Guide rail precision evaluation method and system based on multi-sensor fusion

Abstract

The invention relates to the technical field of multi-sensor information fusion and intelligent manufacturing, and discloses a guide rail precision evaluation method and system based on multi-sensor fusion, and the method comprises the steps: constructing a multi-modal sensor value quantification model based on information entropy, calculating the contribution degree of sensor information and constructing a multi-sensor mutual information coupling model; constructing a self-optimization sensor network configuration system, abstracting a multi-modal sensor system into a graph structure, and generating optimal sensor combination configuration by applying a graph reinforcement learning algorithm; constructing a multi-time scale optimization framework and realizing adaptive data compression based on Huffman coding; constructing a dynamic Bayesian network model to fuse heterogeneous sensor data; constructing a self-evolution evaluation strategy system, and continuously improving an evaluation strategy through historical performance data; through information entropy-driven resource allocation, resources can be dynamically allocated according to the information value of the sensor.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-sensor information fusion and intelligent manufacturing, and more specifically, to a guide rail accuracy evaluation method and system based on multi-sensor fusion. Background Art

[0002] As a key component in high-end equipment manufacturing and precision machining, guide rail accuracy assessment is crucial for ensuring equipment performance and product quality. With the advancement of intelligent manufacturing and automation, guide rail accuracy assessment systems are gradually moving towards multi-sensor and multi-modal fusion. However, existing technologies still have many deficiencies in practical applications, primarily in the following areas: Existing guide rail accuracy assessment systems generally employ a fixed sensor configuration, lacking the ability to dynamically adjust resource allocation based on actual operating conditions and accuracy requirements. In scenarios requiring high accuracy, the system often requires all sensors to be enabled simultaneously and run at a high sampling rate, resulting in increased energy and resource consumption. In scenarios requiring moderate accuracy, however, this fixed configuration results in poor resource utilization, making it difficult to achieve an effective balance between resources and accuracy. For example, some commercial guide rail inspection equipment operates all sensors at full power in all operating states, resulting in energy waste. As the number of sensors increases, the management and maintenance complexity of traditional fixed-configuration sensor systems increases exponentially, making them difficult to scale to meet the demands of large-scale systems. Adding new sensors or changing sensor types often requires manual redesign of data fusion strategies and system reconfiguration, a time-consuming process that compromises system flexibility and maintainability. Existing technologies struggle to balance system optimization with real-time response. While simple data fusion methods can meet certain real-time requirements, they suffer from limitations in accuracy and robustness. While complex optimization algorithms can improve assessment accuracy, they suffer from high computational overhead, making them difficult to meet the demands of online monitoring and real-time feedback. For example, some high-precision guide rail assessment systems require several hours of computing time to complete a full assessment, making it impossible to monitor the guide rail status in real time. Existing guide rail accuracy assessment systems generally lack adaptive mechanisms and are unable to automatically adjust assessment strategies based on different working conditions, guide rail types, and accuracy requirements, resulting in poor system versatility and scalability. Whenever the application scenario changes, the system often needs to be redeveloped and reconfigured, increasing operational and maintenance costs. Traditional methods have difficulty quantifying the actual contribution of different sensor data to accuracy assessments, making it difficult to achieve intelligent resource allocation based on the value of information. This results in a lack of scientific basis for resource allocation and management, making it impossible to fully utilize the synergistic effects of various sensors, and affecting the overall assessment effect.

[0003] In summary, the existing guide rail accuracy assessment technology has problems that need to be improved in terms of dynamic resource allocation, system scalability, real-time performance, adaptability, and information value quantification. It is urgent to propose a new guide rail accuracy assessment method that can realize multimodal information collaborative optimization, self-configuration, and intelligent resource management to meet the higher requirements for guide rail accuracy assessment in the fields of intelligent manufacturing and high-end equipment. Summary of the Invention

[0004] The present invention provides a guide rail accuracy evaluation method and system based on multi-sensor fusion, which solves technical problems in related technologies such as inefficient resource allocation, poor real-time performance and insufficient adaptability.

[0005] The present invention provides a guide rail accuracy evaluation method based on multi-sensor fusion, comprising: Construct a multimodal sensor value quantification model based on information entropy, calculate the sensor information contribution and construct a multi-sensor mutual information coupling model; Based on the output of the mutual information model, a self-optimizing sensor network configuration system is constructed. The multimodal sensor system is abstracted into a graph structure and a graph reinforcement learning algorithm is applied to generate the optimal sensor combination configuration. Based on the optimal sensor combination configuration, multi-level dynamic resource allocation is achieved, a multi-time scale optimization framework is constructed, and adaptive data compression based on Huffman coding is implemented; Based on compressed and optimized data, a dynamic Bayesian network model is constructed to fuse heterogeneous sensor data and achieve adaptive adjustment of the network structure; Based on the adaptive adjustment results, a self-evolving evaluation strategy system is constructed to continuously improve the evaluation strategy through historical performance data.

