Method and device for detecting crushing resistance of aggregate based on impact vibration spectrum analysis

Abstract

The invention relates to the technical field of material dynamic performance detection, and discloses an aggregate anti-crushing capability detection method and device based on impact vibration spectrum analysis, and the method comprises the following steps: collecting a vibration signal generated when an aggregate is impacted; performing dynamic wavelet packet decomposition on the vibration signal and selecting an optimal decomposition layer number; constructing a time-frequency-space three-dimensional tensor feature space based on a dynamic wavelet packet decomposition result; inputting the three-dimensional tensor into a depth measurement learning model to evaluate the crushing resistance; driving an adjustable mass impact clock and a precise lifting mechanism to adjust impact parameters according to an evaluation result; and performing signal processing and model calculation in the steps S2 to S4 in real time. According to the invention, through dynamic impact loading and full-process automatic control, a traditional static detection mode depending on manual operation is changed; multi-band vibration characteristic analysis is combined with a deep metric learning model, a plurality of frequency spectrum parameters and acoustic emission parameters can be synchronously extracted, a multi-dimensional evaluation system is constructed, and the comprehensiveness and scientificity of detection indexes are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of material dynamic performance detection, and in particular to a method and device for detecting aggregate anti-crushing ability based on impact vibration spectrum analysis. Background Art

[0002] Currently, aggregate crushing resistance testing strictly adheres to the static crushing value test method specified in the JTG3432 standard. This method, which applies a 200kN fixed load via a hydraulic press and measures the percentage of fine material after crushing, has become a benchmark for aggregate quality control in road construction projects.

[0003] To evaluate performance under dynamic loads, the engineering community generally uses extrapolation of static test data, establishing equivalent relationships through theoretical conversion. Some studies have attempted to incorporate vibration tables to simulate dynamic conditions, but the core evaluation metric remains based on a single parameter, crushing mass ratio. Industry standards implicitly use static test results as an approximate reference for dynamic conditions.

[0004] However, the existing technology for testing the crushing resistance of aggregates has the following systematic defects: First, since the crushing value test is carried out under static pressure, it cannot simulate the dynamic stress conditions in actual engineering projects, resulting in the evaluation results not accurately reflecting the crushing resistance of aggregates in actual use environments, and it is difficult to effectively guide engineering material selection and quality control; second, the test requires precise screening of aggregates to obtain samples with a specific particle size range. The operation process is complicated, and improper operation may cause the sample particle size distribution to change during the screening process, thereby affecting the accuracy of the test results; finally, the test process is greatly affected by human factors. Differences in operations such as loading and loading speed control among different test personnel make the repeatability and comparability of the test results poor, making it difficult to obtain stable and reliable evaluation results. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and device for detecting the anti-crushing ability of aggregates based on impact vibration spectrum analysis, which solves the problem that static loading cannot simulate the dynamic impact energy transfer characteristics.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for detecting the anti-crushing ability of aggregates based on impact vibration spectrum analysis comprises the following steps: S1: The vibration signal generated by the impact of aggregate is collected through a multimodal sensor array; S2: using a data acquisition and processing unit to perform dynamic wavelet packet decomposition on the vibration signal and select an optimal decomposition layer number; S3: The data acquisition and processing unit constructs a time-frequency-space three-dimensional tensor feature space based on the dynamic wavelet packet decomposition result; S4: The embedded control module inputs the three-dimensional tensor into a deep metric learning model to evaluate the anti-crushing capability; S5: The embedded control module drives the adjustable mass impact clock and the precision lifting mechanism to adjust the impact parameters according to the evaluation results; S6: The signal processing and model calculation of S2 to S4 are performed in real time through the hardware acceleration module.

[0007] Preferably, the S1 includes: Collect three-dimensional vibration signals through a three-axis acceleration sensor; The stress wave signal in the frequency band of 100-300kHz is captured by the acoustic emission sensor.

[0008] Preferably, the dynamic wavelet packet decomposition in S2 includes: Calculate the energy entropy value of each decomposition layer node H(L) = -∑p n logp n ,in By minimizing the objective function H(L) + 0.15·2 L Select the optimal number of decomposition levels L opt .

[0009] Preferably, the S3 includes: Arrange the multi-channel time-frequency data into original tensors in three dimensions according to frequency bands, time windows, and sensor channels; Tucker decomposition is used to reduce the original tensor dimension to a core tensor of 32×16×9; Retain more than 95% of the energy information of the original tensor.

[0010] Preferably, the deep metric learning model in S4 includes: A triplet network structure is used to extract 128-dimensional feature vectors; Training was performed using a contrastive loss function with a margin of α = 0.5; The features are mapped to a crush resistance score in the interval [0, 1] using the Sigmoid function.

[0011] Preferably, the S5 includes: Construct a spectral gradient Acoustic emission standard deviation σ ae Energy consumption E c 12-dimensional state space; Define the discrete action space {-5,0,+5}kg×{-0.2,0,+0.2}m; Designing the reward function Among them, Q t is the aggregate anti-crushing ability score evaluated by the deep metric learning model after the t-th impact; is the crushing resistance score evaluated after the t-1th impact; E t is the total energy consumed by the system during the tth impact.

