy weather, prioritizing construction safety.

[0052] In a specific rainy weather operation scenario, the rainfall intensity measured by the rain gauge was 6.0 mm / h, which falls under the moderate rain level, exceeding the first intensity threshold but not reaching the second. The controller initiated noise reduction processing, performing spectral subtraction with an over-subtraction factor of 2.0 to effectively suppress background noise generated by raindrops impacting the metal casing. Subsequently, a compensation amplification with a reference gain of 1.5 was applied to the frequency band above 400 Hz. After processing, the original feature vector of the vibration signal was extracted, assuming a dominant frequency amplitude of 2.8g and a high-frequency band energy proportion of 28%. The controller looked up the correction coefficients corresponding to the moderate rain level: dominant frequency amplitude 1.25, high-frequency band energy proportion 1.40. After correction, the dominant frequency amplitude was 3.5g, and the high-frequency band energy proportion was 39%, restoring it to a level close to that of dry conditions. The corrected feature vector was input into the support vector machine model, which correctly identified it as strongly weathered gneiss. If the rainfall intensity reaches 10.0 mm / hour, exceeding the second intensity threshold, the controller multiplies the first confidence threshold by 0.8 when generating the tag. This makes the system more inclined to classify the area as rock strata in the boundary transition zone, and initiates thrust reduction control in advance to adapt to the reduced slope stability and lower tolerance for misjudgment during rainy weather. Compared with traditional vibration identification schemes that do not consider the impact of rain, this implementation maintains the basic accuracy of soil and rock identification under rainfall conditions through dual compensation of environmental perception and signal processing. At the same time, it automatically strengthens the safety protection tendency when environmental risks increase through adaptive adjustment of the confidence threshold.

[0053] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Further modifications can be readily implemented by those skilled in the art.

Claims

1. A method for weathered rock strata interface identification and adaptive control of excavation parameters based on vibration spectrum analysis, characterized in that, Includes the following steps: Step 1: Fix a triaxial accelerometer on the surface of the stick or rocker arm structure of the excavator working device, and continuously collect vibration acceleration signals during the excavation operation at a preset sampling frequency; Step 2: Perform a fast Fourier transform on the vibration acceleration signal within the acquisition window to extract the frequency domain feature vector. The frequency domain feature vector includes the main frequency amplitude, the energy proportion of the high-frequency band, the spectral centroid, and the kurtosis value. Step 3: Input the frequency domain feature vector into the pre-trained multi-class support vector machine model and output the category label of the current working medium. The category label should be divided into at least cohesive silt, strongly weathered gneiss and moderately weathered gneiss with interbedded boulders. Step 4: Generate the first and second control commands based on the category labels: When the category label changes from cohesive silt to strongly weathered gneiss, the first control command is used to reduce the input current of the proportional solenoid valve of the excavator's hydraulic main pump to reduce the stick digging thrust, while the second control command is used to activate the breaker hammer operation prompt signal in the cab. When the category label is moderately weathered gneiss with interbedded boulders, if the high-frequency band energy ratio and kurtosis value exceed the preset safe range, a third control command is generated. The third control command is used to limit the pilot oil circuit pressure of the hydraulic breaker to reduce the impact frequency.

2. The method for weathered rock strata interface identification and adaptive control of excavation parameters based on vibration spectrum analysis according to claim 1, characterized in that, Vibration acceleration signals during the excavation operation are continuously acquired at a preset sampling frequency, specifically including: The output signal of the triaxial accelerometer is acquired at the first sampling frequency, and the output signal is amplified by a programmable gain amplifier located between the triaxial accelerometer and the analog-to-digital converter. The gain value of the programmable gain amplifier is determined by the category label of the current working medium. When the category label is cohesive silt, the first gain value is set, and when the category label is strongly weathered gneiss or moderately weathered gneiss with boulders, the second gain value is set to be greater than the first gain value. The signal amplified by the programmable gain amplifier is monitored in real time. When the signal amplitude exceeds the preset impact threshold within N consecutive sampling points, the current acquisition mode is switched from continuous acquisition mode to impact event triggered acquisition mode. In impact event triggered acquisition mode, the signal is acquired at a second sampling frequency higher than the first sampling frequency and amplified with a second gain value until the signal amplitude falls back below the preset impact threshold and continues for M sampling points, then it is restored to the continuous acquisition mode. The value of the second sampling frequency is determined based on the kurtosis value calculated in the previous sampling window. The larger the kurtosis value, the higher the second sampling frequency.