[0006] Furthermore, the step of calculating the contribution of sensor information includes: Obtain the probability distribution of guide rail accuracy parameters in the historical data set and calculate its entropy value; Calculate the conditional entropy of the guide rail accuracy under the condition of known sensor data; The information contribution of the sensor is calculated by the difference between the entropy value and the conditional entropy.

[0007] Furthermore, the step of constructing a sensor multi-sensor mutual information coupling model includes: Calculate mutual information between sensor pairs and identify redundant sensor groups and complementary sensor groups; Calculate high-order mutual information between sensors to capture complex nonlinear complementary relationships; A resource allocation function is established to associate the contribution of sensor information with resource consumption.

[0008] Furthermore, the step of applying the graph reinforcement learning algorithm to optimize the sensor configuration includes: The sensor topology network is represented as a graph structure, where nodes represent sensors and edges represent relationships between sensors; Construct a graph neural network as a function approximator to achieve node feature aggregation and node feature update; A deep Q-learning framework is used for reinforcement learning training, and the sensor configuration is optimized by maximizing the balance objective function between information gain and resource consumption.

[0009] Furthermore, the multi-level dynamic resource allocation includes optimization of the following time scales: Microsecond-level sensor adaptive sampling optimization; Millisecond-level data processing strategy adjustment; Second-level optimization of computing resource allocation; Minute-level sensor network topology reconstruction.

[0010] Furthermore, the step of implementing adaptive data compression based on Huffman coding includes: Count the occurrence probabilities of each value of sensor data; Calculate the negative logarithm of the probability of occurrence of the data value as its encoding length; Construct a Huffman tree to assign variable length codes to each data value; Dynamically update the Huffman tree as data distribution changes.

[0011] Furthermore, the step of constructing a dynamic Bayesian network model includes: Construct a three-layer network structure including sensor layer, feature layer and accuracy parameter layer; Establish four dependency relationships: intra-layer same-time dependency, inter-layer same-time dependency, intra-layer time dependency, and inter-layer time dependency; A network structure adaptive adjustment algorithm is designed based on the conditional information gain principle.

[0012] Furthermore, the steps of constructing the self-evolving evaluation strategy system include: Establish a policy library and performance evaluation mechanism to store evaluation strategies for different scenarios; Implement strategy optimization and evolution mechanisms to continuously improve evaluation strategies through historical performance data; Design a strategy adaptive selection algorithm based on scenario similarity to select the evaluation strategy that best suits the current scenario from the strategy library.

[0013] Furthermore, the resource allocation function quantifies the resource utilization efficiency of the sensor by calculating the ratio of sensor information contribution to resource consumption, and preferentially configures high-efficiency sensors according to the resource efficiency ratio.

[0014] The present invention provides a guide rail accuracy evaluation system based on multi-sensor fusion, which is used to execute the above-mentioned guide rail accuracy evaluation method based on multi-sensor fusion, including: Information entropy quantification module, used to build a multimodal sensor value quantification model based on information entropy; Sensor network configuration module, used to build a self-optimizing sensor network configuration system; Resource allocation module, used to implement multi-level dynamic resource allocation; Data fusion module, used to build a dynamic Bayesian network model to fuse heterogeneous sensor data; Evaluation strategy module, used to build a self-evolving evaluation strategy system.

[0015] The beneficial effects of the present invention are: through information entropy-driven resource allocation, the system can dynamically allocate resources according to the value of sensor information, reduce the use of redundant sensors, reduce system energy consumption, and at the same time increase sensor density in key areas to ensure evaluation accuracy; Based on a self-optimizing sensor network and self-evolving evaluation strategy, the system can automatically adjust the evaluation strategy according to different guide rail types, working conditions and accuracy requirements, improving its ability to adapt to changing working conditions and supporting a unified evaluation platform for multiple types of guide rails. Through multi-level dynamic resource allocation and the edge-fog-cloud three-layer computing architecture, the system maintains comparable assessment accuracy to traditional methods while improving data processing speed, achieving near-real-time accuracy assessment and meeting online monitoring needs. The self-configuration mechanism reduces the exponential growth of sensor network management complexity in traditional systems to linear growth, shortens the time to integrate new sensors from days to hours, reduces the need for manual intervention, and simplifies system maintenance and expansion. Through strategy evolution and experience accumulation, system performance continues to improve with extended runtime, and evaluation efficiency increases logarithmically, forming a virtuous evolutionary cycle. The present invention exhibits emergent properties such as "dynamic balance" and "information resonance", which can automatically maintain the optimal balance between information and resources when operating conditions change, and generate unexpected information value gains through specific sensor combinations. These properties cannot be achieved in traditional fixed configuration systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of a guide rail accuracy evaluation method based on multi-sensor fusion in the present invention; Figure 2 It is a bar chart comparing the contribution of sensor information; Figure 3 is a dendrogram of the mutual information distribution of sensor combinations; Figure 4It is a grouped histogram comparing optimization performance at multiple time scales; Figure 5 It is a line chart of the adaptive data compression effect based on Huffman coding; Figure 6 It is a radar chart comparing the performance of multimodal information entropy optimization systems. DETAILED DESCRIPTION