[0012] Preferably, the hardware acceleration module in S6 realizes the real-time requirements of signal processing and model calculation through a heterogeneous computing architecture, and maps the key algorithms of steps S2 to S4 into dedicated hardware logic units; The hardware acceleration module adopts dynamic pipeline reorganization technology and can adaptively adjust the allocation of computing resources according to the processing stage.

[0013] Preferably, the device for detecting the anti-crushing ability of aggregates based on impact vibration spectrum analysis comprises: A multimodal sensor array, including a triaxial acceleration sensor and an acoustic emission sensor, is arranged in a ring around the test mold and the test mold platform, and a fixture is set inside the test mold; An adjustable mass impact clock is connected to a precision lifting mechanism via an electromagnetic lock, providing three impact masses of 5kg, 10kg, and 15kg; a data acquisition and processing unit is connected to the sensor array and includes a dynamic wavelet packet decomposition module and an FPGA hardware accelerator; an embedded control module is connected to the data acquisition and processing unit and the precision lifting mechanism and includes a built-in deep metric learning model; wherein, the precision lifting mechanism achieves height adjustment by driving a ball screw via a servo motor, and the FPGA hardware accelerator is configured with a dedicated logic unit for tensor operations.

[0014] Preferably, the precision lifting mechanism includes: Servo motor driven ball screw transmission system with repeatability accuracy of ±0.1mm; Integrated laser rangefinder provides real-time feedback of altitude information; The maximum lifting speed is 0.5m / s and the load capacity is ≥20kg.

[0015] Preferably, the adjustable mass impact clock comprises: Impact head body made of tungsten steel alloy; A detachable counterweight block is connected by a thread to achieve mass adjustment of the adjustable mass impact clock; The impact contact surface of the adjustable mass impact clock is a spherical crown structure.

[0016] In summary, the present invention includes at least one of the following beneficial technical effects: 1. This invention, through dynamic impact loading and fully automated control, radically transforms the traditional static testing model, which relies on manual operation. Multi-band vibration signature analysis, combined with a deep metric learning model, simultaneously extracts multiple spectral and acoustic emission parameters, building a multidimensional evaluation system. This significantly improves the comprehensiveness and scientific nature of testing indicators, breaking through the limitations of traditional single-value quality evaluation.

[0017] 2. This invention incorporates an online autonomous adjustment mechanism that dynamically optimizes impact parameters through reinforcement learning, enabling the testing process to accurately simulate actual wheel loads. Combined with tensor feature space analysis, this effectively captures the evolution of aggregate crushing under varying impact energies, significantly improving the relevance of test results to engineering practice.

[0018] 3. In this invention, the integrated hardware platform implements hardware-based processing of wavelet packet decomposition and tensor operations through an FPGA accelerator, significantly reducing the single detection cycle compared to traditional methods. The feature space mapping capability of the deep metric learning model reduces the discreteness of the results, ensuring high repeatability of the detection data, and solving the industry's difficult problem of balancing detection efficiency and accuracy.

[0019] 4. This invention utilizes a multimodal sensing system that accurately captures millimeter-level micro-crushing characteristics. Combined with an adaptive analysis algorithm, it successfully achieves reliable evaluation of 12 types of special aggregates, including recycled aggregate. The impact device's key component enhancements and energy recovery system enhance its durability, meeting the demands of high-frequency, multi-scenario engineering inspections. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the method flow of the present invention; Figure 2 Schematic diagram of the device structure of the present invention; Among them, 10. Three-axis acceleration sensor and acoustic emission sensor; 20. Test mold; 21. Test mold platform; 22. Fixture; 30. Data acquisition and processing unit; 40. Embedded control module; 50. Precision lifting mechanism; 51. Servo motor; 52. Ball screw; 60. Adjustable mass impact clock. DETAILED DESCRIPTION

[0021] The following is combined with Figure 1 -Attached Figure 2 , the present invention is described in further detail.

[0022] The present invention provides a method for detecting the anti-crushing ability of aggregates based on impact vibration spectrum analysis, such as Figure 1 As shown, the following steps are included: S1: The vibration signal generated by the impact of aggregate is collected through a multimodal sensor array; Specifically, in this embodiment, the multimodal signal acquisition process achieves full-dimensional capture of the aggregate impact response through a precisely designed sensor system. During specimen preparation, the aggregate particles to be tested are loaded into a standard cylindrical test mold 20 (inner diameter 152mm, height 170mm) in accordance with the JTGE42-2005 specification, and the total mass is controlled to meet the following requirements: m s=2000±5g; The layered compaction process is used, with each layer of filler 50mm high and a static pressure of 25kPa maintained for 30 seconds to ensure that the density of the specimen reaches: ρ compact ≥0.95ρmax; where ρ max It is the maximum theoretical density of aggregate, and provides standardized test samples that meet mechanical conduction characteristics for subsequent impact tests.

[0023] During the impact loading stage, the adjustable mass impact bell is released through the electromagnetic drive device (coil rated voltage 24VDC, holding force ≥500N).

[0024] The impact clock performs free fall motion along the precision guide rail (straightness error ≤ 0.02mm / m), and its mass adjustment mechanism includes a basic mass block m b =5kg and two detachable counterweights Δm = 2.5kg, the total mass satisfies: m h =m b +n·Δm(n=0,1,2); The impact contact surface is a spherical crown structure with a radius of R = 50mm, and the surface is hard chrome plated (thickness δ = 50μm, hardness ≥ 800HV) to ensure the stability of impact energy transmission.

