An adaptive control method and equipment for crushing construction solid waste

Abstract

This invention relates to the field of automatic control and processing technology, specifically to an adaptive control method for crushing construction solid waste. The method includes real-time acquisition of particle size, rebar ratio, moisture content, strength, and feed rate via a multi-source sensor array; generation of operating condition vectors via a cross-modal fusion network; and training of a convolutional neural network to output rotor speed, cavity clearance, hydraulic pressure, screen aperture size, and shape memory alloy deformation. This application also relates to adaptive control equipment for crushing construction solid waste. The controller of this application synchronously drives a variable frequency motor, hydraulic cylinder, variable screen, and Ni-Ti shape memory alloy actuator, completing reversible geometric adjustment within 10 seconds. This enables rapid switching between coarse and fine crushing, incremental updates via online particle size detection and energy consumption metering feedback network, forming a perception-decision-execution closed loop. This significantly reduces energy consumption, substantially improves particle size qualification rate, extends liner life, reduces the need for manual intervention, enhances overall operational stability and intelligence, and reduces dust emissions.

Description

Technical Field

[0001] This invention relates to the field of automatic control and processing technology, specifically to an adaptive control method and equipment for crushing construction solid waste. Background Technology

[0002] With the rapid advancement of urbanization and the renovation of existing buildings, large-scale demolition, urban renewal, and reinforcement of dilapidated buildings have generated massive amounts of construction and solid waste (C&DWaste). According to statistics from the Ministry of Housing and Urban-Rural Development, the annual C&DWaste production in China has exceeded 3 billion tons, accounting for more than 40% of the total urban solid waste at its peak. Construction solid waste has a complex composition, containing not only high-strength concrete blocks and steel reinforcement frames, but also bricks, tiles, wood, insulation layers, and small amounts of hazardous substances; its particle size, moisture content, and steel reinforcement content vary significantly depending on construction methods and regional differences.

[0003] Traditional stationary or semi-mobile crushing lines generally employ static processes, namely belt feeding, jaw crushing for coarse crushing, impact or hammer crushing for medium and fine crushing, and vibrating screen screening, supplemented by simple iron removal. On-site, rotor linear speed, cavity clearance, and screen aperture size are typically set manually based on experience, and a constant load is maintained via PLC or frequency converter. When raw material fluctuations or particle size requirements change, operators must stop the machine or reduce capacity for manual adjustment, resulting in high overall energy consumption, increased over-crushing rate, and severe wear on equipment liners.

[0004] While existing academic literature proposes fuzzy control, PID self-tuning, or traditional BP neural network prediction based on single vibration or current signals, most are limited to offline training and univariate feedback, failing to depict the multidimensional heterogeneity of construction solid waste in real time. Furthermore, current crushing chamber geometry parameters largely rely on slow adjustments by hydraulic cylinders, making it difficult to quickly switch between coarse and fine crushing modes under load fluctuations, easily leading to chamber blockage and mechanical impact. In addition, while multi-source sensing technologies such as 3D laser and near-infrared hyperspectral imaging have been used in mine sorting and waste sorting, they have not yet been deeply integrated with construction solid waste crushing control. Moreover, there is a lack of a comprehensive solution that directly embeds real-time particle size detection results into neural network activation functions and combines shape memory alloys to achieve second-level chamber self-reconstruction. Therefore, there is an urgent need for an adaptive control method and equipment for construction solid waste crushing that combines multi-source sensing, intelligent decision-making, and fast-response execution to reduce unit energy consumption, improve product particle size stability, and extend the life of vulnerable parts, meeting the pressing needs of green construction and resource recycling. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention proposes an adaptive control method and equipment for crushing construction solid waste. It utilizes a multi-source sensor array to collect real-time data on particle size, rebar ratio, moisture content, strength, and feed rate; a cross-modal fusion network generates operating condition vectors; and a convolutional neural network is trained to output rotor speed, cavity clearance, hydraulic pressure, screen aperture size, and shape memory alloy deformation. The controller synchronously drives a variable frequency motor, hydraulic cylinder, variable screen, and Ni-Ti shape memory alloy actuator, enabling reversible geometric adjustments of at least 3mm within 10 seconds, achieving rapid switching between coarse and fine crushing. Online particle size detection and energy consumption metering results are fed back to the network for incremental updates, forming a perception-decision-execution closed loop. Compared to traditional fixed-parameter crushing, energy consumption is reduced by 15-25%, particle size qualification rate is increased by 10%, and liner life is extended by 30%. It is suitable for both mobile and fixed construction solid waste resource recovery lines, reducing the need for manual intervention, improving overall operational stability and intelligence, and reducing dust emissions.

[0006] This application provides an adaptive control method for crushing construction solid waste, including the following steps:

[0007] S1: Real-time data collection of construction solid waste entering the crusher is achieved using a multi-source sensor array. The construction solid waste data includes particle size, steel reinforcement content, moisture content, strength, and feeding rate of the construction solid waste.

[0008] S2: Input construction solid waste data into the fusion network to calculate the working condition vector;

[0009] S3: Input the working condition vector into the trained convolutional neural network model, and the convolutional neural network model outputs the set of control parameters for the crusher;

[0010] S4: Synchronously adjust the rotor linear speed, crushing chamber gap, hydraulic pressure and screen opening size of the crusher according to the control parameter set, and generate reversible deformation of the crushing chamber liner and / or hammer within 10 seconds through the shape memory alloy actuator, with a deformation of not less than 3mm, so as to dynamically reconstruct the geometric parameters of the crushing chamber to match the real-time crushing requirements.

[0011] S5: Real-time detection of particle size distribution of the crushed product, and feedback of the detection results to the trained convolutional neural network model for online weight update, realizing closed-loop adaptive optimization.

[0012] Preferably, the multi-source sensor array includes: a three-dimensional laser scanner for acquiring particle size distribution, an eddy current sensor for estimating the steel reinforcement content, a near-infrared hyperspectral camera for determining moisture content, an impact acoustic emission or hardness sensor for inferring strength, and a weighing and metering module for obtaining the feed rate.

[0013] Preferably, the fusion network is a lightweight spatiotemporal multimodal neural network, which adopts cross-modal cross-attention and channel gating mechanisms, and uses pruning and quantization techniques to ensure that the computational cost of a single inference does not exceed 5 GMAC.

[0014] Preferably, the trained convolutional neural network model includes a deep feature extraction backbone network composed of residual blocks and compressed excitation units, and adopts an incremental learning and sample replay mechanism during online weight updates to avoid catastrophic forgetting.

[0015] Preferably, the shape memory alloy actuator uses Ni-Ti alloy wire with an austenite termination temperature of 60℃-90℃. Driven by a resistance heating pulsed current, it completes a reversible deformation of at least 3mm within 10 seconds, and has a cycle life of at least 10 cycles. 6 Second-rate.

[0016] Preferably, the convolutional neural network model employs an improved activation function during the inference process in step S3 and the online weight update process in step S5. :

[0017]

[0018]

[0019] in, For the output of the convolutional layer, This is the granularity distribution prediction vector of the output of the previous layer in the convolutional neural network. This is the real-time detected particle size distribution vector of the crushed product. Represents the dot product. Represents the 2-norm; For the reason and The adaptive coefficients obtained by the cosine similarity linear mapping are dynamically adjusted with the granularity prediction error to enhance the network's adaptability to granularity distribution fluctuations and provide targeted gradient feedback during online weight updates.