[0017] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0018] At least one embodiment of the present invention discloses a guide rail accuracy evaluation method based on multi-sensor fusion, such as Figure 1 Shown, including: Step 1: Construct a multimodal sensor value quantification model based on information entropy, calculate the sensor information contribution and construct a multi-sensor mutual information coupling model; This step uses information entropy theory to construct a multimodal sensor data value quantification model to accurately quantify the contribution of different sensors and their combinations to the guide rail accuracy assessment; this step includes the following sub-steps.

[0019] Step 1.1, calculate the contribution of sensor information; For each sensor in the guide rail accuracy evaluation system , the information entropy principle is used to calculate its information contribution, that is, the contribution of the sensor to reducing the uncertainty of guide rail accuracy assessment; Before calculating sensor information contribution, the system preprocesses the data from different sensors. Position sensor data is typically expressed in millimeters or microns, acceleration sensor data in seconds per meter squared, temperature sensor data in degrees Celsius, and vibration sensor data in frequency or amplitude. These data, with varying dimensions and numerical ranges, must first be normalized to bring all sensor data into the same numerical range, ensuring fairness and accuracy in calculating information contribution.

[0020] Specifically, the sensor information contribution is calculated as follows: ; in, Indicates sensor Information contribution, The entropy value of the guide rail accuracy represents the uncertainty of the guide rail accuracy assessment. Indicates that the sensor is known The conditional entropy of the guide rail accuracy under the data condition represents the uncertainty after using the sensor; the higher the information contribution, the greater the value of the sensor for accuracy assessment.

[0021] In actual application, the system first obtains the guide rail accuracy parameters from the historical data set The probability distribution of , calculate its entropy , then for each sensor , calculate the conditional entropy after obtaining its data , and the difference between the two is the information contribution of the sensor.

[0022] It should be understood that in some embodiments, the system may use mutual information Instead of information contribution, it is calculated in the same way, but interpreted from the perspective of information theory as the mutual information between sensor data and accuracy parameters.

[0023] like Figure 2 As shown in the figure, different sensor types contribute varying amounts of information to guide rail accuracy assessment. The figure shows that position sensors and force sensors have the highest information contributions, while temperature sensors contribute relatively little. This result validates the rationale and effectiveness of this application's use of information entropy theory to quantify the value of different sensors.

[0024] Step 1.2, construct the sensor multi-sensor mutual information coupling model; To quantify the information redundancy or complementary relationship between multiple sensors, this application constructs a multi-sensor mutual information coupling model. This model identifies redundant sensor groups and complementary sensor groups by calculating the mutual information between sensor combinations. The specific mutual information calculation is as follows: ; in, Indicates sensor and sensors About guide rail accuracy The mutual information of Indicates sensor Guide rail accuracy Information contribution, Indicates sensor Guide rail accuracy Information contribution, Indicates sensor and sensors Combined effect on guide rail accuracy Information contribution; when When the sensor and sensors There is information redundancy; when When the sensor and sensors There is complementarity, and combined use can produce additional information gain.

[0025] Through this mutual information model, the system can identify and quantify the synergy between sensors, providing a basis for subsequent optimization of sensor configuration.

[0026] It should be noted that, in some embodiments, the system can also calculate high-order mutual information between three or more sensors to capture complex nonlinear complementary relationships, such as: ; in, Indicates sensor ,sensor and sensors About guide rail accuracy The third-order mutual information of Indicates sensor and sensors About guide rail accuracy The mutual information of Indicates sensor and sensors As a whole with the sensor About guide rail accuracy The mutual information of Indicates sensor and sensors About guide rail accuracy The mutual information of Indicates sensor and sensors About guide rail accuracy The third-order mutual information can capture the sensor ,sensor and sensors The complex nonlinear relationships between them are used to discover synergistic effects that cannot be identified by second-order mutual information.