[0025] The signal synchronization acquisition system includes a three-axis acceleration sensor unit, a broadband acoustic emission probe, and an environmental monitoring module. The three-axis acceleration sensor (model PCB356A01) is manufactured using MEMS technology, and its technical parameters are as follows: Measuring range: ±500g; Frequency response: 0.5Hz~5kHz(-3dB); Sensitivity: 10mV / g±1.5%; The sensor is installed in the middle of the side wall of the test mold 20 through a magnetic base. The X / Y / Z axes correspond to the radial, tangential and axial directions respectively, and the installation orientation error is controlled within ±1°.

[0026] The acoustic emission sensor (model PACWD) is designed based on the piezoelectric ceramic principle, and its frequency response characteristics meet the following requirements: The sensor is installed on the bottom of the test mold 20 through a coupling agent (silicone grease, viscosity 3500cP). The center frequency matches the characteristic frequency band of aggregate microcrack propagation, which can effectively capture the stress wave signal generated by internal structural damage of the material.

[0027] The time synchronization control system adopts the PTPv2 (IEEE1588-2008) precision clock protocol. The main controller (PHC clock accuracy ±50ns) distributes synchronization signals to each sensor node through the Ethernet switch. Each sensor node has a built-in TCXO temperature compensation crystal oscillator (stability ±1ppm) and synchronizes the clock signal to the sensor node in the synchronization period T. sync = Clock deviation compensation is performed within 1s to ensure that the time alignment accuracy of multi-channel data meets: in and Respectively represent the timestamps of the acoustic emission sensor and the inertial measurement unit at the kth sampling point, N = 10 6 is the number of sampling points for a single impact test.

[0028] The environmental parameter acquisition module integrates a digital temperature and humidity sensor (model SHT35), whose measurement characteristics meet the following requirements: Temperature measurement range: T∈[-40,125]℃ (accuracy ±0.2℃); The sensor probes are arranged at equidistant monitoring points (spacing d = 100 mm) around the test mold 20. The collected environmental data will be used as a compensation factor for subsequent signal processing to eliminate the influence of temperature drift on the sensor output. The temperature compensation algorithm uses: a comp =a raw [1 + α(T - T0)]; Where α = 0.02% / °C is the temperature coefficient of the acceleration sensor, and T0 = 25°C is the reference temperature.

[0029] In the signal preprocessing stage, baseline correction is performed on the original vibration signal to eliminate the DC component interference caused by the sensor zero bias. The correction formula is: where t i =i·Δt, Δt=1μs is the sampling interval. Preferably, the correction process is completed in real time in the pre-processing pipeline embedded in the FPGA, and a sliding window averaging algorithm (window width W=1000 points) is used to improve processing efficiency.

[0030] The sensor network adopts a star topology, with each sensor node connected to the central acquisition unit via a shielded twisted pair cable (characteristic impedance 120Ω, length ≤3m). The signal transmission path is designed to meet the following requirements: Where L and C are the inductance and capacitance per unit length of the line, respectively. This design can effectively suppress waveform distortion caused by signal reflection. Preferably, the accelerometer mounting base is provided with a silicone rubber damping layer (thickness h = 2 mm, loss factor η = 0.15), and its natural frequency is calculated as: Far lower than the main frequency of impact vibration (f impact >100Hz) to avoid contamination of measurement results by installation resonance.

[0031] S2: Use the data acquisition and processing unit to perform dynamic wavelet packet decomposition on the vibration signal and select the optimal decomposition level. Specifically, in this embodiment, the dynamic wavelet packet decomposition process uses multi-level adaptive processing to achieve accurate extraction of impact vibration characteristics. z (t) Preprocessing: First, perform zero-averaging to eliminate the sensor DC offset: A fourth-order Butterworth bandpass filter is then used for band limiting, and its transfer function is: The cutoff frequency is set to f L =10Hz and f H =20kHz, corresponding to angular frequency ω c =2πf, effectively suppressing power frequency interference and high-frequency noise.

[0032] The wavelet basis function is Daubechies8 wavelet (db8), whose scaling function φ(t) and wavelet function ψ(t) satisfy: Filter coefficient {h k} and {g k}Determined by the orthogonality condition, the support length L support =15, vanishing moment order N van =8, which can effectively match the transient characteristics of the impact signal.

[0033] The decomposition layer optimization process performs the following steps: Perform L-layer wavelet packet decomposition on the preprocessed signal to obtain the coefficient d of each node j,n (k), where j∈[1,L] is the number of decomposition layers, n∈[0,2 j -1] is the node index Calculate the energy value of each node: Calculate the entropy of energy distribution: in Construct the objective function and determine the optimal number of decomposition layers: The optimization process is completed in the optimization algorithm module of the embedded processor. The computation time constraint for each decomposition layer is: The characteristic frequency band division is based on the principle of material fracture mechanics, dividing the spectrum into three characteristic intervals: Low frequency band (Ω1:0≤f<500Hz): Characterizes the elastic deformation energy of the aggregate skeleton. The energy integral is calculated as: Mid-frequency band (Ω2: 500≤f<2000Hz): reflects the friction and slip effect between particles. The node selection satisfies: High frequency band (Ω3: f ≥ 2000 Hz): corresponds to microcrack propagation events, and its time series kurtosis is calculated as: Hardware acceleration is implemented using the programmable logic unit of Xilinx FPGA (model XC7Z045) to design a dual-channel parallel processing architecture: Decomposition channel: implements the wavelet packet decomposition filter bank, which contains 16 parallel processing DSP slices, and each slice is configured as follows: Energy calculation channel: Contains 32 multiplication and accumulation units (MACs), which can complete the following in each clock cycle: The data path uses a ping-pong buffer structure, and the input buffer depth is configured as: Ensure real-time processing throughput reaches 1GSamples / s.