3. The method for weathered rock strata interface identification and adaptive control of excavation parameters based on vibration spectrum analysis according to claim 1, characterized in that, In the frequency domain eigenvectors: The main frequency amplitude is the amplitude corresponding to the point of maximum amplitude in the spectrum after the Fast Fourier Transform; The high-frequency band energy ratio is defined as follows: the entire frequency band from 0 to half of the sampling frequency is divided into a low-frequency band and a high-frequency band, and the sum of the power spectral values ​​corresponding to all frequency points in the high-frequency band is taken as the ratio to the sum of the power spectral values ​​corresponding to all frequency points in the entire frequency band. The centroid of the spectrum is the first moment of the power spectrum, calculated using the following formula: ； Where C is the centroid of the spectrum, f is the frequency, and P(f) is the power spectral amplitude at frequency f. s The sampling frequency; The kurtosis value is the fourth-order normalized statistical moment of the time-domain waveform of the vibration acceleration signal.

4. The method for weathered rock strata interface identification and adaptive control of excavation parameters based on vibration spectrum analysis according to claim 1, characterized in that, Inputting the frequency domain feature vectors into a pre-trained multi-class support vector machine model specifically includes: The multi-class support vector machine model is constructed using a one-to-one multi-class strategy based on a decision binary tree. Its structure includes three binary nodes: the first node uses the dominant frequency amplitude and the proportion of high-frequency band energy as input features to determine whether the current working medium belongs to silt or rock strata; the second node uses the spectral centroid and the kurtosis value as input features to further subdivide the samples that the first node judges as rock strata into strongly weathered gneiss and moderately weathered gneiss with boulders; the third node uses the kurtosis value and the proportion of high-frequency band energy as input features to perform a confirmatory secondary discrimination on the samples that the first node judges as silt, in order to filter out the samples that the first node misclassifies as silt by moderately weathered gneiss with boulders. The training of the multi-class support vector machine model includes: optimizing the kernel function parameters and penalty factor of the support vector machine model with the goal of maximizing the weighted classification accuracy through cross-validation and grid search methods. The weight coefficient of each sample is set according to the construction risk level of its category label. The construction risk level is determined in the order of moderately weathered gneiss with boulders, higher than strongly weathered gneiss, and higher than cohesive silt.

5. The method for weathered rock strata interface identification and adaptive control of excavation parameters based on vibration spectrum analysis according to claim 1, characterized in that, Category labels are generated using the following hierarchical generation mechanism: The multi-class support vector machine model outputs the posterior probability value corresponding to each class label. When the maximum posterior probability value is greater than or equal to the preset first confidence threshold, the class label corresponding to the maximum posterior probability value is directly output as the determined class label of the current working medium. When the maximum a posteriori probability value is less than the first confidence threshold but greater than or equal to the preset second confidence threshold, the category label corresponding to the maximum a posteriori probability value is output as the candidate category label of the current working medium, and a transition state indicator is attached. The transition state indicator triggers the first control instruction in step four to reduce the execution intensity by a preset scaling factor. When the maximum posterior probability value is less than the second confidence threshold, the category label output by the previous acquisition window remains unchanged, and a low confidence flag is added. The low confidence flag prevents the control command update in step four. The first confidence threshold is determined based on the construction risk level of the category label corresponding to the maximum posterior probability value: the construction risk level is determined in the order of moderately weathered gneiss with boulders, higher than strongly weathered gneiss, which is higher than cohesive silt. The higher the construction risk level, the lower the first confidence threshold.