[0027] like Figure 3 The figure shows the distribution of mutual information between different sensor combinations. The size of the area represents the information value of each sensor and its combined effect. It can be intuitively seen that certain sensor combinations (such as acceleration + vibration) have high complementarity and generate information value greater than the sum of their individual contributions. This verifies the "information resonance" phenomenon mentioned in this application, that is, the information value generated by a specific sensor combination far exceeds the sum of its individual contributions.

[0028] Step 1.3, establish a balance model between information entropy and resource efficiency; Associate information value with resource consumption and establish a resource allocation function.

[0029] Before establishing a resource allocation function, the system preprocesses data for various heterogeneous resource metrics. Information contribution is a dimensionless probability value typically between 0 and 1. Computational resource consumption may be expressed as a percentage of CPU utilization, megabytes of memory occupied, or milliseconds of processing time. Energy consumption is typically measured in watts or joules. These metrics have different dimensions and numerical ranges. Directly calculating their ratios can lead to over-inflating or under-inflating certain metrics. Therefore, it is necessary to standardize each resource consumption metric and convert it into dimensionless values ​​on the same scale to effectively calculate resource efficiency ratios.

[0030] The resource allocation function is implemented as follows: First, the information contribution of each sensor to accuracy assessment is obtained. Then, its computing resource and energy consumption are calculated separately, and weighted coefficients are set based on the actual application scenario. The information contribution is used as the numerator, and the weighted sum of resource consumption is used as the denominator. The two are divided to obtain the resource efficiency ratio. The weighted coefficients can be adjusted based on the importance attached to computing resources, energy, and time. Finally, the system prioritizes sensors based on this ratio to achieve optimal resource allocation.

[0031] The system evaluates the resource efficiency of the sensor by calculating the ratio of sensor information contribution to resource consumption. Specifically, the information contribution value of the sensor to the guide rail accuracy assessment is first obtained as the numerator. Then, the computing resource consumption value and energy consumption value of the sensor are obtained, multiplied by the corresponding weight coefficients, and added as the denominator. Finally, the numerator is divided by the denominator to obtain the resource efficiency ratio of the sensor. The weight coefficient can be adjusted according to the relative importance of computing resources and energy in the application scenario. The higher the resource efficiency ratio, the greater the value of the information provided by the sensor per unit resource consumption, and the system will give priority to configuring sensors with high resource efficiency.

[0032] This resource allocation function quantifies the resource utilization efficiency of the sensor by calculating the information contribution under unit resource consumption, providing a quantitative basis for resource optimization allocation; the system can sort the resource efficiency of each sensor according to this function and give priority to configuring high-efficiency sensors.

[0033] Optionally, in some implementations, the resource allocation function may be extended to take time cost into consideration.

[0034] When calculating resource efficiency, the system not only considers computing resource consumption and energy consumption, but also takes the time required to acquire and process sensor data as an additional factor. Specifically, in the denominator of resource consumption, in addition to computing resource consumption multiplied by the first weight coefficient and energy consumption multiplied by the second weight coefficient, the time cost multiplied by the third weight coefficient is also added. The third weight coefficient is adjusted according to the importance of the time factor in the application scenario, and can be appropriately increased in scenarios with high real-time requirements.

[0035] Step 2: Based on the output of the mutual information model, a self-optimizing sensor network configuration system is constructed. The multimodal sensor system is abstracted into a graph structure and a graph reinforcement learning algorithm is applied to generate the optimal sensor combination configuration. Based on the information entropy quantification model established in step 1, this step constructs a self-optimizing sensor network configuration system to achieve automatic optimization configuration of the sensor network; this step includes the following sub-steps.

[0036] Step 2.1, construct the sensor topology network model; Abstract the multimodal sensor system into a graph structure: ; in Represents the multimodal sensor system graph structure, represents the set of sensor nodes, Represents the set of relationship edges between sensors; The weight of the edge is determined by the degree of information complementarity or redundancy calculated by the mutual information model; through this topological model, the system can intuitively represent the sensor network structure and the relationship between sensors, providing a basis for network optimization.

[0037] In specific implementation, the system first initializes the sensor topology map, treating each sensor as a node in the map. Then, based on the mutual information value calculated in step 1.2, it establishes edge connections between nodes. The edge weight is set to the absolute value of the mutual information. Edges with positive mutual information (indicating redundancy) are marked red, and edges with negative mutual information (indicating complementarity) are marked green, which facilitates the system to perform visual analysis and optimize decisions.