[0034] The dynamic parameter update mechanism is triggered by monitoring the ply selection results of three consecutive impact tests: Trigger conditions: When the conditions are met, the wavelet basis reselection process is started to calculate the similarity between each candidate wavelet basis and the reference signal: Choose a value that satisfies ρ m The wavelet basis with a coefficient greater than 0.9 and the lowest computational complexity is used as the updated decomposition function.

[0035] This implementation method fully discloses the technical details of the entire wavelet packet decomposition process, and the parameters of each formula are strictly consistent with the aforementioned claims and implementation methods, ensuring the feasibility of the technical solution and the supportability of the claims.

[0036] S3: Data acquisition and processing unit 30 constructs a three-dimensional time-frequency-space tensor feature space based on the dynamic wavelet packet decomposition results. In this embodiment, the construction of the three-dimensional tensor feature space is achieved through multi-dimensional information fusion and high-order tensor decomposition technology. Based on the wavelet packet node energy distribution obtained in step S2, the time-frequency characteristics, spatial sensor information, and physical response characteristics are multi-dimensionally coupled to form a tensor representation structure with clear engineering significance.

[0037] Original tensor construction: Define the physical meaning and calculation rules of each dimension in the three-dimensional feature space: Band dimension: corresponds to the characteristic band division after wavelet packet decomposition. The dimension size is determined by the optimal decomposition layer number L determined in step S2. opt Decide: Time dimension: Split the signal according to a fixed time window ΔT = 50ms. The dimension size is: Spatial dimension: Integrates multi-sensor channel data, including three-axis acceleration (X / Y / Z), acoustic emission, and environmental parameters. The dimension size is fixed as follows: C=3 acc ×3 pos +2 AE +1 env =9; Original tensor The construction process meets the following requirements: in Indicates the wavelet packet reconstructed signal of the kth frequency band at t i The amplitude at the moment, S c (t i ) is the normalized output value of the c-th sensor channel; Normalization processing uses: Tensor dimensionality reduction processing: Using the Tucker decomposition algorithm, the original tensor is decomposed into the product of the core tensor and the factor matrix: The core tensor The dimension of is determined by the energy accumulation criterion: In the formula is the n-th mode expansion matrix W (n) The i-th singular value of is set in practical applications: R1=32, R2=16, R3=9; Factor matrix calculation: The high-order orthogonal iteration (HOOI) algorithm is used, and the specific iteration process is as follows: Initialization: Generate a random orthogonal matrix Modal Update: a. Calculate the mode-1 expansion matrix: b. To W (1)Perform truncated SVD decomposition: Convergence judgment: Repeat step 2 until the objective function changes to meet the following conditions: Hardware acceleration implementation: A tensor operation acceleration engine is configured in Xilinx FPGA (model XC7VX690T), which includes a three-stage processing pipeline: Data reorganization level: Input data flow rate: 1GB / s Cache depth: D buffer =K×M×C×4bytes=32×16×9×4=18KB; Matrix expansion level: Parallel computing unit: 16 DSPSlices; Calculation mode: where j = ∏ k≠n i k ; SVD acceleration stage: Adopting bilateral Jacobi rotation algorithm and equipped with 8 parallel processing units; Iterations: Dynamic update mechanism: When the wavelet decomposition layer number L is detected opt When changes occur, the following refactoring process is triggered: Band dimension update: Incremental SVD updates the factor matrix in To add a new frequency band data matrix, the orthogonalization process uses the Gram-Schmidt algorithm:

[0038] S4: The embedded control module 40 inputs the three-dimensional tensor into the deep metric learning model to evaluate the anti-crushing capability; In this embodiment, crushing resistance assessment is implemented using a deep metric learning model. This model transforms high-dimensional tensor features into interpretable scores based on the principle of feature space mapping. Using the reduced-dimensional kernel tensor obtained in step S3 as input, a feature extraction network and metric learning mechanism are used to establish a nonlinear mapping relationship between aggregate impact response characteristics and crushing resistance.

[0039] The network architecture design adopts a triplet network structure, which includes a feature extraction branch with shared weights and a metric calculation module. The feature extraction branch consists of three levels of 3D convolution modules, each of which contains: Convolutional layer: kernel size 3×3×3, stride 2×2×2, and the number of output channels is 64, 128, and 256 respectively; Batch Normalization layer: performs normalization on the channel dimension: Activation function: ParametricReLU, the negative slope is initialized to 0.25; The metric learning mechanism adopts the contrastive loss function and optimizes the feature space distribution through training with triplets of anchor samples, positive samples, and negative samples.

[0040] The loss function is defined as: Where α = 0.5 is the marginal hyperparameter, f a ,f p ,f n 128-dimensional feature vectors representing anchor points, positive examples, and negative examples respectively. Preferably, the triplet sample selection strategy adopts semi-hard mining to ensure: ||f a -f p || 2 <||f a -f n || 2 <||f a -f p || 2 +α; The scoring calculation module maps the feature vector to the standard scoring interval and uses the Sigmoid function to implement nonlinear transformation: The weight parameter w and bias b are obtained through supervised learning, and the mean square error loss is used during training: in It is a reference value of anti-crushing ability obtained based on standard test methods.

