A component size detection device and detection method for rail transportation

Abstract

The present application discloses a component size detection device for rail transit and a detection method thereof, comprising: collecting multi-source sensor data, wherein the multi-source sensor data comprises a laser ranging data stream, an image data stream and an ultrasonic data stream; based on the standardized multi-source sensor data, performing segmentation calculations through a multi-level pulse neural morphological network to obtain a segmentation result of the rail transit component, wherein the segmentation result comprises contour features and regional features; based on the segmentation result, performing multi-objective optimization using a gradient descent optimizer to obtain an optimized size parameter set, wherein the multi-objective optimization comprises maximizing measurement accuracy, minimizing calculation time and minimizing resource consumption; based on the optimized size parameter set, performing time difference learning through a distributed error calculation unit to obtain a final size measurement result of the rail transit component.

Description

Technical Field

[0001] The present invention belongs to the field of rail transit detection, and in particular relates to a component size detection device and a detection method for rail transit. Background Art

[0002] In the field of rail transit, component size detection is an important part of ensuring product quality and operational safety. At present, there are mainly the following technical solutions for the size detection of rail transit components:

[0003] One solution is the traditional method based on manual measurement, which is performed by using tools such as calipers and rulers. This method is inefficient and prone to human errors.

[0004] Another solution is an automated measurement system based on a single sensor. Although it improves efficiency, it is difficult to cope with multi-dimensional measurement needs in complex environments.

[0005] Another solution uses a simple machine vision system, but due to the lack of support from intelligent algorithms, it is difficult to ensure measurement accuracy in complex environments. Summary of the invention

[0006] In view of the above problems, the present invention provides a rail transit component size detection device and a detection method thereof. Multi-source sensor data are processed using multi-level pulse neuromorphic computing technology, multi-objective optimization is achieved through a gradient descent optimizer, and a distributed error calculation unit is introduced for time difference learning, thereby improving the accuracy and efficiency of detection.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] A method for detecting dimensions of rail transit components, comprising:

[0009] Collect multi-source sensor data, including laser ranging data stream, image data stream and ultrasonic data stream;

[0010] Perform noise filtering and data standardization on the laser ranging data stream, perform distortion correction and illumination compensation on the image data stream, perform signal enhancement and denoising on the ultrasonic data stream, and obtain standardized multi-source sensor data;

[0011] Based on standardized multi-source sensor data, segmentation calculation is performed through a multi-level pulse neuromorphic network to obtain the segmentation results of rail transit components, where the segmentation results include contour features and regional features;

[0012] Based on the segmentation results, a gradient descent optimizer is used to perform multi-objective optimization to obtain an optimized set of size parameters, where the multi-objective optimization includes maximizing measurement accuracy, minimizing computing time, and minimizing resource consumption.

[0013] Based on the optimized set of dimensional parameters, time difference learning is performed through a distributed error calculation unit to obtain the final dimensional measurement results of the rail transit components.

[0014] Wherein, the step of performing segmentation calculation by a multi-level spiking neuromorphic network comprises:

[0015] Converting the standardized multi-source sensor data into a time domain pulse sequence includes:

[0016] Convert the laser ranging data stream into a distance coded pulse sequence,

[0017] converting the image data stream into a grayscale coded pulse sequence,

[0018] converting the ultrasonic data stream into an intensity coded pulse sequence;

[0019] Using the time-domain pulse sequence to perform feature extraction through the multi-level pulse neuromorphic network;

[0020] Based on the result of the feature extraction, the contour and area segmentation of the rail transit component are achieved.

[0021] Wherein, the multi-level pulse neuromorphic network comprises:

[0022] An input layer for receiving the standardized multi-source sensor data,

[0023] a pulse coding layer for converting the multi-source sensor data into a time domain pulse sequence,

[0024] A feature extraction layer for extracting features using the time domain pulse sequence,

[0025] A segmentation decision layer is used to implement component segmentation based on the result of the feature extraction.

[0026] Wherein, the method of using the time domain pulse sequence to extract features through the multi-level pulse neuromorphic network includes:

[0027] Extracting basic edge and contour features through the first layer of neurons of the multi-level spiking neuromorphic network;

[0028] Extracting structural features of corner points and surfaces through the second layer of neurons of the multi-level spiking neuromorphic network;

[0029] The extracted features are combined and fused through the third layer of neurons of the multi-level pulse neuromorphic network to form a complete feature representation of the rail transit component.

[0030] Wherein, the step of using gradient descent optimizer to perform multi-objective optimization includes:

[0031] Construct a joint optimization objective function of measurement accuracy, computing time and resource consumption;

[0032] Constructing a gradient descent optimizer based on the joint optimization objective function;

[0033] Multi-objective optimization is achieved by iteratively updating the parameters of the gradient descent optimizer.

[0034] Wherein, the multi-objective optimization is achieved by iteratively updating the parameters of the gradient descent optimizer, including:

[0035] Perform forward propagation calculations, calculate the values ​​of each objective function based on the current parameter configuration, and evaluate the overall performance of each objective function;

[0036] Perform reverse update and adjust parameter configuration according to the gradient information of the objective function;

[0037] A dynamic annealing mechanism is introduced to explore a larger solution space through random perturbations to avoid falling into the local optimal solution.