[0038] Step 2.2, apply graph reinforcement learning algorithm to optimize sensor configuration; The graph reinforcement learning algorithm is applied to optimize the sensor topology network and automatically generate the optimal sensor combination configuration. The algorithm models the sensor configuration problem as a Markov decision process, and its objective function is: ; in is the objective function, which represents the expected cumulative benefit; is the expectation operator, which represents the expected value of all possible state transitions; For the sum symbol; is the termination time step of the decision process; is a discount factor used to balance the weights of current and future rewards; For the The information gain provided by the combination of sensors at each moment; The weight coefficient for balancing information gain and resource consumption; For the Resource consumption at all times.

[0039] The system finds the optimal balance between information gain and resource consumption by maximizing the objective function.

[0040] The specific implementation of the graph reinforcement learning algorithm is as follows.

[0041] First, the system represents the sensor topology network as a graph structure: ; in It is a graph structure representation of the sensor topology network; is a node set, representing all sensor nodes; is an edge set, representing the relationship edges between sensors; It is the node feature matrix, and each row contains the sensor’s type, location, accuracy, energy consumption and other attributes.

[0042] Before constructing the node feature matrix, the system preprocesses data for various sensor attributes. Because sensor types are categorical data (such as position sensors, accelerometers, and temperature sensors), sensor positions are continuous numerical data that may be represented as coordinates, sensor accuracy is typically expressed as a percentage or absolute error value, and sensor energy consumption is expressed in units of power, categorical data requires encoding conversion to convert non-numeric data such as sensor type into numerical data, such as using one-hot encoding or label encoding. Furthermore, because the numerical ranges of various attributes vary significantly, normalization is required to ensure the stability and convergence of graph neural network training.

[0043] Then, the system constructs a graph neural network (GNN) as a function approximator, whose structure includes multiple layers of graph convolutional layers and fully connected layers; the information transmission mechanism of the graph convolutional layer is as follows.

[0044] Node feature aggregation: For nodes , from its neighbor nodes Information collected: ; in For slave nodes To Node Information messages; For the message function, define how to aggregate information; For the Layer Node The eigenvector of For the Layer Node The eigenvector of For nodes and The edge features between .

[0045] Message Function The specific implementation is to integrate the characteristics of the current node, the characteristics of neighboring nodes, and the characteristics of the edges between the two nodes to generate a message for information transmission. Typically, the integration method is to splice the various features and input them into a multi-layer perceptron or linear transformation, and output a new set of feature vectors for subsequent node state updates.

[0046] Node feature update: Update node features based on aggregated information: ; in For the Layer Node The eigenvector of is the feature update function; For the Layer Node The eigenvector of For the sum symbol; For nodes The set of neighbor nodes of For slave nodes To Node Information message.

[0047] Feature update function The specific implementation is: integrate the features of the current node with the messages transmitted by its neighboring nodes, usually using splicing, weighted summation or gating mechanism, and input the integrated results into a nonlinear transformation (such as activation function) to obtain the feature representation of the node in the new layer.

[0048] Based on the constructed graph neural network, the algorithm uses the Deep Q-Network (DQN) framework for reinforcement learning training; the specific steps are as follows.

[0049] State representation: The current sensor network configuration state is represented as a node feature matrix, which includes parameters such as the switch state and sampling frequency of each sensor.

[0050] Action definition: The action space includes: Activate / deactivate specific sensors; Adjust the sensor sampling frequency; Adjust the data sharing configuration between sensors.

[0051] Reward calculation: Calculate the reward based on the information gain and resource consumption of the current configuration: ; in For the The reward value at the moment; For the Information gain at each moment; is the weight coefficient, which adjusts the balance between information gain and resource consumption; For the Time-normalized resource consumption.

[0052] Experience Replay: Maintaining the Experience Buffer , storage transfer samples , used for batch training, where is the current state; for the actions taken; For the rewards received; is the next state after executing the action; is a state transition sample.

[0053] Target Network: Use the target network Stable training, periodically from the main network Copy the parameters.

[0054] Loss function: Minimize the temporal difference (TD) error: ; in is the loss function; Experience buffer zone Sampling The expected value of the sample; For rewards; is the discount factor; For all possible actions Take the maximum value; The next state of the target network and actions Q value estimation; The current state of the main network and actions The Q-value estimation of .

[0055] Training process: Initialize the main network and target network ; For each training cycle: according to - Greedy strategy selects actions; Perform actions, observe rewards and next states; Store experience in the buffer ; from Sampling small batches of data for training; Calculate TD target and loss function; Update main network parameters; Every Update the target network once, The target network synchronization period.

[0056] Through repeated training iterations, the system gradually learns the optimal sensor configuration strategy and can automatically generate the optimal sensor combination based on different working conditions and accuracy requirements; in actual applications, the system can be pre-trained in a simulation environment, and then the trained model can be deployed to the actual environment and adapted to the characteristics of the actual environment through online fine-tuning.