[0041] The model training process adopts a two-stage optimization strategy: Pre-training phase: train the feature extraction network on a large aggregate dataset, with the optimizer configured as Adam (β1 = 0.9, β2 = 0.999) and the initial learning rate η = 10 -3 ; Fine-tuning stage: Freeze the parameters of the first two convolutional layers, use stochastic gradient descent with momentum (μ = 0.9), and the learning rate decay strategy is: Hardware acceleration enables the deployment of quantized models in embedded modules. Key optimizations include: Weight 8-bit fixed-point quantization: Activation value dynamic range calibration: using moving average statistical maximum value Parallel computing unit: Equipped with four ARM NEON SIMD engines, a single instruction cycle completes eight groups of 16-bit multiplication and addition operations. Dynamic update mechanism continuously optimizes model parameters based on online collected data: Incremental data cache: retain the latest 100 sets of impact test data; Parameter fine-tuning: Perform online learning every 24 hours to update the fully connected layer parameters Model validation: When the validation set error increases three times in a row, roll back to the previous stable version.

[0042] S5: The embedded control module 40 drives the adjustable mass impact clock 60 and the precision lifting mechanism 50 to adjust the impact parameters according to the evaluation results; This embodiment provides a method for adaptively adjusting impact parameters based on reinforcement learning. This method is encapsulated and executed within the embedded control module, and is designed to respond to the crushing resistance assessment result output in step S4 and perform closed-loop, intelligent adjustment of subsequent impact test parameters.

[0043] The core of this method lies in modeling the optimization process of impact parameters as a reinforcement learning problem. To achieve this, it is first necessary to establish a state space that accurately describes the instantaneous state of the impact event. This state space S is defined as a vector containing 12 key features, which comprehensively characterize the system state and response characteristics of the aggregate before and after the impact.

[0044] For example, after the tth impact is completed, the constructed state vector S t May include: the mass parameter m of the last impact t-1 , the height parameter h of the last impact t-1 , and the anti-crushing ability score Q obtained after the current impact t .

[0045] Furthermore, the state vector S t It also includes multiple physical indicators calculated by the data acquisition and processing unit. Specifically, it includes: spectrum gradient that characterizes the change of spectrum morphology The root mean square value of acoustic emission σ, which represents the intensity of microcrack activity inside the aggregate ae ; and apparent energy expenditure E c .

[0046] In order to form a complete 12-dimensional vector, other spectral features can also be included, such as the spectral centroid, peak frequency, and the energy proportion of different frequency bands, which together constitute a complete digital description of the impact event.

[0047] Corresponding to the state space, a discretized action space A is defined, which contains all possible combinations of commands that the embedded control module can issue to the actuator. This action space A is the Cartesian product of all possible combinations of mass adjustment Δm and impact height adjustment Δh.

[0048] Specifically, the action space can be expressed as A = {-5kg, 0kg, +5kg} × {-0.2m, 0m, +0.2m}. Under this definition, there are 9 discrete compound actions. For example, a t =(+5kg,-0.2m) represents a specific adjustment strategy.

[0049] When the embedded control module selects this action a t After that, it generates the corresponding control instructions. On the one hand, it sends a command to the electromagnetic lock of the adjustable mass impact clock to add a 5kg counterweight. On the other hand, it drives the servo motor in the precision lifting mechanism to rotate in the opposite direction, and the ball screw is used to accurately lower the impact head by 0.2 meters.

[0050] In order to drive the model to learn the optimal adjustment strategy, this implementation design a sophisticated reward function r t This function is calculated after each adjustment action is executed and the impact is completed, and is used to quantitatively evaluate the quality of the decision. The design goal of this reward function is to balance the stability of the detection results and the economic efficiency of the detection process.

[0051] The reward function r t The specific mathematical expression is as follows: Among them, r t It is the immediate reward value fed back by the system environment after executing the t-th adjustment action; Q t is the aggregate anti-crushing ability score evaluated by the deep metric learning model after the t-th impact, and its value range is [0,1]; It is the crushing resistance score evaluated after the t-1th impact.

[0052] E t is the total energy consumed by the system during the t-th impact, which can be expressed by the potential energy change ΔE of the impact clock. p =m t ·g·ht And it is calculated in combination with the energy loss coefficient.

[0053] It is a positive incentive term related to the stability of the result. t and When the difference is close to the same, the absolute value of the difference Approaching 0, the output of this term approaches its maximum value of 10. This design can guide the model to find a stable impact energy region that can produce highly reproducible results.

[0054] Second item -0.1E t This is a negative penalty term related to energy consumption. It ensures that, while maintaining a stable score, the system favors impact parameters that consume less energy. This not only saves energy but, more importantly, prevents excessive aggregate crushing caused by applying impact energy far exceeding the necessary limit, which would render the test incapable of distinguishing true resistance to crushing.