[0038] Wherein, the step of performing time difference learning by a distributed error calculation unit comprises:

[0039] Construct multiple parallel error calculation units;

[0040] Calculate the time difference between the current measurement result and the historical best result;

[0041] updating a learning parameter based on the time difference;

[0042] Output the final size measurement result of the rail transit component.

[0043] Wherein, the method further comprises:

[0044] Performing real-time verification of the final size measurement result;

[0045] When a significant deviation is detected in the final dimension measurement result, an alarm is triggered and an automatic calibration procedure is started.

[0046] Wherein, the outputting of the final size measurement result of the rail transit component includes:

[0047] Output various dimensional data of the rail transit components;

[0048] Output measurement confidence information;

[0049] Output measurement error estimation information;

[0050] Generates a measurement report containing metrics for evaluating the quality of the measurement.

[0051] The present application also provides a component size detection device for rail transit, comprising:

[0052] Multi-source sensor array, including:

[0053] The laser distance measuring sensors arranged around the detection area are used to collect the distance information of the rail transit components.

[0054] An industrial camera is arranged above the detection area, and is used to collect image information of the rail transit components.

[0055] An ultrasonic sensor array disposed at the bottom of the detection area is used to collect thickness information of the rail transit component;

[0056] A processor is used to execute the component size detection method for rail transit as described above.

[0057] Beneficial effects:

[0058] 1) The use of multi-level pulse neuromorphic computing technology improves the system's ability to process multi-source sensor data and enhances the robustness of detection;

[0059] 2) Multi-objective optimization is achieved through gradient descent optimizer, balancing multiple objectives such as measurement accuracy, computing time and resource consumption;

[0060] 3) The introduction of distributed error calculation unit and time difference learning mechanism achieves continuous optimization of measurement accuracy;

[0061] 4) The system has good adaptability and can meet the detection requirements of different types of rail transit components. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the disclosed embodiments in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.

[0063] Figure 1 A schematic diagram of the overall structure of a rail transit component size detection device provided by an embodiment of the present invention;

[0064] Figure 2 This is a schematic diagram of a multi-source sensor data acquisition process in an embodiment of the present invention;

[0065] Figure 3 Schematic diagram of the structure of a multi-level pulse neuromorphic network in an embodiment of the present invention;

[0066] Figure 4Schematic diagram of the workflow of the gradient descent optimizer in an embodiment of the present invention;

[0067] Figure 5 Schematic diagram of the structure of a distributed error calculation unit in an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0069] It should be clear that the following embodiments of the present disclosure are described by specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other in the absence of conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.

[0070] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein may be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described herein may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described herein may be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein may be used to implement this device and / or practice this method.

[0071] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The drawings only show components related to the present disclosure rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0072] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the aspects described may be practiced without these specific details.

[0073] like Figure 1As shown, the rail transit component size detection device provided by the embodiment of the present invention includes a multi-source sensor array and a processor. The multi-source sensor array includes laser ranging sensors arranged around the detection area, an industrial camera arranged above the detection area, and an ultrasonic sensor array arranged at the bottom of the detection area.

[0074] Specifically,

[0075] Laser distance measuring sensors are evenly arranged around the detection area to collect distance information of components;

[0076] Install a high-resolution industrial camera above the inspection area to collect image information of the components;

[0077] An ultrasonic sensor array is arranged at the bottom of the detection area to collect thickness information of the component.

[0078] The present application embodiment includes the following steps:

[0079] Collecting multi-source sensor data, wherein the multi-source sensor data includes a laser ranging data stream, an image data stream, and an ultrasonic data stream;

[0080] Performing noise filtering and data standardization processing on the laser ranging data stream, performing distortion correction and illumination compensation processing on the image data stream, and performing signal enhancement and denoising processing on the ultrasonic data stream to obtain standardized multi-source sensor data;

[0081] Based on the standardized multi-source sensor data, segmentation calculation is performed through a multi-level pulse neuromorphic network to obtain a segmentation result of the rail transit component, wherein the segmentation result includes contour features and regional features;

[0082] Based on the segmentation results, a gradient descent optimizer is used to perform multi-objective optimization to obtain an optimized size parameter set, wherein the multi-objective optimization includes maximizing measurement accuracy, minimizing calculation time, and minimizing resource consumption;

[0083] Based on the optimized set of dimensional parameters, time difference learning is performed through a distributed error calculation unit to obtain a final dimensional measurement result of the rail transit component.

[0084] The specific implementation process is described below:

[0085] like Figure 2 As shown in Figure 1, the process of collecting and preprocessing multi-source sensor data includes the following steps:

[0086] First, the data of rail transit components are collected through a multi-source sensor array. Among them, the laser ranging sensor collects the distance information of the component, the industrial camera collects the image information of the component, and the ultrasonic sensor array collects the thickness information of the component. In this embodiment, the sampling frequency of the laser ranging sensor is 1kHz, the resolution of the industrial camera is 4096×3072 pixels, and the operating frequency of the ultrasonic sensor array is 40kHz.