[0057] Optionally, in some embodiments, the system can adopt a policy gradient method (such as DDPG, PPO) instead of DQN to process the continuous action space and achieve more refined sensor parameter adjustment.

[0058] Step 2.3, realize adaptive reconstruction of sensor network; According to one embodiment of the present application, the system constructs an adaptive reconstruction mechanism for the sensor network, enabling the system to automatically adjust the network structure according to environmental changes, changes in task requirements or changes in sensor status; the mechanism includes three parts: trigger condition detection, reconstruction plan generation and smooth transition implementation.

[0059] Trigger conditions include: changes in working conditions (such as guide rail load or speed changes exceeding a threshold), changes in accuracy requirements (such as increased or decreased accuracy requirements), sensor failure or performance degradation, and new sensors added to the system. When the system detects these conditions, the network reconstruction process will be initiated.

[0060] The reconstruction scheme is generated based on the current sensor state and the latest information entropy quantization results. The graph reinforcement learning algorithm in step 2.2 is applied to quickly generate a sensor configuration scheme that adapts to the new situation. To ensure system stability, a smooth transition mechanism is introduced and a gradual adjustment strategy is adopted to avoid sudden changes in the configuration from causing impact on the system.

[0061] In addition, according to another embodiment of the present application, the system can build a preset reconstruction solution library, pre-calculate the optimal configuration solution for common scene changes, and directly load the corresponding solution when the trigger condition is met, further improving the reconstruction speed.

[0062] Step 3: Based on the optimal sensor combination configuration, realize multi-level dynamic resource allocation, build a multi-time scale optimization framework and implement adaptive data compression based on Huffman coding; Based on the sensor value quantification model and self-optimizing network configuration established in the first two steps, this step implements multi-level dynamic allocation of system resources to ensure maximum evaluation accuracy under limited resource conditions; this step includes the following sub-steps.

[0063] Step 3.1, construct a multi-timescale optimization framework; According to one embodiment of the present application, the system constructs a multi-timescale optimization framework to optimize resource allocation at different time granularities to achieve efficient utilization of system resources; the framework includes: Microsecond level: Sensor adaptive sampling optimization, real-time adjustment of sampling frequency; Millisecond level: data processing strategy adjustment, dynamic selection of processing algorithm and accuracy; Seconds: Optimizes computing resource allocation and dynamically schedules tasks between processing units. Minute level: Reconstruct the sensor network topology and adjust the sensor combination configuration.

[0064] Optimization at each level works in tandem with each other. Low-level optimization is constrained by high-level configurations, while high-level optimization is adjusted based on low-level feedback, forming a closed-loop optimization system. For example, when the system detects abnormal guide rail accuracy in a certain area, microsecond-level optimization will immediately increase the sampling frequency of sensors in that area, millisecond-level optimization will allocate more precise processing algorithms to the data in that area, second-level optimization will allocate more computing resources for data analysis in that area, and minute-level optimization may reconstruct the sensor network and increase the number or types of sensors in that area.

[0065] Optionally, in some embodiments, the system may add hourly and daily optimization levels for historical data pattern analysis and long-term optimization strategy adjustment, respectively, to cope with changes on a larger time scale.

[0066] like Figure 4 The figure below shows the performance of the proposed multi-timescale optimization framework across various performance metrics. It's clear from the figure that microsecond-level optimization achieves the best real-time performance, while minute-level optimization offers significant advantages in accuracy and resource efficiency. This result demonstrates the effectiveness of optimizing at different timescales in improving system performance across different aspects, demonstrating that the multi-timescale optimization framework can achieve comprehensive improvements in system performance.

[0067] Step 3.2, implement adaptive data compression based on Huffman coding; The principle of information entropy is applied to achieve adaptive compression of sensor data and improve data transmission and storage efficiency. Specifically, the coding length is dynamically allocated to different sensor data based on the information entropy.

[0068] The system first counts the probability of occurrence of each value of the sensor data, and then for each data value, calculates the negative logarithm (base 2) of its occurrence probability as its encoding length; in this way, data values ​​that appear frequently will be assigned shorter codes, and data values ​​that appear infrequently will be assigned longer codes, thereby minimizing the encoding length as a whole; for example, if the probability of occurrence of a sensor data value is 0.5, then its encoding length is approximately 1 bit; if the probability of occurrence is 0.25, then the encoding length is approximately 2 bits; if the probability of occurrence is 0.125, then the encoding length is approximately 3 bits, and so on.

[0069] The system first calculates the probability distribution of each sensor data, then constructs a Huffman tree and assigns a variable-length code to each data value. As the data distribution changes, the system dynamically updates the Huffman tree to maintain optimal compression efficiency. This adaptive compression method based on information entropy can reduce the amount of data transmitted, reduce network bandwidth and storage space requirements, while ensuring lossless compression without affecting evaluation accuracy.