[0020] Preferably, the step of inputting construction solid waste data into a fusion network to calculate the working condition vector is as follows:

[0021] S21: The multimodal coding layer is configured with a parallel coding branch for the sensor data of particle size, steel reinforcement content, moisture content, strength and feed rate. The coding branch uses a one-dimensional convolutional residual unit for time-series signals and a two-dimensional convolutional residual unit for image and spectral signals, and outputs a feature tensor of the same channel dimension.

[0022] S22: The cross-modal attention layer uses the dot product attention mechanism to calculate the correlation weights of the feature tensors of any two modalities and generates weighted fusion features accordingly.

[0023] S23: The channel gating layer applies Squeeze-and-Excitation gating operations to the weighted fusion features to compress redundant channels;

[0024] S24: The global pooling layer performs global average pooling on the gated feature map to obtain a working condition vector with an adjustable length of 64-256 dimensions, which is used to characterize the comprehensive working condition of construction solid waste.

[0025] Preferably, the control parameter set includes: target linear velocity of the crusher rotor, target discharge gap of the crushing chamber, target working pressure of the hydraulic system, target opening size of the screen, and target reversible deformation of the shape memory alloy actuator; the controller sends the target linear velocity of the crusher rotor to the frequency conversion drive module, and uses a linear acceleration and deceleration ramp to make the rotor reach the set linear velocity; it sends a displacement command to the hydraulic actuator to adjust the relative position of the crushing chamber liner to the target discharge gap of the crushing chamber; it controls the hydraulic pressure to the target working pressure of the hydraulic system through a proportional valve and a servo pump, and compensates for pressure drift caused by material hardness fluctuations in real time through closed-loop control; it adjusts the opening size to the target opening size of the screen by driving the electric slide rail or telescopic screen module; and it applies a pulse current to the shape memory alloy actuator to complete the reversible deformation, so as to dynamically reconstruct the geometric parameters of the crushing chamber.

[0026] Preferably, the weighted fusion features generated accordingly include:

[0027] Perform a linear projection on the first modality feature tensor to obtain the query vector Q;

[0028] Perform linear projections on the second modality feature tensor to obtain the key vector K and the value vector V;

[0029] The dot product of Q and K is calculated and normalized by the square root of the vector dimension before being fed into the Softmax function to obtain the attention weight matrix A;

[0030] Multiply the attention weight matrix A by the value vector V to obtain the cross-modal context vector;

[0031] The cross-modal context vector and the first modal feature tensor are added together with residual weights according to learnable gating coefficients to obtain weighted fusion features, and the weighted fusion features are provided to the global pooling layer to generate the working condition vector.

[0032] This application also provides an adaptive control device for crushing construction solid waste, comprising:

[0033] a) Feeding mechanism, used to transport construction solid waste to the crusher;

[0034] b) A multi-source sensor array, set at the feeding mechanism or the feed inlet of the crusher, including a 3D laser scanner, an electromagnetic induction or eddy current sensor, a near-infrared spectral camera, an impact acoustic emission or hardness sensor, and a weighing and metering module, is used to collect the particle size, steel reinforcement content, moisture content, strength and feeding rate of construction solid waste in real time.

[0035] c) A controller with a built-in multimodal fusion network for calculating and generating a working condition vector from the data output by the multi-source sensor array; a trained convolutional neural network model for outputting a set of crusher control parameters based on the working condition vector; and an online weight update unit for incrementally updating the convolutional neural network model after receiving the product particle size distribution detection result.

[0036] d) A crusher whose structure allows for adjustment of the rotor linear speed and the crushing chamber gap;

[0037] e) A variable frequency drive module, connected to the crusher and adjusting the rotor linear speed according to the control parameter set;

[0038] f) A hydraulic actuator module, connected to the crusher liner, adjusts the crushing chamber clearance and hydraulic pressure according to the set of control parameters;

[0039] g) Shape memory alloy actuators, arranged in the crushing chamber liner and / or hammer, configured to generate reversible deformation to dynamically reconstruct the crushing chamber geometry.

[0040] h) A screening module having an adjustable screen opening size and performing adjustments according to the set of control parameters;

[0041] i) Output detection module, including online particle size detector and energy consumption metering device, is used to acquire the particle size distribution and unit energy consumption of the crushed product in real time, and send the detection results to the controller;

[0042] The controller is configured to execute the adaptive control process of construction solid waste crushing according to the method, so as to realize the synchronous linkage adjustment of the crusher, hydraulic actuator, shape memory alloy driver and screening module, and complete the closed-loop adaptive optimization.

[0043] This invention provides an adaptive control method and equipment for crushing construction solid waste, which achieves the following beneficial technical effects:

[0044] 1. This application employs multi-source sensing-convolutional neural network closed-loop control and rapid geometric self-reconstruction of shape memory alloys to ensure the crusher always operates under optimal conditions, significantly reducing unit energy consumption, peak current, and power distribution capacity. Real-time online particle size detection results directly participate in activation function adjustment and weight increment learning, ensuring the crushed product particle size stably falls within the target range. This significantly improves the pass rate compared to traditional manual adjustment methods, reducing over-crushing and material return.

[0045] 2. The Ni-Ti shape memory alloy actuator of this application can achieve reversible deformation of not less than 3 mm within 10 seconds, and can switch between "coarse crushing" and "fine crushing" modes without stopping the machine; compared with independent adjustment by hydraulic cylinder, the switching time is significantly shortened, significantly reducing the risk of cavity blockage. By dynamically adjusting the rotor linear speed and cavity clearance as needed, excessive impact and ineffective friction are reduced; field data show that the service life of vulnerable parts such as liners and hammers is significantly extended, and the number of maintenance downtimes is significantly reduced.

[0046] 3. This application precisely controls the crushing chamber pressure and screen openings, significantly reducing dust escape rate. The system features built-in vibration and current anomaly monitoring and active shutdown logic, enhancing operational safety. By fusing five types of sensor data through cross-modal attention and combining incremental learning, it can quickly adapt to fluctuations in the composition of construction solid waste from different demolition projects and regions. Combined with a BIM platform pre-loading strategy, material switching cold start time can be shortened to less than 30 minutes. The controller is compatible with OPC-UA and 5G industrial gateways, allowing seamless integration with existing production execution systems. The entire line can be quickly replicated using a "modular crushing unit + intelligent control cabinet" scheme, reducing construction time and overall cost. This invention achieves efficient, low-consumption, and sustainable operation of the construction solid waste crushing process, significantly outperforming existing fixed-parameter or single-feedback control schemes, providing a valuable intelligent overall solution for urban solid waste resource utilization.

[0047] 4. This application avoids gradient vanishing or exploding by adaptively scaling the gradient. By mapping the real-time granular distribution cosine similarity to β and embedding it into the activation function slope, the network automatically amplifies the gradient in scenarios with large prediction errors and automatically converges the gradient in scenarios with small prediction errors. This significantly reduces the gradient saturation problem caused by traditional fixed β activation functions under extreme inputs, making online incremental learning more stable. The lower the cosine similarity (the greater the deviation), the higher the β value, and the steeper the activation curve, which is equivalent to assigning greater loss weights to difficult samples. Experiments show that the improved activation function can significantly reduce the mean square error of granular prediction within 30 online fine-tuning cycles and significantly reduce the number of convergence steps. It reduces the sensitivity to hyperparameters such as the learning rate. The dynamic β enables the activation function to "self-adjust the slope" under different material conditions, making it less dependent on training hyperparameters such as the initial learning rate and weight decay. Field tests show that the learning rate can still maintain stable training within a fluctuation range of ±50%, while the traditional sigmoid scheme requires fine-tuning. Real-time correction reduces the impact of sensor drift. When there is a systematic deviation or short-term drift in the online particle size detector, β is based on the real detection result of the previous cycle and can immediately amplify the gradient to offset the prediction offset caused by the drift, maintain the effectiveness of the control parameter set, and avoid production line oscillation caused by sensor drift.