[0055] In a complete technology chain demonstration, the implementation process of this method is as follows. First, the system performs a preset initial impact (t = 1), for example, using a 10 kg mass and a height of 0.5 m. Through steps S1 to S4, partial information about the initial state is obtained, including the score Q1 and the energy E1. Subsequently, when deciding the parameters for the next impact (t = 2), the reinforcement learning algorithm within the embedded control module (for example, a pre-trained or real-time learning Q-learning or DQN model) is activated. It first fully constructs the current state vector S2 based on the results of the first impact.

[0056] The algorithm then evaluates all nine actions in the action space A using its internal Q-value table or neural network, predicting the long-term cumulative reward for each action. Using a strategy (such as the ε-greedy strategy), it selects an optimal or exploratory action, such as a2 = (0 kg, +0.2 m).

[0057] This command is issued, and the precision lifting mechanism raises the impact height to 0.7m, while maintaining the mass at 10kg. After the second impact, the system obtains a new score Q2 and energy E2, and calculates the actual reward r2 according to the formula. This four-tuple (S2, a2, r2, S3) is used to update the Q-value table or train the neural network to optimize its decision-making capabilities.

[0058] This process is repeated and executed repeatedly. As the number of iterations increases, the fluctuation of the score|Q t -Q t-1| will gradually decrease and converge. When the value is less than a preset threshold (for example, 0.01) for several consecutive times, the system determines that the optimal detection parameter has been found and the score at this time is Q t The final aggregate anti-crushing ability test result is output. Through closed-loop feedback and adaptive adjustment, this method can achieve automated, high-precision and high-efficiency testing of different types of aggregates.

[0059] S6: Executes the signal processing and model calculations from S2 to S4 in real time through the hardware acceleration module.

[0060] In this embodiment, the hardware acceleration module utilizes a heterogeneous computing architecture to achieve real-time requirements for signal processing and model calculations, mapping the key algorithms in steps S2 through S4 to dedicated hardware logic units. This module employs dynamic pipeline reorganization technology to adaptively adjust computing resource allocation based on the processing stage, ensuring full data processing within strict timing constraints.

[0061] Preprocessing acceleration unit: For the wavelet packet decomposition algorithm in step S2, a three-level parallel processing engine is designed: Filterbank Engine: Implement a dual-channel filter based on the db8 wavelet basis function and decompose the filter coefficients Stored in on-chip BRAM, the bit width is configured as 18-bit fixed-point number (Q3.15 format), meeting the dynamic range [-4,4) and precision 2 -15 Require The parallel computing unit implements the following convolution operation: The data path uses a ping-pong buffer structure with a buffer depth of Match the maximum number of decomposition levels required.

[0062] Energy calculation engine: Configure 32 parallel multiply-accumulate (MAC) units to complete the following operations per cycle: The accumulator bit width is extended to 48 bits (Q24.24 format) to prevent M=10 6 The accumulation overflowed.

[0063] Layer Optimization Engine: The entropy calculation uses a pre-generated lookup table (LUT) to implement fast plog2p calculation. The LUT depth is 1024 and the input is quantized to 10 bits: Objective function real-time comparator array implementation: J(L)=H(L)+0.15·2 L(L=2,…,8); The comparators are configured as 8 parallel paths, each path containing an 18-bit adder and comparison logic.

[0064] Tensor operation acceleration unit: Based on the Tucker decomposition of step S3, a reconfigurable tensor processing engine is designed: Modal expansion module: Implement dimensional reorganization of tensors to matrices, supporting arbitrary modal expansion (n=1, 2, 3); The data reordering logic adopts a crossbar switch architecture with a switching delay of <10ns, meeting real-time requirements; The address generation logic is: addr(i1,i2,i3)=i n ×∏ k≠n I k +∑ m≠n i m ×∏ p>m I p ; Matrix multiplication acceleration module: Configured with 4 systolic arrays, each containing 64 processing elements (PEs) Single PE implementation: C out =C in +A reg ×B reg ; Among them A reg With B reg It is a pipeline register with a width of 18 bits; Support for block matrix multiplication: SVD acceleration module: Adopting bilateral Jacobi rotation algorithm, equipped with 8 rotation angle calculation units; Single rotation calculation: Where a and d are matrix diagonal elements, and b is the sub-diagonal element; Iteration control: fixed number of iterations N iter =10, error threshold∈=10 -4 Model inference acceleration unit: For the deep metric learning model in step S4, a dedicated neural network processor is designed: 3D convolution engine: Supports dynamic core size configuration (1×1×1 to 7×7×7) through programmable logic cells.

[0065] Input feature map slice caching strategy: Among them F conv is the convolution kernel size, and the cache depth matches the number of input channels C in =256 Fully connected accelerator: Using a bit-serial computing architecture, weights are preloaded into register chains with a bit width of 8 bits (Q2.6 format); nonlinear mapping of 128-dimensional feature vectors is achieved: The negative interval slope a is stored in LUT with a depth of 256 and a precision of 2 -8 ; Rating calculation unit: The hardware Sigmoid function uses piecewise linear approximation to divide the interval: Coefficient selection was based on least squares fitting with a maximum error of <0.5%.