[0087] Then, the collected multi-source sensor data is preprocessed. Kalman filtering is used to filter out noise from the laser ranging data stream, and the minimum-maximum normalization method is used to standardize the data; Zhang's camera model is used to correct distortion of the image data stream, and the adaptive histogram equalization algorithm is used for illumination compensation; wavelet transform is used to enhance the signal of the ultrasonic data stream, and median filtering is used for denoising.

[0088] like Figure 3 As shown in FIG. 1 , the specific implementation process of the multi-level pulse neuromorphic network includes:

[0089] In the input layer, preprocessed multi-source sensor data is received. Each input neuron is configured with an adjustable synaptic weight and a dynamic threshold mechanism to adaptively adjust the pulse triggering conditions.

[0090] In the pulse coding layer, the multi-source sensor data is converted into a time domain pulse sequence. Specifically, the laser ranging data is encoded into a pulse sequence according to the size of the distance value. The larger the distance value, the longer the pulse interval; the gray value of the image data is mapped to the frequency of the pulse sequence. The higher the gray value, the faster the pulse frequency; the intensity value of the ultrasonic data is converted into the pulse amplitude. The larger the intensity value, the higher the pulse amplitude.

[0091] In the feature extraction layer, the first layer of neurons mainly extracts basic edge and contour features through time domain correlation analysis; the second layer of neurons extracts complex features such as corners and surfaces through dynamic adjustment of the weights of synaptic connections; the third layer of neurons combines and fuses the extracted features to form a complete component feature representation.

[0092] In the segmentation decision layer, a competitive neural network mechanism is used to accurately divide the component area. When the feature activation of a certain area exceeds the set threshold, the corresponding neuron generates an output response to complete the area segmentation.

[0093] The above process can be described in the following steps:

[0094] Step A1: Construct a multi-level spiking neural network, including:

[0095] Input layer: receives multi-source sensor data preprocessed in step 1;

[0096] Pulse coding layer: converts sensor data into time domain pulse trains;

[0097] Feature extraction layer: use time domain pulses to extract features;

[0098] Segmentation decision layer: Component segmentation is achieved based on the extracted features.

[0099] Step A2: Implement time domain encoding of pulse signals and convert the multi-source sensor data in step 1 into a pulse sequence:

[0100] The laser ranging data is converted into a distance coded pulse sequence PL;

[0101] Image data is converted into grayscale coded pulse sequence PI;

[0102] The ultrasonic data is converted into an intensity-coded pulse sequence PU.

[0103] Step A3: Neuromorphic computing based on time-domain pulse trains:

[0104] Use pulse sequences PL, PI, and PU for feature extraction;

[0105] The preliminary segmentation of component contours and regions is achieved through the collaborative calculation of multi-level neurons;

[0106] The output segmentation result S includes the contour features and regional features of the component.

[0107] In this step A1-A3, the system first inputs the pulse sequences PL, PI and PU from three different sensors into the neuromorphic processing unit. The processing unit uses a bionic neuron model that can simulate the pulse response characteristics of biological neurons. Each neuron is equipped with an adjustable synaptic weight and a dynamic threshold mechanism to adaptively adjust the pulse triggering conditions. When the cumulative potential of the input pulse sequence exceeds the threshold, the neuron generates an output pulse and triggers the calculation of the next layer of neurons.

[0108] In the feature extraction stage, the system adopts a multi-layer neuromorphic network structure. The first layer of neurons is mainly responsible for extracting basic edge and contour features. Through the time domain correlation analysis of the pulse sequence, the basic geometric features of the component can be effectively identified. The second layer of neurons focuses on extracting higher-level structural features, such as corner points, surfaces and other complex features. These features are continuously optimized and refined through the dynamic adjustment of the weights of synaptic connections. The third layer of neurons combines and fuses the extracted features to form a complete component feature representation.

[0109] In the final segmentation calculation stage, the system uses the feature information extracted previously to accurately divide the component area through a special decision layer. This layer uses a competitive neural network mechanism, and different neurons enhance the accuracy of the segmentation boundary through lateral inhibition. When the feature activation of a certain area exceeds the set threshold, the corresponding neuron will produce a strong output response, thereby marking the area as a valid part of the component. This segmentation mechanism based on time domain pulses has strong noise robustness and can effectively cope with various interferences in the actual detection environment.

[0110] The entire neuromorphic computing process is a dynamic evolutionary process. The system gradually optimizes the segmentation results through continuous pulse interaction and weight adjustment. The final output segmentation result S not only contains the precise contour information of the component, but also contains the semantic features of the region, providing a reliable basis for subsequent dimensional measurement. Compared with traditional digital computing methods, this computing paradigm based on the biological nervous system shows better real-time and adaptability when processing time series data and pattern recognition tasks.