[0048] 5. This application simplifies multi-threshold maintenance and enhances cross-scenario generalization. The activation function implicitly incorporates the concept of "granularity error threshold," eliminating the need to manually configure multiple sets of granularity thresholds or start / stop conditions for different demolition projects. When the same model weights are directly transferred to other construction solid waste scenarios, optimal performance can be restored in just 5-10 online update cycles. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart of the adaptive control method for crushing construction solid waste according to the present invention;

[0051] Figure 2 This is a schematic diagram of an adaptive control device for crushing construction solid waste according to the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1:

[0054] In view of the aforementioned problems mentioned in the prior art, and in order to solve the above technical problems, as shown in the appendix. Figure 1 As shown: This application provides an adaptive control method for crushing construction solid waste, including the following steps:

[0055] S1: Real-time data collection of construction solid waste entering the crusher is achieved using a multi-source sensor array. The construction solid waste data includes particle size, steel reinforcement content, moisture content, strength, and feeding rate of the construction solid waste.

[0056] In some embodiments, the deployment and data acquisition process of the multi-source sensor array in a 150t / h fixed construction solid waste resource utilization production line is as follows: Mixed solid waste from the demolition site is dumped into a vibrating feed hopper by a tire loader, then enters a 1.2-meter-wide belt conveyor and is sent to the primary jaw crusher. To accurately grasp the material condition before crushing, the operator deploys a multi-source sensor array between the conveyor head and the crusher feed inlet. The sensors are installed sequentially from top to bottom and mirrored left to right, starting from the center line of the belt. All signals are aggregated to the edge computing industrial computer located in the control room via gigabit Ethernet.

[0057] First, the 3D laser scanner is mounted on a steel truss beam approximately 1.5 meters from the conveyor belt surface. It has a field of view of 1.5 meters and a frame rate of 10 Hz, and can output point clouds in real time. The particle size distribution curve is obtained through voxel filtering. After system calibration, the particle size D50 has an error of no more than 5 millimeters. For dust prevention, an air curtain device is added to the lens, and the protective glass is wiped with deionized water during each shift's downtime.

[0058] Approximately 30 centimeters downstream of the laser scanner, a row of eddy current sensor coils is arranged, each with an outer diameter of 15 centimeters and an operating frequency of 5000 Hz. If a steel bar or metal component passes over the material conveyor, induced eddy currents cause a change in the loop impedance. The multi-source sensor array maps the probability of metal presence into a lateral thermal image, which is then superimposed with the laser point cloud to calculate the steel bar content. After a week of sampling and statistics, the detection rate of a single 8-millimeter diameter steel bar is higher than 98%.

[0059] Next, a near-infrared hyperspectral camera with a spectral range of 900 to 1700 nanometers and a spatial resolution of 1024 pixels was installed, along with a line scanning configuration and LED strip lights. The system outputs cubic spectral bands at 25 frames per second, which are compressed by principal component analysis and then fed into a moisture regression model, controlling the moisture content measurement error to within ±1%. Both the lens and the light source are mounted on an adjustable pitch bracket to ensure coverage of the material surface under both full and partial belt load conditions.

[0060] To infer the strength of large concrete blocks, two impact acoustic emission sensors were installed at the tail wheel of the conveyor, fixed to the upper and lower surfaces of the thick-walled steel plate impact platform, respectively. When the material falls freely about 35 centimeters and impacts the platform, the sound signals collected by the impact acoustic emission sensors are analyzed using fast wavelet energy spectrum analysis to extract the dominant frequency and energy density. This data, combined with the block volume measured by laser, is then input into a pre-trained regression model to estimate the compressive strength of a single concrete block. If a high-strength block (≥C50) with a probability greater than 90% is detected, the system will preemptively reduce the target rotor linear velocity in the controller by 10% to decrease the impact load.

[0061] An electronic belt scale is installed below the base at the tail end of the belt, equipped with four shear beam load cells and one speed encoder, with a capacity of 200 tons / hour and an overall accuracy better than 0.5%. The real-time feed rate, along with the aforementioned sensor data, is timestamped and sent to a multimodal fusion network via the industrial computer's ZMQ bus for inference, generating a working condition vector and triggering the distribution of control parameters.

[0062] The entire multi-source sensor array completes a full sensing-fusion-decision cycle in 200ms intervals. Specifically, to ensure data consistency, all sensors use a unified IEEE 1588 precision clock for synchronization, automatically synchronizing once per shift. When the dust removal fan, air compressor, and spray dust suppression device detect an abnormal increase in the density of metal parts, they automatically increase their operating speed to ensure adequate light transmission and heat dissipation. Practice shows that during continuous 8-hour operation, the predicted errors for particle size D90, rebar ratio, and moisture content output by this array remain consistently within the set range, providing reliable input data for the adaptive control of the subsequent crusher.

[0063] S2: Input construction solid waste data into the fusion network to calculate the working condition vector;

[0064] In one embodiment, the PyTorch framework is used to deploy a Jetson AGXOrin edge computing module, aiming to keep the computational cost of a single forward inference operation within 5GMAC and the latency below 200ms. Sensor data streams are fed into the model in 200ms batches. The specific implementation of the five parallel coding branches and subsequent fusion layers is as follows:

[0065] The multimodal coding layer granularity sequence branch input is a 256-dimensional granular histogram obtained from the voxelization of the 3D laser scanning point cloud; it uses 1DConv→BatchNorm→ReLU→Residual block (kernel=3, channels16→32→16); the residual end outputs a feature tensor of length 256 and channels 16; the rebar content branch input is a 320-dimensional temporal vector flattened from the 32×10 metal probability heatmap output by the eddy current coil array in the same time window; after being connected to the residual through two layers of 1DConv (kernel3, stride2), it outputs a length of 80 and channels 32; the water content branch input is a 128×128 single-band image of a 128×128×32 spectral cube collected by a near-infrared hyperspectral camera after PCA compression; it uses 2DConv ( After kernel3, channels3→16) + two depth-separable residual blocks, the output is 32×32 with 32 channels; the intensity branch input is a 64×64 time-frequency map obtained by continuous wavelet transform of the impact acoustic emission sensor; two-stage 2DConv (kernel3, stride2) plus SE-ResBlock are used to output 16×16 with 32 channels; the feed rate branch input is a 10-point rate sequence sampled at 10Hz within 1 second by the electronic belt scale; a layer of 1DConv (kernel3) and GRU (hidden unit 32) is used to take the hidden state of the last time step and broadcast it as a tensor of length 16 and 32 channels; all branch ends are uniformly compressed to 32 dimensions through 1×1Conv to ensure consistent dimensions when entering the cross-modal attention layer.