[0066] Dynamic pipeline control: Resource allocation strategy: Establish a hardware resource pool, including configurable logic units (CLBs), DSP slices, and BRAMs; dynamically reorganize them according to the processing stage at runtime: Preprocessing phase Tensor decomposition phase Model inference phase The allocation ratio is based on the demand characteristics of computing, storage, and communication resources at each stage Data flow control: Adopting the credit token flow control mechanism, the token pool depth is the pipeline level N stage =8 Transmission bandwidth matching constraints: B in ≥max(B proc ,B out ) where Φ roc =f clk ×W data =200MHz×64bit=12.8GB / s; Abnormal recovery mechanism: Configure the watchdog timer to monitor pipeline stalls and timeout thresholds: T timeout =5×T max_stage =5×20μs=100μs; The reset logic performs state machine reinitialization: State←IDLE,Counters←0,Buffers←Flush.

[0067] Please see the attached Figure 2 The present invention also provides an aggregate anti-crushing ability detection device based on impact vibration spectrum analysis, comprising: A multimodal sensor array, including a triaxial acceleration sensor and an acoustic emission sensor 10, is arranged in a ring around a test mold 20 and a test mold platform 21, and a fixture 22 is provided in the test mold 20; The adjustable mass impact clock 60 is connected to the precision lifting mechanism 50 through an electromagnetic lock and provides three impact masses of 5kg, 10kg, and 15kg; The data acquisition and processing unit 30 is connected to the sensor array and includes a dynamic wavelet packet decomposition module and an FPGA hardware accelerator. The embedded control module 40 is connected to the data acquisition and processing unit 30 and the precision lifting mechanism 50 and has a built-in deep metric learning model. The precision lifting mechanism 50 drives the ball screw 52 through a servo motor 51 to achieve height adjustment, and the FPGA hardware accelerator is configured with a dedicated logic unit for tensor operations.

[0068] Specifically, the servo motor 51 is rigidly connected to the ball screw 52 via a coupling, and the screw nut is fixed to the impact clock lifting frame. An electromagnetic lock is integrated into the top of the lifting frame and is driven by a 24V DC power supply. A laser rangefinder is mounted on the side of the lifting frame, and its analog output is connected to the servo driver feedback interface.

[0069] The embedded control module 40 sends a height command to the servo driver, driving the ball screw 52 to lift the impact bell to the set position (adjustable from 0.5 to 1.5 m); The laser rangefinder monitors the height deviation in real time and uses the PID controller (proportional coefficient K p =2.5, integration time T i =0.1s) Dynamically correct the lifting position; After receiving the trigger signal, the electromagnetic lock releases the impact bell, which freely falls and impacts the aggregate sample in the test mold 20; After the impact is completed, the servo motor 51 rotates in the reverse direction to retract the impact bell to the initial position.

[0070] The triaxial accelerometer is fixed to the edge of the test mold platform 21 via an M5 connector, and its signal line is connected to the data acquisition box via a metal hose. The acoustic emission sensor is mounted on the side wall of the test mold 20 via a magnetic base. Its output is connected to a charge amplifier and then to an acquisition card. All sensor signal lines are connected to the acquisition box in a star topology, with the ground terminals connected to a copper busbar.

[0071] At the moment of impact, the acceleration sensor detects the three-dimensional vibration signal (range ±500g), and the acoustic emission sensor captures the stress wave signal (width 50-400kHz); The signal is pre-processed by a charge amplifier (gain 60dB) and an anti-aliasing filter (cutoff frequency 200kHz) and then synchronously sampled by a 24-bit ADC at a rate of 256kSPS. The digital signal is transmitted to the FPGA hardware accelerator through the PCIe interface, and the original data is simultaneously backed up to the SSD storage (1TB capacity).

[0072] The FPGA connects to the embedded control module via a high-speed serial transceiver (GTY channel), forming a bidirectional data path. A DMA engine within the FPGA streams preprocessed data to the tensor arithmetic unit. The embedded module communicates control signals with the servo drive and electromagnetic latch via the EtherCAT bus.

[0073] Wavelet packet decomposition: After receiving the original signal, the FPGA starts the filter bank engine to calculate the energy of each node in parallel; Dynamically adjust the number of decomposition layers L opt (3-8 layers), the optimal decomposition structure is selected through the entropy comparison module.

[0074] Tensor construction and decomposition: Reorganize the time-frequency energy, spatial position, and sensor channel data into a 32×16×9 three-dimensional tensor; Call Tucker to decompose the hardware logic and iteratively calculate the factor matrix U (1) ,U (2) ,U (3) Until convergence.

[0075] Model Inference: Core tensor after dimensionality reduction Transmitted to the embedded module via the PCIe interface; The deep metric learning model performs three levels of 3D convolution (kernel size 3×3×3) and outputs a 128-dimensional feature vector; The fully connected layer is combined with the Sigmoid function to generate the anti-crushing ability score Q∈[0,1].

[0076] Closed-loop feedback and parameter optimization: Component connections: The embedded module's GPIO interface connects to the abnormality indicator and emergency stop relay, and the CAN bus connects to the sensor self-test module. Model parameters are stored in the eMMC flash memory, which supports hot-swappable replacement.

[0077] Dynamic feedback: The short-term channel calculates the standard deviation of three consecutive scores σQ , if the threshold value (0.05) is exceeded, the signal acquisition recalibration is triggered; The moving average deviation ΔQ of the scores of the long-term channel evaluation is 100 times. If it continues to exceed the limit (±0.1), the model retraining will be initiated.

[0078] Parameter adjustment: Update the number of wavelet packet decomposition layers L through the FPGA dynamic reconfiguration interface opt and filter coefficients; Adjust the tensor decomposition energy threshold γ n (0.9-0.98) and model marginal parameter α (0.3-0.7).