[0111] Exemplarily, at the input layer, preprocessed multi-source sensor data is received. Each input neuron is configured with an adjustable synaptic weight and a dynamic threshold mechanism for adaptively adjusting the pulse triggering condition. Taking laser ranging data as an example, when the distance value at a certain moment is d(t), the membrane potential V(t) of the input neuron is updated as follows:

[0112] V(t) = V(t-1) × α + w × d(t)

[0113] Wherein, α is the membrane potential attenuation coefficient, the value range is [0,1], and in this embodiment, it is 0.8; w is the synaptic weight. When V(t) exceeds the threshold θ, the neuron generates a pulse output and resets the membrane potential to 0. The threshold θ is dynamically adjusted according to the statistical characteristics of the input data:

[0114] θ(t) = θ(t-1) + β × (V(t) - V_target)

[0115] Wherein, β is the learning rate, and its value is 0.01; V_target is the membrane potential corresponding to the target average firing rate, which is set to 0.5 in this embodiment.

[0116] In the pulse coding layer, the multi-source sensor data is converted into a time domain pulse sequence. Taking image data as an example, for the gray value I(x,y), its corresponding pulse emission time T(x,y) is encoded as follows:

[0117] T(x,y) = round(Tmax × (1 - I(x,y) / 255))

[0118] Wherein, Tmax is the encoding window length, which is 1000ms in this embodiment; round is the rounding function. The higher the gray value, the earlier the release time, thereby encoding the gray information of the image in the time domain.

[0119] In the feature extraction layer, a hierarchical feature extraction strategy is adopted:

[0120] The first layer of neurons is used for edge feature extraction, using 8-directional Gabor filters as synaptic connection patterns, and the filter parameters are:

[0121] Direction: 0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°;

[0122] Wavelength: 4 pixels;

[0123] Phase: 0° and 90°;

[0124] Gaussian envelope: σ = 2;

[0125] The second layer of neurons is used for structural feature extraction, including:

[0126] Corner detection: Use the improved Harris operator, the response function is;

[0127] R = det(M) - k×trace(M)²

[0128] Where M is a 2×2 autocorrelation matrix, k=0.04;

[0129] Surface features: Classified by the ratio of the principal curvatures k1 and k2;

[0130] If |k1 / k2| < 0.1, it is considered a plane;

[0131] If 0.1 ≤ |k1 / k2| ≤ 0.9, it is determined to be a cylinder;

[0132] If |k1 / k2| > 0.9, it is considered a sphere;

[0133] The third layer of neurons uses the attention mechanism to fuse features:

[0134] Attention(Q,K,V) = softmax(QK^T / d)V

[0135] Where Q, K, V are query, key, and value matrices respectively, and d is the feature dimension.

[0136] In the segmentation decision layer, a competitive neural network mechanism is adopted. For the i-th neuron, its winning probability Pi is calculated as follows:

[0137] Pi = exp(ai / τ) / Σexp(aj / τ)

[0138] Wherein, ai is the activation of the neuron, τ is the temperature parameter, the initial value is set to 1.0 and gradually decreases with the progress of training. When Pi exceeds the threshold (set to 0.8 in this embodiment), the neuron wins and the corresponding area is determined to be part of the target area.

[0139] like Figure 4 As shown in Figure 2, the implementation process of the gradient descent optimizer includes:

[0140] First, the joint optimization objective function is constructed. The measurement accuracy objective function is in the form of mean square error; the computing time objective function is constructed by monitoring the time consumption of key computing nodes; and the resource consumption objective function is established based on hardware resource indicators such as CPU occupancy and memory usage.

[0141] Then, a dynamic gradient field is constructed. In the parameter space, each point corresponds to a specific set of measurement parameter configurations, and the gradient direction indicates the optimal path for parameter optimization. By performing intelligent search in the gradient field, the optimal parameter configuration area can be quickly located.

[0142] Then, multi-objective optimization is achieved through iterative updates. A larger step size is used in the early stage of optimization to speed up convergence, and the step size is automatically reduced when approaching the optimal solution to improve accuracy. At the same time, a dynamic annealing mechanism is introduced to avoid falling into the local optimal solution through random perturbations.

[0143] The above process can be described in the following steps:

[0144] Step B1: Construct a multi-objective optimization model. The optimization objectives include:

[0145] Maximize measurement accuracy;

[0146] Minimize computation time;

[0147] Minimize resource consumption.

[0148] Step B2: Based on the segmentation result S output above, perform parameter optimization using gradient simulation calculation:

[0149] Build a gradient descent optimizer;

[0150] Iteratively update model parameters;

[0151] Output the optimized size parameter set D.

[0152] Specifically, steps B1-B2 take the segmentation result S output by steps A1-A3 as input, construct a multi-objective optimization framework, comprehensively consider multiple optimization objectives such as measurement accuracy, computational efficiency, and resource consumption, and finally output the optimized size parameter set D.

[0153] In the construction phase of the multi-objective optimization model, the system first defines three key optimization objective functions. The measurement accuracy objective function f1 is in the form of mean square error, which is used to evaluate the deviation between the measurement result and the actual size; the calculation time objective function f2 is constructed by monitoring the time consumption of key computing nodes to ensure that the system can meet the real-time requirements; the resource consumption objective function f3 is based on hardware resource indicators such as CPU occupancy rate and memory usage to achieve efficient use of resources. These three objective functions constitute a multi-dimensional optimization space, and the system needs to find the optimal balance point in this space.