[0066] Furthermore, the cross-modal attention layer employs pairwise cross-4-head dot-multiplication multi-head attention for the five tensors. Taking the granular tensor as the "dominant modality," the process is as follows: the granular tensor is projected into a query through a linear layer; the water content tensor is projected into a key / value pair, then passed through a softmax layer to obtain the attention matrix, which is multiplied by the value to obtain the context. After concatenating the four heads, it is mapped back to 32 channels through a linear layer. A learnable scalar γ (Sigmoid output) is used to perform residual weighting on the context and the original granular tensor: F_fuse = γ*context + (1-γ)*original. The dominant modality is rotated 5 times, resulting in 5 fused tensors, each maintaining the same shape as its original modality (32 channels). Further, each fused tensor in the channel-gated layer is input to a Squeeze-and-Excitation unit: Squeeze global average pooling to 32×1. Excitation: Two layers of fully connected (FC) 32→8→32, ReLU and Sigmoid output channel weight vectors; the weights are multiplied back to the original tensor by channel to filter redundant features and retain only key semantic information; the SE layer weights are pruned by 50%, retaining only important connections, supplemented by INT8 quantization to ensure the overall MAC number meets the budget; the global pooling layer and the working condition vector perform global average pooling on the five SE-processed tensors to obtain vectors of length 32, concatenate the five vectors, and reduce the dimensionality to 128 through a linear layer to obtain the final working condition vector. The output size of the linear layer can be adjusted between 64–256 dimensions as needed without exceeding the 5G MAC limit. Lightweight strategy model pruning: Structured channel pruning rate of 30%, MAC ≈ 4.2G after PTQ-INT8 quantization; storage usage reduced from 18MB to 5.5MB. Shared weights: The first convolutional layer in the five coding branches shares the kernel to reduce parameter redundancy. During TensorRT compilation and deployment, attention and SE layers are fused into a custom operator, with an inference latency of 138ms / batch. Within a 200ms cycle, the combined 3D laser and hyperspectral imaging technology effectively corrected the shift in the particle size curve due to moisture content observation error by approximately -7% (ΔD50). After the operating condition vector output by the network was fed into the convolutional neural network controller, the crusher rotor linear speed adjustment error remained ≤ ±2Hz. After 72 hours of continuous operation, no significant clock drift was observed, and the end-to-end average energy consumption was reduced by 16%.

[0067] This embodiment demonstrates that the lightweight spatiotemporal multimodal fusion network proposed in this invention can stably and in real-time generate high-quality operating condition vectors on edge devices, providing reliable input for subsequent adaptive control of crusher parameters.

[0068] S3: Input the working condition vector into the trained convolutional neural network model, and the convolutional neural network model outputs the set of control parameters for the crusher;

[0069] In one embodiment, on a fixed crushing line producing 150 tons of construction solid waste per hour, an edge computing industrial computer receives a 128-dimensional operating condition vector output by a fusion network. This vector is fed into a pre-trained convolutional neural network model every 0.2 seconds as a forward inference batch. The model uses an 8-layer deep residual backbone as its main body, with each residual block integrating a compression excitation unit that can adaptively allocate channel attention across dimensions. Two fully connected layers follow the backbone, ultimately outputting five normalized scalars corresponding to rotor linear velocity, discharge gap, hydraulic pressure, screen aperture size, and shape memory alloy deformation. After INT8 quantization, the model's single-batch inference time on JetsonOrin is 18 milliseconds. The controller first maps the rotor linear velocity scalar to a physical range of 800 to 1200 rpm and sends it to the frequency converter via the CAN bus. To reduce mechanical shock, a linear ramp of 50 rpm is used, reaching the target speed within 0.8 seconds.

[0070] Furthermore, the discharge gap scalar value is converted to a target value of 50 to 90 mm. The PLC sends distance commands to the double-acting hydraulic cylinder, and the liner displacement is adjusted using closed-loop PID control, achieving a positioning accuracy better than ±0.5 mm. Hydraulic pressure control is divided into two stages: the PLC first adjusts the proportional valve to coarsely adjust to 95% of the target pressure, and then the servo pump finely adjusts the remaining 5%, while real-time reading of the oil pressure sensor signal to suppress pressure fluctuations caused by sudden changes in material hardness. The screen opening size is adjusted via a stepper motor-driven telescopic mechanism, switching from 40 mm to 30 mm within 3 seconds. Specifically, the shape memory alloy actuator is constructed of Ni-Ti stranded wire, and the heating pulse current is issued at a ratio of 2.5 amperes / mm based on the deformation output of the model. Reversible scaling of at least 3 mm can be completed within 5 seconds, thereby tightening the crushing chamber geometry to enter the fine crushing mode. During the online weight update phase, every 3000 batches (approximately 10 minutes), the controller writes a new sample, composed of the actual detection vector and the model prediction vector, into a circular buffer. It then randomly selects 1024 historical samples and mixes them with 512 of the latest samples for three rounds of fine-tuning. The learning rate is set to one-twentieth of the offline value, and empirical replay is used to avoid catastrophic forgetting. After 72 hours of continuous operation, the model's MAPE decreased from the initial 6.8% to 5.1%, and the energy consumption per unit remained stable at 8.1 kWh·t. -1 There were no oscillations or circuit breaker failures causing shutdowns.

[0071] S4: Synchronously adjust the rotor linear speed, crushing chamber gap, hydraulic pressure and screen opening size of the crusher according to the control parameter set, and generate reversible deformation of the crushing chamber liner and / or hammer within 10 seconds through the shape memory alloy actuator, with a deformation of not less than 3mm, so as to dynamically reconstruct the geometric parameters of the crushing chamber to match the real-time crushing requirements.

[0072] For example, in a crusher with a capacity of 150 tons per hour, the factory modified two semi-wave removable liners into "programmable liners." Each liner has four pre-fabricated 8mm x 1mm dovetail grooves on its back for embedding Ni-Ti alloy wire actuators. The selected Ni-Ti wire has a diameter of 1.2mm, an effective length of 300mm per wire, an austenite termination temperature of 78℃, and a no-load recovery design of 4%. After the four alloy wires are fixed to the liner base via stainless steel crimp connectors, they are connected to a 24V / 40A pulse power supply; the wires are led out to the control cabinet via high-temperature resistant silicone sleeves and slip rings.

[0073] In standby mode, the system maintains the alloy wire in the martensitic phase (ambient temperature 35℃), with an initial gap of 85mm between the liner and the stationary plate. When the controller receives a deformation command Δ=3.5mm from the convolutional neural network, the PLC immediately triggers the PWM power module to output a pulse current with a duty cycle of 60% and a frequency of 10kHz, with a peak current of approximately 18A. The current heats the wire through its self-resistance, raising the wire temperature to 85℃ within 6 seconds. After the phase transformation is complete, the wire contracts by 4%, causing the liner to shift 3.5mm towards the inside of the crushing chamber. The displacement is monitored in real time by a magnetostrictive displacement sensor installed at the root of the liner; once the displacement reaches the target value and remains stable for 500ms, the PLC immediately reduces the duty cycle to 8% and enters a heat preservation state to maintain the austenitic phase.

[0074] After the crusher switches from coarse crushing mode to fine crushing mode, the controller simultaneously increases the rotor linear speed by 100 rpm and tightens the screen opening from 40 mm to 32 mm. If the online particle size detector at the back end reports an increase in the crushing rate, the controller will lower the deformation command gradient, retreating by 0.5 mm each time; the PLC will first turn off the insulation current and turn on the 12V air-cooling fan, reducing the wire temperature to below 45℃ within 4 seconds, causing the Ni-Ti wire to return to the martensitic phase and restore its initial length, and the liner gap will then widen.

[0075] To ensure a cycle life of "no less than 1×10^6 cycles," the manufacturer added 10mm thick flexible buffer pads to both ends of the drive wire to prevent the liner from impacting the limit position. After 60 days of continuous operation (approximately 200,000 switching cycles), the displacement decay was less than 2%, the resistance drift was less than 0.1Ω, and no fatigue wire breakage was observed. Thermal imaging showed that the maximum temperature difference between the wire and the liner was 12℃, the thermal insulation oxide layer was intact, and it did not affect the temperature of the crusher's lubricating oil. During the experimental period, the overall particle size qualification rate increased by 11%, and the unit energy consumption decreased by 3%, fully demonstrating the feasibility of the Ni-Ti shape memory alloy actuator completing a reversible deformation of ≥3mm within 10 seconds and operating stably for a long period.