[0070] It should be noted that, in some embodiments, the system can perform hierarchical compression based on the importance of the data, using lossless compression for key data to ensure accuracy, and using lossy compression for secondary data to further improve the compression rate; in addition, the system can combine predictive coding technology, use time series models to predict data values, and only transmit prediction errors to further improve compression efficiency.

[0071] like Figure 5 The figure shows the effectiveness of the adaptive data compression method based on Huffman coding proposed in this application on compressing different types of sensor data. Compared with traditional fixed-length coding (compression ratio of 0), all sensor data achieved high compression ratios, with optical sensor data achieving the highest compression ratio, retaining only 38% of the original data. This result fully demonstrates the advantages of the adaptive compression method proposed in this application in improving data transmission and storage efficiency.

[0072] Step 3.3: Build the edge-fog-cloud three-layer computing architecture; According to another embodiment of the present application, the system constructs a three-layer edge-fog-cloud computing architecture, and dynamically allocates computing tasks to the most appropriate computing layer based on task computing needs and timeliness requirements.

[0073] Edge layer: Deployed at sensor nodes or nearby processing units, it handles tasks with high real-time requirements and low computational load, such as data filtering and simple feature extraction. Fog layer: Deployed in field-level processing units, it handles tasks with medium computational load and timeliness requirements, such as preliminary fusion analysis and short-term trend forecasting. Cloud layer: Deployed on a central server or cloud platform to handle complex tasks with large computational load and non-real-time processing, such as deep model training and historical data mining.

[0074] The specific implementation of the computing task allocation mechanism is as follows.

[0075] The system performs each computing task Define three-dimensional eigenvectors: ; in For the task The eigenvector of For the task The computational complexity (the amount of computing resources required); For the task Real-time requirements (maximum allowed delay time); For the task The data dependencies (the amount of data that needs to be accessed).

[0076] Before defining the task feature vector, the system preprocesses data in three dimensions. Computational complexity can be expressed in terms of floating-point operations, algorithmic time complexity, and so on. Real-time requirements require latency thresholds expressed in time units (milliseconds, seconds). Data dependency is expressed in terms of data size (KB, MB, GB). These three dimensions have completely different scales and numerical ranges, requiring normalization to convert them to values ​​on the same scale to ensure equal weighting for calculating the matching score.

[0077] Task feature extraction function 、 、 The specific implementation is to calculate the amount of computing resources required by the task, the maximum allowed delay time, and the amount of data the task needs to access. These values ​​are obtained by analyzing parameters such as task type, input and output data size, and real-time requirements, and are used for subsequent task allocation and scheduling.

[0078] At the same time, for each computing layer Define the resource property vector: ; in For the computing layer The resource characteristic vector of For the computing layer computing power; For the computing layer response time; For the computing layer of available bandwidth.

[0079] Before defining the compute layer resource feature vectors, the system preprocesses the resource feature data. Computing power can be expressed in FLOPS, core count, or processing speed, response time can be expressed in milliseconds (latency), and available bandwidth can be expressed in Mbps or Gbps (network transmission rate). These metrics vary in scale and value range. When calculating the matching between tasks and compute layers, each resource feature must be normalized to prevent metrics with large value ranges from dominating the matching results.

[0080] Computational layer resource feature extraction function 、 、 The specific implementation is to measure the computing power, response time, and available bandwidth of each computing layer. These parameters are obtained by detecting the hardware configuration, network status, and current load of the computing nodes, and are used to evaluate the resource status of each computing layer.

[0081] The system calculates the matching score between the task and the computing layer based on the task feature vector and the resource characteristics of each computing layer.

[0082] The system determines the best execution layer for the computing task by calculating the comprehensive matching score. Specifically, it first calculates the ratio of the computing power of the computing layer to the computing complexity of the task and multiplies it by the first weight coefficient. Then, it calculates the ratio of the task real-time requirement to the response time of the computing layer and multiplies it by the second weight coefficient. Next, it calculates the ratio of the computing layer bandwidth to the task data dependency and multiplies it by the third weight coefficient. Finally, these three items are added together to obtain the matching score between the task and the computing layer. The weight coefficient can be adjusted according to the application scenario. The system selects the computing layer with the highest matching score to execute the corresponding computing task, and considers the current load situation of each layer through the load balancing factor to ensure reasonable allocation of resources.