[0154] In order to effectively handle this multi-objective optimization problem, the system adopts a gradient-based simulation calculation method. This method first constructs a dynamic gradient field to simulate the changing trend of the optimization target in the parameter space. Each point in the gradient field corresponds to a specific set of measurement parameter configurations, and the gradient direction indicates the optimal path for parameter optimization. By performing intelligent search in this gradient field, the system can quickly locate the optimal parameter configuration area. In particular, the system introduces an adaptive step size mechanism, which uses a larger step size in the early stage of the search to speed up the convergence speed, and automatically reduces the step size to improve accuracy when approaching the optimal solution.

[0155] During the parameter optimization process, the system uses an improved gradient descent algorithm for iterative optimization. Each iteration includes two stages: forward propagation and reverse update. In the forward propagation stage, the system calculates the value of each objective function based on the current parameter configuration and evaluates the overall performance; in the reverse update stage, the system adjusts the parameter configuration to improve performance based on the gradient information of the objective function. In order to avoid falling into the local optimum, the system also introduces a dynamic annealing mechanism to explore a larger solution space through random perturbations. This optimization strategy can not only effectively balance multiple optimization objectives, but also has strong convergence and stability.

[0156] A key innovation in the optimization process is the introduction of a constraint adjustment mechanism. This mechanism can dynamically adjust the weights of each optimization objective according to the actual application scenario, so that the system can maintain good performance under different working conditions. For example, in scenarios with high precision requirements, the system will automatically increase the weight of the measurement accuracy objective; while in scenarios with high real-time requirements, the weight of the computational efficiency objective will be increased accordingly. This flexible weight adjustment strategy enables the system to better adapt to different application requirements.

[0157] Through the above optimization process, the system finally outputs the optimized dimensional parameter set D. This parameter set not only contains the dimensional data of the component, but also contains important information such as measurement confidence and error estimation, providing a reliable basis for subsequent precision optimization. This gradient-based multi-objective optimization method successfully achieves the coordinated optimization of measurement accuracy, computational efficiency, and resource consumption, significantly improving the overall performance of the system.

[0158] The following is a specific example:

[0159] First, a joint optimization objective function is constructed. This objective function comprehensively considers three aspects: measurement accuracy, computing time and resource consumption:

[0160] F(θ) = w1 × Accuracy(θ) + w2 × Time(θ) + w3 × Resource(θ)

[0161] Among them, θ is the optimization parameter vector, which contains the key parameters of the measurement system; w1, w2, and w3 are weight coefficients, and their initial values ​​are set to 0.5, 0.3, and 0.2, respectively, which can be dynamically adjusted according to actual application requirements.

[0162] The measurement accuracy objective function adopts the weighted mean square error form:

[0163] Accuracy(θ) = -1 × (1 / N)Σ(yi - ŷi)²× wi

[0164] Among them, yi is the true value, ŷi is the measured value, wi is the weight of the measurement point, and N is the number of measurement points.

[0165] The computation time objective function is defined as the time cost of the key nodes:

[0166] Time(θ) = -1 × (t_preprocess + t_network + t_postprocess)

[0167] Among them, t_preprocess is the data preprocessing time, t_network is the network calculation time, and t_postprocess is the post-processing time.

[0168] The resource consumption objective function considers CPU and memory usage:

[0169] Resource(θ) = -1 × (α × CPU_usage + β × Memory_usage)

[0170] Among them, α and β are balance coefficients, which are 0.6 and 0.4 respectively.

[0171] Then, a dynamic gradient field is constructed. For any point θt in the parameter space, its gradient is calculated as follows:

[0172] ∇F(θt) = [∂F / ∂θ1, ∂F / ∂θ2, ..., ∂F / ∂θn]

[0173] To improve computational efficiency, a small batch stochastic gradient method is used:

[0174] ∇F_batch(θt) = (1 / M)Σ∇F_m(θt)

[0175] Wherein, M is the batch size, which is 32 in this embodiment.

[0176] In the parameter update stage, an adaptive learning rate strategy with momentum is adopted:

[0177] Step 1: Calculate the momentum term:

[0178] mt = β1 × mt-1 + (1-β1) × ∇F_batch(θt)

[0179] vt = β2 × vt-1 + (1-β2) × (∇F_batch(θt))²

[0180] Among them, β1=0.9, β2=0.999

[0181] Step 2: Bias correction:

[0182] m̂t = mt / (1-β1^t)

[0183] v̂t = vt / (1-β2^t)

[0184] Step 3: Parameter update:

[0185] θt+1 = θt - η × m̂t / (v̂t + ε)

[0186] Among them, η is the basic learning rate, and its initial value is set to 0.001; ε is the stability coefficient, and its value is 1e-8.

[0187] At the same time, a dynamic annealing mechanism is introduced, and the learning rate decays as follows:

[0188] η_t = η_0 × (1 + γ × t)^(-0.5)

[0189] Among them, γ is the annealing coefficient, and its value is 0.0001.

[0190] To avoid falling into the local optimum, random perturbations are injected with probability p in each iteration:

[0191] θt = θt + N(0, σ²)

[0192] Where p = 0.1, σ = 0.01 × (1-t / T), and T is the total number of iterations.