6. The method for weathered rock strata interface identification and adaptive control of excavation parameters based on vibration spectrum analysis according to claim 5, characterized in that, When the category label changes from cohesive silt to strongly weathered gneiss, the first control command controls the input current of the proportional solenoid valve to gradually decrease from the current operating current value to the preset target current value according to the preset first decay rate. The first decay rate is determined based on the maximum a posteriori probability value before the category label change: the higher the maximum a posteriori probability value, the slower the first decay rate, so as to prioritize the smoothness of the action when the identification confidence is high and prioritize the timeliness of the response when the identification confidence is low. The target current value is determined based on a preset percentage corresponding to the strongly weathered gneiss category label, with the preset percentage being 40% to 60% of the current operating current value; The second control command determines the flashing frequency or sound alert intensity of the warning signal in the cockpit based on the real-time change rate of the main frequency amplitude and the energy ratio of the high-frequency band: when the increase of the main frequency amplitude in K consecutive sampling points exceeds the preset increase threshold, or the real-time value of the energy ratio of the high-frequency band exceeds the preset energy threshold, the flashing frequency or sound alert intensity is increased to guide the operator to switch the breaker first. When the category label includes a transition state indicator, the first control command reduces the attenuation of the input current of the proportional solenoid valve by a proportional factor, the value of which ranges from 0.5 to 0.

8.

7. The method for weathered rock strata interface identification and adaptive control of excavation parameters based on vibration spectrum analysis according to claim 5, characterized in that, The preset safe zone is jointly defined by the upper limit threshold of the high frequency band energy proportion and the upper limit threshold of the kurtosis value. The upper limit threshold of the high frequency band energy proportion and the upper limit threshold of the kurtosis value are determined according to the geomechanical parameters of the slope of the construction area. The geomechanical parameters include the uniaxial compressive strength of weathered gneiss, the density of fracture development and the depth of groundwater level. When the high-frequency band energy ratio exceeds the upper limit threshold of the high-frequency band energy ratio and the kurtosis value exceeds the upper limit threshold of the kurtosis value, the third control command controls the input current of the electro-hydraulic proportional pressure reducing valve installed on the pilot oil circuit of the breaker to gradually decrease from the current working current value according to the preset second decay rate. The output pressure of the electro-hydraulic proportional pressure reducing valve decreases as the input current decreases, thereby continuously reducing the impact frequency of the breaker. The target value for reducing the input current of the electro-hydraulic proportional pressure reducing valve is determined according to the degree of kurtosis exceeding the limit. The greater the kurtosis exceeds the upper limit threshold, the lower the target value and the lower the corresponding impact frequency of the hydraulic breaker. When the kurtosis value exceeds the preset low-frequency safety lockout threshold, the third control command controls the input current of the electro-hydraulic proportional pressure reducing valve to reduce to zero, so as to prevent the breaker from being activated. At the same time, the low-frequency safety lockout warning signal in the cockpit is activated. The low-frequency safety lockout threshold is 1.5 to 2.0 times the upper limit of the kurtosis value.

8. The method for weathered rock strata interface identification and adaptive control of excavation parameters based on vibration spectrum analysis according to claim 1, characterized in that, Before extracting the frequency domain feature vector in step two, the following steps are also included: A rain sensor is installed on the top of the excavator cab or the upper slewing body, and a humidity sensor is installed on the mounting base surface of the triaxial accelerometer. When the rainfall intensity detected by the rain sensor exceeds a preset first intensity threshold, the vibration acceleration signal is subjected to noise reduction processing and high-frequency energy compensation processing in sequence. The noise reduction intensity parameter is determined according to the rainfall intensity; the greater the rainfall intensity, the greater the noise reduction intensity, in order to suppress the impact noise generated by raindrops hitting the surface of the equipment. The high-frequency energy compensation processing applies a frequency-dependent compensation gain to the digital signal after analog-to-digital conversion that is higher than the preset boundary frequency. The compensation gain increases with the frequency, and the reference value of the compensation gain is determined according to the rainfall intensity level at which the rainfall intensity is located. The rainfall intensity is input into a pre-established rainfall intensity-feature correction mapping table. The extracted dominant frequency amplitude and high-frequency band energy proportion are multiplied by the corresponding correction coefficients to obtain the corrected dominant frequency amplitude and high-frequency band energy proportion. The rainfall intensity-feature correction mapping table is pre-established by conducting calibration excavation experiments on known soil and rock types under dry conditions and different rainfall intensities. After replacing the original values ​​with the corrected main frequency amplitude and the corrected high-frequency band energy ratio, they together with the spectral centroid and kurtosis value form the corrected frequency domain feature vector, which is then input into the multi-class support vector machine model. When the rainfall intensity detected by the rain sensor exceeds the preset second intensity threshold, the first confidence threshold used when generating the category label is multiplied by a preset scaling factor, with the scaling factor ranging from 0.7 to 0.9.

Images