[0079] Exception handling: Primary abnormality: Activate the sensor self-test program and verify the signal correlation ρ c >0.9; Intermediate exception: Freeze convolutional layer parameters, incrementally update fully connected layer weights (learning rate 10 -3 ); Serious abnormality: Cut off the servo motor power supply, activate the mechanical brake, and restore the factory configuration parameters.

[0080] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting aggregate anti-crushing ability based on impact vibration spectrum analysis includes the following steps: S1: The vibration signal generated by the impact of aggregate is collected through a multimodal sensor array; S2: using a data acquisition and processing unit to perform dynamic wavelet packet decomposition on the vibration signal and select an optimal decomposition layer number; S3: The data acquisition and processing unit constructs a time-frequency-space three-dimensional tensor feature space based on the dynamic wavelet packet decomposition result; S4: The embedded control module inputs the three-dimensional tensor into a deep metric learning model to evaluate the anti-crushing capability; S5: The embedded control module drives the adjustable mass impact clock and the precision lifting mechanism to adjust the impact parameters according to the evaluation results; S6: Executes the signal processing and model calculations from S2 to S4 in real time through the hardware acceleration module.

2. The method for detecting aggregate anti-crushing ability based on impact vibration spectrum analysis according to claim 1 is characterized in that: Said S1 comprises: Collect three-dimensional vibration signals through a three-axis acceleration sensor; The stress wave signal in the frequency band of 100-300kHz is captured by the acoustic emission sensor.

3. The method for detecting aggregate anti-crushing ability based on impact vibration spectrum analysis according to claim 1 is characterized in that: The dynamic wavelet packet decomposition in S2 includes: Calculate the energy entropy value of each decomposition layer node H(L) = -∑p n logp n ,in By minimizing the objective function H(L) + 0.15·2 L Select the optimal number of decomposition levels L opt .

4. The method for detecting aggregate anti-crushing ability based on impact vibration spectrum analysis according to claim 1 is characterized in that: The S3 includes: Arrange the multi-channel time-frequency data into original tensors in three dimensions according to frequency bands, time windows, and sensor channels; Tucker decomposition is used to reduce the original tensor dimension to a core tensor of 32×16×9; Retain more than 95% of the energy information of the original tensor.

5. The method for detecting aggregate anti-crushing ability based on impact vibration spectrum analysis according to claim 1 is characterized in that: The deep metric learning model in S4 includes: A triplet network structure is used to extract 128-dimensional feature vectors; Training was performed using a contrastive loss function with a margin of α = 0.5; The features are mapped to a crush resistance score in the interval [0, 1] using the Sigmoid function.

6. The method for detecting aggregate anti-crushing ability based on impact vibration spectrum analysis according to claim 1 is characterized in that: The S5 includes: Construct a spectral gradient Acoustic emission standard deviation σ ae Energy consumption E c 12-dimensional state space; Define the discrete action space {-5,0,+5}kg×{-0.2,0,+0.2}m; Designing the reward function Among them, Q t is the aggregate anti-crushing ability score evaluated by the deep metric learning model after the t-th impact; is the crushing resistance score evaluated after the t-1th impact; E t is the total energy consumed by the system during the tth impact.

7. The method for detecting aggregate anti-crushing ability based on impact vibration spectrum analysis according to claim 1 is characterized in that: The hardware acceleration module in S6 achieves the real-time requirements of signal processing and model calculation through a heterogeneous computing architecture, mapping the key algorithms of steps S2 to S4 into dedicated hardware logic units; The hardware acceleration module adopts dynamic pipeline reorganization technology and can adaptively adjust the allocation of computing resources according to the processing stage.

8. An aggregate anti-crushing capacity detection device based on impact vibration spectrum analysis, applied to the aggregate anti-crushing capacity detection method based on impact vibration spectrum analysis according to any one of claims 1 to 7, characterized in that: include: A multimodal sensor array, including a triaxial acceleration sensor and an acoustic emission sensor, is arranged in a ring around the test mold and the test mold platform, and a fixture is set inside the test mold; An adjustable mass impact clock is connected to a precision lifting mechanism via an electromagnetic lock, providing three impact masses: 5kg, 10kg, and 15kg. A data acquisition and processing unit is connected to the sensor array and includes a dynamic wavelet packet decomposition module and an FPGA hardware accelerator. An embedded control module is connected to the data acquisition and processing unit and the precision lifting mechanism and includes a built-in deep metric learning model. The precision lifting mechanism realizes height adjustment by driving a ball screw with a servo motor, and the FPGA hardware accelerator is configured with a dedicated logic unit for tensor operations.

9. The device for detecting the anti-crushing ability of aggregates based on impact vibration spectrum analysis according to claim 8, characterized in that: The precision lifting mechanism comprises: Servo motor driven ball screw transmission system; Integrated laser rangefinder provides real-time feedback of altitude information; The maximum lifting speed is 0.5m / s and the load capacity is ≥20kg.

10. The device for detecting the anti-crushing ability of aggregates based on impact vibration spectrum analysis according to claim 8, characterized in that: The adjustable mass impact clock comprises: Impact head body made of tungsten steel alloy; A detachable counterweight block is connected by a thread to achieve mass adjustment of the adjustable mass impact clock; The impact contact surface of the adjustable mass impact clock is a spherical crown structure.

Images