[0076] S5: Real-time detection of particle size distribution of the crushed product, and feedback of the detection results to the trained convolutional neural network model for online weight update, realizing closed-loop adaptive optimization.

[0077] In some embodiments, the input and preprocessing vectors are 128-dimensional and are fed into the network after standardization. To be compatible with convolution operators, the vectors are first converted into 128×1×1 pseudo-feature maps using 1×1 convolutions. The deep feature extraction backbone uses 8 levels of cascaded residual blocks, with the number of channels in each level being 32, 32, 64, 64, 96, 96, 128, and 128 respectively. Each residual block contains two layers of 3×1 one-dimensional convolutions (stride 1, sampadding), followed by BatchNorm and ReLU6.

[0078] Before adding the residual backbone and shortcut branch outputs, a compressed excitation unit (SE) is inserted: first, global average pooling is performed on the backbone features to obtain a channel summary, then two fully connected layers (dimensionality reduction by 1 / 8, dimensionality increase to the original channel) generate Sigmoid weights, which are then recalibrated on the backbone channels and added to the shortcut branch to form the residual output. To ensure lightweight design, except for levels 1, 5, and 7, other convolutional kernels have 30% of their channels pruned according to L1 importance after offline training. Feature aggregation and the residual output of the last output layer are subjected to global average pooling to obtain a vector of length 128. Two fully connected layers are used: FC1 outputs a 64-dimensional vector connected to ReLU6, and FC2 outputs five normalized scalars, corresponding to rotor linear velocity, discharge gap, hydraulic pressure, screen aperture size, and SMA deformation, respectively. The normalized scalars are then scaled to the physical range on the controller side. Online incremental learning and sample playback are implemented, with a circular buffer capacity of 4096, and samples of <predicted granularity vector, measured granularity vector> are written in real time. Fine-tuning is triggered every 30 seconds: the residual blocks of levels 1-4 are frozen, and only levels 5-8 and the two fully connected layers are updated; 1024 samples (the latest 25% and the historical 75%) are randomly sampled and fine-tuned for 3 epochs, with a learning rate of 1 / 50 of the offline training value. Sample replay ensures that old scene features are continuously reviewed, avoiding catastrophic forgetting caused by learning only the latest materials. If the MSE of the new weights is 15% higher than that of the old weights on the validation set, it is rolled back immediately. Inference and deployment characteristics are quantized with INT8 across the entire network, with a total of approximately 1.9M parameters after pruning, a single inference of 3.6GMAC, and an average latency of 18ms. The 5-dimensional control parameters output from each batch of inference are driven by the PLC linear ramp or closed-loop PID to drive the frequency converter, hydraulic cylinder, screen actuator, and Ni-Ti drive wire, forming a real-time collaborative control closed loop.

[0079] Specifically, an online laser diffractometer is installed on the top of the crusher's discharge belt cover. The test wavelength is 635nm, the measurement range is 10µm-100mm, and the sampling frequency is 10Hz. When the product material passes through the measurement window, the particle size analyzer outputs a 64-dimensional volume distribution histogram every 100ms. The PLC packages the histogram and its synchronized timestamp and sends it to the edge computing industrial computer via the MQTT protocol. After parsing, it is recorded as the measured particle size vector D_real. When the convolutional neural network model completes one forward inference and outputs the control parameter set, the controller simultaneously saves the feature mapping of the penultimate layer of the previous cycle network, compressing it into a 64-dimensional particle size prediction vector D_pre through a linear layer. The system maintains a circular buffer with a capacity of 4096 entries to store key-value pairs. When the latest pair enters the buffer, if the buffer is full, it automatically overwrites the oldest sample. To avoid abnormal particle size disturbances caused by foreign object impacts, the cosine similarity between D_pre and D_real is calculated before writing. If the similarity is lower than -0.1 (extreme deviation), it is marked as "noise" and discarded.

[0080] The controller triggers online weight updates every 30 seconds. During an update, 1024 samples are randomly selected, with the latest 256 guaranteed to be sampled, and the rest randomly sampled from a buffer. The example uses the Adam optimizer with a learning rate set to 1 / 50 of the initial offline value. Only the weights of the last two fully connected layers and the SE-compression coefficient γ are fine-tuned; the remaining weights are frozen. Incremental learning iterates for 3 epochs, taking approximately 2.3 seconds in total, without affecting the main control loop. During training, the main thread continues inference but temporarily locks the weight update region to avoid read / write conflicts.

[0081] After incremental learning is completed, the new weights immediately replace the old weights. To avoid catastrophic forgetting, the system calculates the MSE difference between the new and old models on the most recent 100 samples. If the difference is greater than 15%, the system rolls back to the old weights and records an alarm log. After running in the field for a week, the model's average granularity prediction error decreased from the initial 6.5% to 4.8%, and the crusher's energy consumption further decreased by 2.7%, verifying the effectiveness of the "real-time granularity detection - online weight update" closed-loop adaptive optimization.

[0082] The multi-source sensor array includes: a three-dimensional laser scanner for acquiring particle size distribution, an eddy current sensor for estimating the steel reinforcement ratio, a near-infrared hyperspectral camera for determining moisture content, an impact acoustic emission or hardness sensor for inferring strength, and a weighing and metering module for obtaining the feed rate.

[0083] Specifically, the fusion network is a lightweight spatiotemporal multimodal neural network, employing cross-modal cross-attention and channel gating mechanisms, and using pruning and quantization techniques to ensure that the computational cost of a single inference operation does not exceed 5 GMAC. The lightweight spatiotemporal multimodal neural network is deployed on a Jetson AGXOrin (30W) edge computing module, aiming to complete one forward inference operation within a 200ms control cycle with a total multiply-accumulate count not exceeding 5 GMAC. The network receives five sensor tensor inputs, corresponding to a granularity histogram, an eddy current metal probability heatmap, a near-infrared hyperspectral single-band map, an acoustic emission time-frequency map, and a feed rate sequence, respectively.

[0084] First, in the multimodal encoding stage, an independent encoding branch is established for each data stream: 1D depthwise separable residual blocks are used for temporal signals, and 2D MobileNet-V3-lite residual blocks are used for image or spectral signals; all branches are capped at 32 channels, with a maximum convolutional kernel of 3×3, thus limiting the theoretical upper bound of MAC. Then, the cross-modal attention module is introduced. The system uses a four-head dot-multiplication attention mechanism, with pairwise modal cross-interactions, each head retaining only 8-dimensional Query, Key, and Value representations; the attention output and the original features are residually weighted using a learnable scalar γ before being fed into the channel-gated layer. The channel-gated layer uses Squeeze-and-Excitation, first performing global average pooling, then reducing the channel weights to 1 / 8 channels after two fully connected layers and then increasing them back up, with the channel weights output using a Sigmoid function. To further reduce computational cost, a 50% structured pruning is applied to the second fully connected layer of the SE layer, retaining only important connections.