[0193] The conditions for judging convergence are:

[0194] ||∇F(θt)|| < ε1 or |F(θt) - F(θt-1)| < ε2

[0195] Among them, ε1=1e-5, ε2=1e-6.

[0196] In practical applications, the optimizer can effectively balance three objectives:

[0197] Measurement accuracy: relative error is controlled within 0.1%;

[0198] Computation time: single optimization iteration time is less than 10ms;

[0199] Resource consumption: The peak CPU usage does not exceed 80%.

[0200] like Figure 5 As shown, the implementation process of the distributed error calculation unit includes:

[0201] First, multiple parallel error calculation units are constructed. The shape error unit is responsible for the geometric deviation analysis of the component contour; the size error unit is responsible for the accuracy evaluation of key size parameters such as length, width, and height; and the posture error unit is responsible for the position and direction error analysis of the component in three-dimensional space.

[0202] Then, the distributed propagation of error signals is realized. Each error calculation unit is connected to each other through a high-speed communication network, which can not only calculate the local error independently, but also sense the error status of other units.

[0203] Next, the measurement accuracy is optimized based on time difference learning. The system maintains a time series database of historical measurement results, and identifies and handles possible measurement anomalies by analyzing the law of error changes over time.

[0204] Finally, the measurement results are output. In addition to the dimensional data of the component, it also includes measurement confidence information, error estimation information, and a complete report containing measurement quality evaluation indicators.

[0205] In a specific application scenario, the rail transit component size detection device of the present invention is used for the quality inspection of high-speed railway sleepers. Experimental results show that under standard working conditions, the measurement accuracy of the system can reach ±0.1mm, the single complete detection time is less than 1 second, and the average CPU occupancy rate is kept below 60%. Compared with traditional detection methods, the detection efficiency is improved by 200%, the measurement accuracy is improved by 50%, and the false detection rate is reduced by 80%.

[0206] The above process can be expressed in the following steps:

[0207] Step C1: Design a distributed error feedback network, including:

[0208] Construct multiple parallel error calculation units;

[0209] Realize distributed propagation of error signals.

[0210] Step C2: Optimizing measurement accuracy based on time difference learning:

[0211] Calculate the time difference between the current measurement result and the historical best result;

[0212] Update learning parameters;

[0213] Output the final component size measurement result R.

[0214] Among them, step C takes the optimized dimensional parameter set D output by step B as input, realizes continuous optimization of measurement accuracy through distributed error signal processing and time difference learning mechanism, and finally outputs high-precision component dimensional measurement result R.

[0215] In the design of the distributed error feedback network, the system innovatively adopts a hierarchical error calculation unit structure. Each error calculation unit is equipped with a dedicated data processing module, which is responsible for error analysis in a specific dimension. Among them, the shape error unit focuses on the geometric deviation analysis of the component contour; the size error unit is responsible for the accuracy evaluation of key size parameters such as length, width, and height; the posture error unit focuses on the position and direction errors of the component in three-dimensional space. These distributed error calculation units are interconnected through a high-speed communication network to form a collaboratively optimized whole. Each unit can not only independently calculate the local error, but also perceive the error status of other units, thereby achieving optimal control of the global error.

[0216] The temporal difference learning mechanism is another core innovation of this step. Based on the traditional reinforcement learning framework, the system introduces time series correlation analysis. Specifically, the system maintains a time series database of historical measurement results, recording the measurement results at different time points and their corresponding error information. By conducting in-depth analysis of these time series data, the system can identify the patterns of measurement errors over time. For example, some errors may be related to the diurnal changes in ambient temperature, and some errors may be related to the working hours of the equipment. This recognition of time series correlation provides the system with richer optimization basis.

[0217] In the actual learning process, the system uses an improved time difference algorithm to update parameters. First, the system calculates the time difference between the current measurement result and the historical optimal result. This difference includes not only the numerical deviation, but also the information of trend changes. Then, based on this difference information, the system constructs a time difference reward function. The special thing about this reward function is that it not only considers the reduction of immediate error, but also pays attention to the continuity and stability of error improvement. This design enables the system to avoid short-term over-optimization and achieve a more stable and lasting performance improvement.

[0218] In order to improve learning efficiency and robustness, the system also introduces an adaptive learning rate adjustment mechanism. When the system detects that the measurement error shows a significant downward trend, it will appropriately increase the learning rate to speed up the optimization; when the error changes tend to be flat, it will automatically reduce the learning rate to avoid over-adjustment. At the same time, the system has also established an anomaly detection mechanism that can promptly identify and handle possible measurement anomalies to ensure the stability of the learning process.

[0219] The final output component size measurement result R not only contains high-precision size data, but also comes with a complete quality assessment report. The report records in detail the various indicators of the measurement process, including key information such as measurement confidence, error distribution, and optimization trajectory. This information can not only be used to evaluate the reliability of the current measurement, but can also be stored as historical data for continuous optimization and performance improvement of the system. This method based on distributed error signals and time difference learning successfully realizes the adaptive optimization of measurement accuracy and significantly improves the reliability and stability of the system.

[0220] In particular, this step also implements a real-time verification and feedback mechanism for the measurement results. The system will compare the final measurement results with the preset standard specifications. If a significant deviation is found, the alarm mechanism will be immediately triggered and the automatic calibration procedure will be started. This closed-loop quality control mechanism ensures that the measurement system can always maintain the best working condition and provide reliable guarantee for the quality inspection of rail transit components.

[0221] Exemplarily, the implementation process of the distributed error calculation unit includes:

[0222] First, build multiple parallel error calculation units, including shape error unit, size error unit and posture error unit:

[0223] 1. Shape error unit:

[0224] For the analysis of geometric deviations of component contours, the Hausdorff distance is used for evaluation:

[0225] dH(X,Y) = max{sup(inf(d(x,y))), sup(inf(d(x,y)))}

[0226] x∈X y∈Y y∈Y x∈X

[0227] Among them, X is the measured contour point set, Y is the standard contour point set, and d(x,y) is the Euclidean distance.

[0228] Shape similarity score calculation:

[0229] Score_shape = exp(-dH(X,Y) / σ)

[0230] Here, σ is a scale parameter, and in this embodiment, its value is 1.0.

[0231] 2. Dimension error unit:

[0232] Calculate the relative errors of the three dimensions of length (L), width (W), and height (H) respectively:

[0233] εL = |Lm - Ls| / Ls

[0234] εW = |Wm - Ws| / Ws

[0235] εH = |Hm - Hs| / Hs

[0236] The subscript m represents the measured value, and s represents the standard value.

[0237] Comprehensive dimensional error score:

[0238] Score_dim = exp(-λ × (wL×εL + wW×εW + wH×εH))

[0239] Among them, λ is the normalization coefficient, which is 5.0; wL, wW, and wH are weight coefficients, which are dynamically allocated according to the component type.

[0240] 3. Attitude error unit:

[0241] Quaternions are used to represent component posture and calculate posture deviation:

[0242] δq = qm⊗qs^(-1)

[0243] Among them, qm is the measured attitude quaternion, qs is the standard attitude quaternion, and ⊗ represents quaternion multiplication.

[0244] Attitude Error Score:

[0245] Score_pose = (1 + cos(θ)) / 2

[0246] where θ = 2arccos(|δq_w|), where δq_w is the scalar part of δq.

[0247] Then, the distributed propagation of the error signal is realized. Each error calculation unit is connected to each other through a ring topology structure, using an asynchronous communication mechanism:

[0248] 1. Local error calculation:

[0249] Each unit independently computes a local error statistic:

[0250] μi = (1 / K)ΣScore_i_k

[0251] σi² = (1 / K)Σ(Score_i_k - μi)²

[0252] Among them, i represents the unit type and k represents the most recent K measurement results.

[0253] 2. Error information broadcast:

[0254] Each unit periodically broadcasts a message to neighboring units containing the following information:

[0255] Local error statistics (μi, σi²);

[0256] Anomaly detection flag (flag_anomaly);

[0257] Timestamp;

[0258] 3. Error information fusion:

[0259] After receiving the error information of adjacent units, the weighted average method is used for fusion:

[0260] μ_global = Σ(wi × μi)

[0261] Among them, the weight wi is dynamically adjusted according to the new and old timestamps:

[0262] wi = exp(-Δti / τ) / Σexp(-Δtj / τ)

[0263] Among them, Δti is the time difference, τ is the time constant, and its value is 1000ms.

[0264] Next, the measurement accuracy is optimized based on time difference learning:

[0265] 1. Maintaining the historical database:

[0266] For each measurement result (R_t), record the following information:

[0267] Three types of error scores (Score_shape, Score_dim, Score_pose);

[0268] Global error statistics (μ_global);

[0269] System configuration parameters (Config_t);

[0270] 2. Time difference analysis:

[0271] Calculate the state difference between adjacent moments:

[0272] ΔR_t = R_t - R_(t-1)

[0273] ΔScore_t = Score_t - Score_(t-1)

[0274] 3. Parameter update strategy:

[0275] Adjust system parameters according to the difference results:

[0276] If ΔScore_t > 0:

[0277] Config_(t+1) = Config_t + α × ΔConfig_t

[0278] Else:

[0279] Config_(t+1) = Config_t - β × ΔConfig_t

[0280] Among them, α is the positive learning rate (0.1) and β is the negative learning rate (0.05).

[0281] Finally, the measurement result report is output, which contains the following contents:

[0282] 1. Dimensions:

[0283] Main size parameters and their confidence intervals;

[0284] Statistical distribution characteristics (mean, variance, skewness, kurtosis);

[0285] 2. Reliability indicators:

[0286] Measure confidence: Conf = min(Score_shape, Score_dim, Score_pose)

[0287] System stability: Stab = 1 - σ_global / μ_global

[0288] Anomaly probability: P_anomaly = Count(flag_anomaly) / N

[0289] 3. Quality Assessment:

[0290] Radar chart of scores in each dimension;

[0291] Time series trend analysis;

[0292] Anomaly marking and cause analysis.

[0293] The present application also provides a component size detection device for rail transit, comprising:

[0294] Multi-source sensor array, including:

[0295] The laser distance measuring sensors arranged around the detection area are used to collect the distance information of the rail transit components.