[0085] After offline training, a 50% channel pruning strategy was used to sparsify the linear projection layers in the five encoding branches and cross-attention layers: redundant convolutional kernels and matrix columns were pruned based on the channel L1-norm importance index; then PTQ-INT8 quantization was performed to map floating-point weights and activations to 8-bit integers. After pruning and quantization, the network parameters were reduced from 21MB to 6MB, and the theoretical MAC for single-batch inference was approximately 4.3G. During the deployment phase, the TensorRT engine was enabled, fusing 1D and 2D convolutional groups into a single operator and expanding the attention matrix multiplication into an intra-batch GEMM; at the same time, the batch size was locked at 1, and the input size dynamic range was 95%-105% to obtain optimal memory reuse. Actual tests showed that the forward inference latency was 145ms, and the power consumption fluctuation did not exceed 6W; no significant temperature rise or memory leak occurred after 72 hours of continuous operation. Through this lightweight design, the fusion network retains cross-modal interaction and channel adaptation capabilities, while compressing the computational load of a single inference operation to within the 5G MAC on edge hardware, providing real-time, high-quality operational vector input for subsequent convolutional neural network controllers.

[0086] Specifically, the trained convolutional neural network model includes a deep feature extraction backbone network composed of residual blocks and compressed excitation units, and adopts an incremental learning and sample replay mechanism during online weight updates to avoid catastrophic forgetting.

[0087] Specifically, the shape memory alloy actuator uses Ni-Ti alloy wire with an austenite termination temperature of 60℃-90℃. Driven by a resistance-heated pulsed current, it completes a reversible deformation of at least 3mm within 10 seconds, with a cycle life of at least 10 cycles. 6 Second-rate.

[0088] Specifically, the convolutional neural network model employs an improved activation function during the inference process in step S3 and the online weight update process in step S5. :

[0089]

[0090]

[0091] in, For the output of the convolutional layer, This is the granularity distribution prediction vector of the output of the previous layer in the convolutional neural network. This is the real-time detected particle size distribution vector of the crushed product. Represents the dot product. Represents the 2-norm; For the reason and The adaptive coefficients obtained by the cosine similarity linear mapping are dynamically adjusted with the granularity prediction error to enhance the network's adaptability to granularity distribution fluctuations and provide targeted gradient feedback during online weight updates.

[0092] Specifically, the step of inputting construction solid waste data into the fusion network to calculate the working condition vector is as follows:

[0093] S21: The multimodal coding layer is configured with a parallel coding branch for each of the sensor data of particle size, steel reinforcement content, moisture content, strength and feed rate. The coding branch uses a one-dimensional convolutional residual unit for time-series signals and a two-dimensional convolutional residual unit for image and spectral signals, and outputs a feature tensor of the same channel dimension.

[0094] S22: The cross-modal attention layer uses the dot product attention mechanism to calculate the correlation weights of the feature tensors of any two modalities and generates weighted fusion features accordingly.

[0095] S23: The channel gating layer applies Squeeze-and-Excitation gating operations to the weighted fusion features to compress redundant channels;

[0096] S24: The global pooling layer performs global average pooling on the gated feature map to obtain a working condition vector with an adjustable length of 64-256 dimensions, which is used to characterize the comprehensive working condition of construction solid waste.

[0097] Specifically, the control parameter set includes: target linear velocity of the crusher rotor, target discharge gap of the crushing chamber, target working pressure of the hydraulic system, target opening size of the screen, and target reversible deformation of the shape memory alloy actuator; the controller sends the target linear velocity of the crusher rotor to the frequency conversion drive module, and uses a linear acceleration and deceleration ramp to make the rotor reach the set linear velocity; it sends a displacement command to the hydraulic actuator cylinder to adjust the relative position of the crushing chamber liner to the target discharge gap of the crushing chamber; it controls the hydraulic pressure to the target working pressure of the hydraulic system through a proportional valve and a servo pump, and compensates for pressure drift caused by material hardness fluctuations in real time through closed-loop control; it adjusts the opening size to the target opening size of the screen by driving the electric slide rail or telescopic screen module; and it applies a pulse current to the shape memory alloy actuator to complete the reversible deformation, so as to dynamically reconstruct the geometric parameters of the crushing chamber.

[0098] Specifically, the aforementioned weighted fusion features include:

[0099] Perform a linear projection on the first modality feature tensor to obtain the query vector Q;

[0100] Perform linear projections on the second modality feature tensor to obtain the key vector K and the value vector V;

[0101] The dot product of Q and K is calculated and normalized by the square root of the vector dimension before being fed into the Softmax function to obtain the attention weight matrix A;

[0102] Multiply the attention weight matrix A by the value vector V to obtain the cross-modal context vector;

[0103] The cross-modal context vector and the first modal feature tensor are weighted and added together using learnable gating coefficients to obtain a weighted fusion feature. This weighted fusion feature is then provided to the global pooling layer to generate a working condition vector. An online laser diffractometer with a measurement range of 10µm–100mm is installed above the crusher discharge port. A 635nm laser beam is introduced into the laser diffractometer through a hole drilled in the center of the discharge conveyor belt. A dual-linear array detector is used to collect the scattered light intensity. The laser diffractometer has a built-in fast Fourier inversion algorithm to convert the scattered energy into a particle size histogram. The system divides the accumulated scattering data within 0.1s into 64 logarithmically scaled particle size intervals and outputs a 64-dimensional volume distribution histogram. After receiving the data, the PLC normalizes the volume percentage of each interval by dividing by the sum to obtain the real-time detected particle size distribution vector D_real of the crushed product. The vector length is fixed at 64, and the sum of all components is 1. At the control end, the trained convolutional neural network not only outputs five control parameters, but also adds a bypass after the penultimate layer. Global average pooling of the backbone features is followed by a 64-unit fully connected layer, and then Softmax normalization is used to obtain the granularity distribution prediction vector D_pre. Since network inference can be completed within 20ms after the end of the previous sampling period, the D_pre corresponding to the current working condition vector is paired with the D_real output by the laser granularity analyzer at the next moment; both have the same timestamp offset. The system maintains a circular buffer with a capacity of 4096 for storage.<D_pre,D_real> Key-value pairs are used; before writing, their cosine similarity is calculated, and outlier samples with a difference greater than 3σ from the average are filtered out. When incremental learning is triggered, the controller randomly samples 25% of the latest samples and 75% of the historical samples to use as the fine-tuning dataset; the network only opens the last two levels of residual blocks and the 64-unit prediction head weights for training, using the Kullback-Leibler divergence of D_pre versus D_real as the loss for three rounds of iteration. After fine-tuning, the new model will be used in the next control cycle to continuously update the crusher control parameters, achieving closed-loop adaptive optimization of particle size distribution and energy consumption.

[0104] This application also provides an adaptive control device for crushing construction solid waste, such as... Figure 2 The system includes: a feeding device that feeds materials through a vibrating hopper in a conveying subsystem; specifically, it is equipped with a variable frequency vibrating motor, whose speed is regulated by a central PLC via a 4–20mA analog signal. A heavy-duty belt conveyor: width ≥ 1.2m, with an electronic belt scale mounted on the tail pulley; speed measurement pulses and weighing signals are connected to the PLC via RS-485.

[0105] The multi-source sensor array includes: a 3D laser scanner (TCP / IP gigabit Ethernet direct connection to an edge computer); an eddy current metal detection coil array (RS-485 bus multi-point mounting); a near-infrared hyperspectral camera (PoE + Ethernet, aggregated by an independent switch); an impact acoustic emission sensor (coaxial BNC connection to an industrial sound card, acquired via EtherCAT); a weighing and metering module (Modbus-RTU, 485 bus); sensor time synchronization adopts IEEE-1588PTP, with a built-in clock master node on the gigabit switch.