[0296] An industrial camera is arranged above the detection area, and is used to collect image information of the rail transit components.

[0297] An ultrasonic sensor array disposed at the bottom of the detection area is used to collect thickness information of the rail transit component;

[0298] A processor is used to execute the component size detection method for rail transit as described above.

[0299] The system of the embodiments of the present disclosure can execute the method provided by the embodiments of the present disclosure, and the implementation principles are similar. The actions performed by each module in the system of each embodiment of the present disclosure correspond to the steps in the method of each embodiment of the present disclosure. For the detailed functional description of each module of the system, please refer to the description in the corresponding method shown in the previous text, which will not be repeated here.

[0300] The above is only an optional implementation method for some implementation scenarios of the present disclosure. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present disclosure, other similar implementation methods based on the technical ideas of the present disclosure are also within the protection scope of the embodiments of the present disclosure.

Claims

1. A component size detection method for rail transit, characterized in that , including the following steps: Collecting multi-source sensor data, wherein the multi-source sensor data includes a laser ranging data stream, an image data stream, and an ultrasonic data stream; Performing noise filtering and data standardization processing on the laser ranging data stream, performing distortion correction and illumination compensation processing on the image data stream, and performing signal enhancement and denoising processing on the ultrasonic data stream to obtain standardized multi-source sensor data; Based on the standardized multi-source sensor data, segmentation calculation is performed through a multi-level pulse neuromorphic network to obtain a segmentation result of the rail transit component, wherein the segmentation result includes contour features and regional features; Based on the segmentation results, a gradient descent optimizer is used to perform multi-objective optimization to obtain an optimized size parameter set, wherein the multi-objective optimization includes maximizing measurement accuracy, minimizing calculation time, and minimizing resource consumption; Based on the optimized size parameter set, a distributed error calculation unit is used to perform time difference learning to obtain a final size measurement result of the rail transit component; The step of performing segmentation calculation by a multi-level pulse neuromorphic network includes: converting the standardized multi-source sensor data into a time domain pulse sequence, including: Converting the laser ranging data stream into a distance coded pulse sequence, converting the image data stream into a grayscale coded pulse sequence, and converting the ultrasonic data stream into an intensity coded pulse sequence; The time-domain pulse sequence is used to perform feature extraction through the multi-level pulse neuromorphic network; based on the result of the feature extraction, the contour and region segmentation of the rail transit component is achieved.

2. The method according to claim 1, characterized in that The multi-level pulse neuromorphic network includes: an input layer for receiving the standardized multi-source sensor data, a pulse coding layer for converting the multi-source sensor data into a time-domain pulse sequence, a feature extraction layer for extracting features using the time-domain pulse sequence, and a segmentation decision layer for implementing component segmentation based on the result of the feature extraction.

3. The method according to claim 1, characterized in that , the use of the time domain pulse sequence to extract features through the multi-level pulse neuromorphic network includes: extracting basic edge and contour features through the first layer of neurons of the multi-level pulse neuromorphic network; extracting structural features of corner points and surfaces through the second layer of neurons of the multi-level pulse neuromorphic network; combining and fusing the extracted features through the third layer of neurons of the multi-level pulse neuromorphic network to form a complete feature representation of the rail transit component.

4. The method according to claim 1, characterized in that ,The steps of using the gradient descent optimizer for multi-objective optimization include: constructing a joint optimization objective function of measurement accuracy, computing time and resource consumption; building a gradient descent optimizer based on the joint optimization objective function; realizing multi-objective optimization by iteratively updating the parameters of the gradient descent optimizer.

5. The method according to claim 4, characterized in that , the multi-objective optimization is achieved by iteratively updating the parameters of the gradient descent optimizer, including: performing forward propagation calculations, calculating the values ​​of each objective function based on the current parameter configuration, and evaluating the overall performance of each objective function; performing reverse updates, adjusting the parameter configuration according to the gradient information of the objective function; introducing a dynamic annealing mechanism, exploring a larger solution space through random perturbations, and avoiding falling into a local optimal solution.

6. The method according to claim 1, characterized in that The step of performing time difference learning through a distributed error calculation unit includes: constructing multiple parallel error calculation units; calculating the time difference between the current measurement result and the historical optimal result; updating the learning parameters based on the time difference; and outputting the final size measurement result of the rail transit component.

7. The method according to claim 6, characterized in that The method further includes: performing real-time verification on the final size measurement result; and triggering an alarm and starting an automatic calibration procedure when a significant deviation is detected in the final size measurement result.

8. The method according to claim 6, characterized in that The output of the final dimensional measurement result of the rail transit component includes: outputting various dimensional data of the rail transit component; outputting measurement confidence information; outputting measurement error estimation information; and generating a measurement report containing measurement quality evaluation indicators.

9. A component size detection device for rail transit, characterized in that , comprising: a multi-source sensor array, comprising: laser ranging sensors arranged around a detection area, for collecting distance information of the rail transit component, an industrial camera arranged above the detection area, for collecting image information of the rail transit component, an ultrasonic sensor array arranged at the bottom of the detection area, for collecting thickness information of the rail transit component; a processor, for executing the component size detection method for rail transit as described in any one of claims 1 to 8.

Images