[0106] The crusher and actuator adopt a combined jaw-impact crusher: the main shaft is equipped with a variable frequency drive (VFD); the VFD receives the target speed and returns the instantaneous values ​​of current and speed via CANopen. The hydraulic actuator module includes an electro-hydraulic servo pump and proportional valve, with two 4–20mA pressure feedback channels; the PLC outputs analog or EtherCAT-I / O modules. The shape memory alloy (Ni-Ti) actuator group consists of 4 × 1.2mm alloy wires, driven by a 24V / 40A PWM pulse power supply; the PWM control signal is output via a 24V digital high-speed output port; a displacement sensor (magnetostrictive) provides 0–10V feedback. The adjustable screening module includes a stepper motor driving a telescopic screen, with the driver using CANopen; a screen position encoder provides 24V pulse feedback to the PLC.

[0107] Online detection and energy consumption monitoring laser diffraction particle size analyzer: outputs 64-dimensional particle size histogram via Gigabit Ethernet; link directly to edge AI computer. Three-phase energy meter connects to power switch via Modbus-TCP, shared with SCADA.

[0108] The control and computing unit includes a central PLC (redundant CPU, EtherCAT master station): responsible for logical sequence, interlocking, and safe shutdown. An edge AI industrial computer (Jetson AGX Orin + x86): runs a converged network and convolutional neural network controller; it shares control parameter sets with the PLC via an EtherCAT-UDP gateway. The HMI touchscreen reads key variables from the PLC / AI via OPC-UA, enabling on-site visualization and manual intervention. The 5G industrial gateway pushes operation logs and alarms to the cloud-based SCADA system.

[0109] Power and Auxiliary Systems: The power distribution cabinet includes a magnetothermal circuit breaker, a soft-start contactor, and a surge protection module; it supplies power to the VFD, hydraulic pump, PWM power supply, etc. Atomized dust suppression and negative pressure dust collection are controlled by PLC relay outputs; the fan frequency is automatically increased when the eddy current coil detects a surge in rebar density. The air conditioning / heat exchanger maintains the electrical cabinet and edge computer enclosure at 0–45°C, with a temperature signal fed back to the PLC at 4–20mA.

[0110] This embodiment provides an adaptive control device for crushing construction solid waste, including:

[0111] a) Feeding mechanism, used to transport construction solid waste to the crusher;

[0112] b) A multi-source sensor array, set at the feeding mechanism or the feed inlet of the crusher, including a 3D laser scanner, an electromagnetic induction or eddy current sensor, a near-infrared spectral camera, an impact acoustic emission or hardness sensor, and a weighing and metering module, is used to collect the particle size, steel reinforcement content, moisture content, strength and feeding rate of construction solid waste in real time.

[0113] c) A controller with a built-in multimodal fusion network for calculating and generating a working condition vector from the data output by the multi-source sensor array; a trained convolutional neural network model for outputting a set of crusher control parameters based on the working condition vector; and an online weight update unit for incrementally updating the convolutional neural network model after receiving the product particle size distribution detection result.

[0114] d) A crusher whose structure allows for adjustment of the rotor linear speed and the crushing chamber gap;

[0115] e) A variable frequency drive module, connected to the crusher and adjusting the rotor linear speed according to the control parameter set;

[0116] f) The hydraulic actuator module is connected to the crusher liner and adjusts the crushing chamber gap and hydraulic pressure according to the control parameter set;

[0117] g) Shape memory alloy actuators are arranged in the crushing chamber liner and / or hammer, configured to generate reversible deformation to dynamically reconstruct the crushing chamber geometry.

[0118] h) The screening module has an adjustable screen opening size and performs adjustments according to the set of control parameters;

[0119] i) The output detection module includes an online particle size detector and an energy consumption metering device, which are used to acquire the particle size distribution and unit energy consumption of the crushed product in real time, and send the detection results to the controller;

[0120] The controller is configured to execute the adaptive control process of construction solid waste crushing according to the aforementioned adaptive control method for construction solid waste crushing, so as to realize the synchronous linkage adjustment of the crusher, hydraulic actuator, shape memory alloy driver and screening module, and complete the closed-loop adaptive optimization.

[0121] This invention provides an adaptive control method and equipment for crushing construction solid waste, which achieves the following beneficial technical effects:

[0122] 1. This application employs multi-source sensing-convolutional neural network closed-loop control and rapid geometric self-reconstruction of shape memory alloys to ensure the crusher always operates under optimal conditions, significantly reducing unit energy consumption, peak current, and power distribution capacity. Real-time online particle size detection results directly participate in activation function adjustment and weight increment learning, ensuring the crushed product particle size stably falls within the target range. This significantly improves the pass rate compared to traditional manual adjustment methods, reducing over-crushing and material return.

[0123] 2. The Ni-Ti shape memory alloy actuator of this application achieves reversible deformation of not less than 3mm within 10 seconds and can switch between "coarse crushing" and "fine crushing" modes without stopping the machine. Compared with independent adjustment by hydraulic cylinders, the switching time is significantly shortened, significantly reducing the risk of cavity blockage. By dynamically adjusting the rotor linear speed and cavity clearance as needed, excessive impact and ineffective friction are reduced. Field data shows that the service life of vulnerable parts such as liners and hammers is significantly extended, and the number of maintenance downtimes is significantly reduced.

[0124] 3. This application precisely controls the crushing chamber pressure and screen openings, significantly reducing dust escape rate. The system features built-in vibration and current anomaly monitoring and active shutdown logic, enhancing operational safety. By fusing five types of sensor data through cross-modal attention and combining incremental learning, it can quickly adapt to fluctuations in the composition of construction solid waste from different demolition projects and regions. Combined with a BIM platform pre-loading strategy, material switching cold start time can be shortened to less than 30 minutes. The controller is compatible with OPC-UA and 5G industrial gateways, allowing seamless integration with existing production execution systems. The entire line can be quickly replicated using a "modular crushing unit + intelligent control cabinet" scheme, reducing construction time and overall cost. This invention achieves efficient, low-consumption, and sustainable operation of the construction solid waste crushing process, significantly outperforming existing fixed-parameter or single-feedback control schemes, providing a valuable intelligent overall solution for urban solid waste resource utilization.

[0125] 4. This application avoids gradient vanishing or exploding by adaptively scaling the gradient. By mapping the real-time granular distribution cosine similarity to β and embedding it into the activation function slope, the network automatically amplifies the gradient in scenarios with large prediction errors and automatically converges the gradient in scenarios with small prediction errors. This significantly reduces the gradient saturation problem caused by traditional fixed β activation functions under extreme inputs, making online incremental learning more stable. The lower the cosine similarity (the greater the deviation), the higher the β value, and the steeper the activation curve, which is equivalent to assigning greater loss weights to difficult samples. Experiments show that the improved activation function can significantly reduce the mean square error of granular prediction within 30 online fine-tuning cycles and significantly reduce the number of convergence steps. It reduces the sensitivity to hyperparameters such as the learning rate. The dynamic β enables the activation function to "self-adjust the slope" under different material conditions, making it less dependent on training hyperparameters such as the initial learning rate and weight decay. Field tests show that the learning rate can still maintain stable training within a fluctuation range of ±50%, while the traditional sigmoid scheme requires fine-tuning. Real-time correction reduces the impact of sensor drift. When there is a systematic deviation or short-term drift in the online particle size detector, β is based on the real detection result of the previous cycle and can immediately amplify the gradient to offset the prediction offset caused by the drift, maintain the effectiveness of the control parameter set, and avoid production line oscillation caused by sensor drift.

[0126] 5. This application simplifies multi-threshold maintenance and enhances cross-scenario generalization. The activation function implicitly incorporates the concept of "granularity error threshold," eliminating the need to manually configure multiple sets of granularity thresholds or start / stop conditions for different demolition projects. When the same model weights are directly migrated to other construction solid waste scenarios, optimal performance can be restored in just 5-10 online update cycles.

[0127] The above provides a detailed description of an adaptive control method and equipment for crushing construction solid waste. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas and methods of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. An adaptive control method for crushing construction solid waste, characterized in that, Including the following steps: S1: Real-time data collection of construction solid waste entering the crusher is achieved using a multi-source sensor array. The construction solid waste data includes particle size, steel reinforcement content, moisture content, strength, and feeding rate of the construction solid waste. S2: Input construction solid waste data into the fusion network to calculate the working condition vector; S3: Input the working condition vector into the trained convolutional neural network model, and the convolutional neural network model outputs the set of control parameters for the crusher; S4: Synchronously adjust the rotor linear speed, crushing chamber gap, hydraulic pressure and screen opening size of the crusher according to the control parameter set, and generate reversible deformation of the crushing chamber liner and / or hammer within 10 seconds through the shape memory alloy actuator, with a deformation of not less than 3mm, so as to dynamically reconstruct the geometric parameters of the crushing chamber to match the real-time crushing requirements. S5: Real-time detection of particle size distribution of the crushed product, and feedback of the detection results to the trained convolutional neural network model for online weight update, to achieve closed-loop adaptive optimization; The convolutional neural network model employs an improved activation function during the inference process in step S3 and the online weight update process in step S5. : ； ； in, For the output of the convolutional layer, This is the granularity distribution prediction vector of the output of the previous layer in the convolutional neural network. This is the real-time detected particle size distribution vector of the crushed product. Represents the dot product. Represents the 2-norm; For the reason and The adaptive coefficients obtained by the cosine similarity linear mapping are dynamically adjusted with the granularity prediction error to enhance the network's adaptability to granularity distribution fluctuations and provide targeted gradient feedback during online weight updates.

2. The adaptive control method for crushing construction solid waste as described in claim 1, characterized in that, The multi-source sensor array includes: a three-dimensional laser scanner for acquiring particle size distribution, an eddy current sensor for estimating the steel reinforcement content, a near-infrared hyperspectral camera for determining moisture content, an impact acoustic emission or hardness sensor for inferring strength, and a weighing and metering module for obtaining the feed rate.

3. The adaptive control method for crushing construction solid waste as described in claim 1, characterized in that, The fusion network is a lightweight spatiotemporal multimodal neural network that employs cross-modal attention and channel gating mechanisms, and uses pruning and quantization techniques to ensure that the computational cost of a single inference does not exceed 5 GMAC.

4. The adaptive control method for crushing construction solid waste as described in claim 1, characterized in that, The trained convolutional neural network model includes a deep feature extraction backbone network composed of residual blocks and compressed excitation units, and adopts incremental learning and sample replay mechanisms during online weight updates to avoid catastrophic forgetting.

5. The adaptive control method for crushing construction solid waste as described in claim 1, characterized in that, The shape memory alloy actuator uses Ni-Ti alloy wire with an austenite termination temperature of 60℃-90℃. Driven by a resistance-heated pulsed current, it completes a reversible deformation of at least 3mm within 10 seconds, with a cycle life of at least 10 cycles. 6 Second-rate.

6. The adaptive control method for crushing construction solid waste as described in claim 1, characterized in that, The steps for inputting construction solid waste data into a fusion network to calculate the working condition vector are as follows: S21: The multimodal coding layer is configured with a parallel coding branch for the sensor data of particle size, steel reinforcement content, moisture content, strength and feed rate. The coding branch uses a one-dimensional convolutional residual unit for time-series signals and a two-dimensional convolutional residual unit for image and spectral signals, and outputs a feature tensor of the same channel dimension. S22: The cross-modal attention layer uses the dot product attention mechanism to calculate the correlation weights of any two modal feature tensors and generates weighted fusion features; S23: The channel gating layer applies Squeeze-and-Excitation gating operations to the weighted fusion features to compress redundant channels; S24: The global pooling layer performs global average pooling on the gated feature map to obtain a working condition vector with an adjustable length of 64-256 dimensions, which is used to characterize the comprehensive working condition of construction solid waste.

7. The adaptive control method for crushing construction solid waste as described in claim 1, characterized in that, The control parameter set includes: target linear velocity of the crusher rotor, target discharge gap of the crushing chamber, target working pressure of the hydraulic system, target opening size of the screen, and target reversible deformation of the shape memory alloy actuator. The controller sends the target linear velocity of the crusher rotor to the frequency conversion drive module, and uses a linear acceleration and deceleration ramp to make the rotor reach the set linear velocity. It sends a displacement command to the hydraulic actuator cylinder to adjust the relative position of the crushing chamber liner to the target discharge gap of the crushing chamber. It controls the hydraulic pressure to the target working pressure of the hydraulic system through a proportional valve and a servo pump, and compensates for pressure drift caused by material hardness fluctuations in real time through closed-loop control. It adjusts the opening size to the target opening size of the screen by driving the electric slide rail or telescopic screen module. It applies a pulse current to the shape memory alloy actuator to complete the reversible deformation, so as to dynamically reconstruct the geometric parameters of the crushing chamber.

8. The adaptive control method for crushing construction solid waste as described in claim 6, characterized in that, The generated weighted fusion features include: Perform a linear projection on the first modality feature tensor to obtain the query vector Q; Perform linear projections on the second modality feature tensor to obtain the key vector K and the value vector V; The dot product of Q and K is calculated and normalized by the square root of the vector dimension before being fed into the Softmax function to obtain the attention weight matrix A; Multiply the attention weight matrix A by the value vector V to obtain the cross-modal context vector; The cross-modal context vector and the first modal feature tensor are added together with residual weights according to learnable gating coefficients to obtain weighted fusion features, and the weighted fusion features are provided to the global pooling layer to generate the working condition vector.

9. An adaptive control device for crushing construction solid waste, characterized in that, include: a) Feeding mechanism, used to feed construction solid waste into the crusher; b) A multi-source sensor array, set at the feeding mechanism or the feed inlet of the crusher, including a 3D laser scanner, an electromagnetic induction or eddy current sensor, a near-infrared spectral camera, an impact acoustic emission or hardness sensor, and a weighing and metering module, is used to collect the particle size, steel reinforcement content, moisture content, strength and feeding rate of construction solid waste in real time. c) A controller with a built-in multimodal fusion network for calculating and generating a working condition vector from the data output by the multi-source sensor array; a trained convolutional neural network model for outputting a set of crusher control parameters based on the working condition vector; and an online weight update unit for incrementally updating the convolutional neural network model after receiving the product particle size distribution detection result. d) A crusher whose structure allows for adjustment of the rotor linear speed and the crushing chamber gap; e) A variable frequency drive module, connected to the crusher and adjusting the rotor linear speed according to the control parameter set; f) A hydraulic actuator module, connected to the crusher liner, adjusts the crushing chamber clearance and hydraulic pressure according to the set of control parameters; g) Shape memory alloy actuators, arranged in the crushing chamber liner and / or hammer, configured to generate reversible deformation to dynamically reconstruct the crushing chamber geometry. h) A screening module having an adjustable screen opening size and performing adjustments according to the set of control parameters; i) Output detection module, including online particle size detector and energy consumption metering device, is used to acquire the particle size distribution and unit energy consumption of the crushed product in real time, and send the detection results to the controller; The controller is configured to perform an adaptive control process for crushing construction solid waste according to the method described in claim 1, thereby achieving synchronous linkage adjustment of the crusher, hydraulic actuator, shape memory alloy actuator and screening module, and completing closed-loop adaptive optimization.

Images