[0083] The system dynamically determines the task allocation strategy based on task characteristics, current resource status, and network conditions. For example, for real-time monitoring of guide rail accuracy, basic data filtering and feature extraction are completed at the edge layer, multi-sensor data fusion and preliminary evaluation are completed at the fog layer, and complex accuracy trend analysis and prediction are completed at the cloud layer, forming a layered and collaborative computing system.

[0084] In addition, an asynchronous communication mechanism is adopted between each layer, so that the lower layer can work independently when the upper layer calculations are not completed, ensuring real-time responsiveness; the system also implements a task migration mechanism, which can dynamically migrate tasks to other layers for processing when the network connection is interrupted or resources are insufficient, ensuring the continuous operation of the system.

[0085] Step 4: Based on the compressed and optimized data, a dynamic Bayesian network model is constructed to fuse heterogeneous sensor data and achieve adaptive adjustment of the network structure; To solve the problem of heterogeneous sensor data fusion, this step constructs a dynamic Bayesian network model to achieve effective integration of different types of sensor data; this step includes the following sub-steps.

[0086] Step 4.1, establish a dynamic Bayesian network structure; According to one embodiment of the present application, the system constructs a dynamic Bayesian network (DBN) structure adapted for guide rail accuracy evaluation, which includes three layers: a sensor layer, a feature layer, and an accuracy parameter layer; the sensor layer nodes represent the raw data of various sensors, the feature layer nodes represent the features extracted from the raw data, and the accuracy parameter layer nodes represent the various indicators of guide rail accuracy.

[0087] Before building a dynamic Bayesian network, the system preprocesses heterogeneous data from various sensors. Continuous sensor data (such as temperature, displacement, and velocity) has different numerical ranges and units, discrete sensor data (such as switch states and mode selections) requires encoding, and time series feature data may have different sampling frequencies and time scales. To ensure that the Bayesian network can effectively learn the conditional dependencies between variables, all input data must be standardized, and categorical variables must be appropriately encoded to make them suitable for probabilistic inference calculations.

[0088] The edges in the network structure represent conditional dependencies between nodes, and the system learns the conditional probability distribution of these dependencies based on historical data. Unlike static Bayesian networks, dynamic Bayesian networks add a time dimension and can model temporal dependencies, making them more suitable for dynamic evaluation of guide rail accuracy over time.

[0089] The specific implementation of the dynamic Bayesian network is as follows.

[0090] First, define the network node type: Sensor node collection: ; in, represents the set of sensor nodes, 、 、 Represents the first, second, and sensor nodes, is the total number of sensor nodes.

[0091] Feature node collection: ; in, Represents a set of feature nodes, 、 、 Represents the first, second, and feature nodes, is the total number of feature nodes.

[0092] Precision parameter node collection: ; in, Represents a set of precision parameter nodes, 、 、 Represents the first, second, and precision parameter nodes, is the total number of precision parameter nodes.

[0093] Then, for each time slice , establish dependency types between nodes: Intra-layer temporal dependency: dependencies between nodes in the same layer, such as features and Conditional dependencies in the same time slice; Inter-layer temporal dependency: dependencies between nodes in different layers, such as sensors and features The mapping relationship; Intra-layer time dependency: the dependency of the same node between adjacent time slices, such as and The temporal relationship of Inter-layer time dependency: the dependency between nodes in different layers in adjacent time slices, such as and predictive relationship.

[0094] For each dependency, the system needs to learn its conditional probability distribution; for example, for the mapping from the feature layer to the accuracy layer, the system needs to learn the conditional probability ,in Indicates the precision parameter The parent node set may include feature nodes and the accuracy parameter node of the previous moment; is the conditional probability function.

[0095] The initial network structure is set based on expert knowledge and includes known sensor-feature-accuracy dependencies. The system then continuously optimizes the network topology through structural learning algorithms to discover potential dependencies.

[0096] In the actual guide rail accuracy evaluation system, the sensor layer may include nodes such as position sensors, acceleration sensors, and temperature sensors; the feature layer may include nodes such as average value, standard deviation, spectrum characteristics, and time domain characteristics; and the accuracy parameter layer may include nodes such as straightness, parallelism, and flatness.

[0097] Optionally, in some embodiments, the system can construct a hierarchical Bayesian network and establish multiple sub-networks at different abstraction levels to handle accuracy assessment problems at different scales.

[0098] Step 4.2, realize adaptive adjustment of network structure; Based on the principle of conditional information gain, this application designs an adaptive adjustment algorithm for the Bayesian network structure, which enables the network to automatically adjust its structure as data distribution and working conditions change. The adjustment mechanism includes: Calculate the conditional information gain for each potential edge (node ​​pair); Add connecting edges to node pairs whose information gain exceeds the threshold; Prune existing edges whose information gain is lower than the threshold; Apply network simplification algorithms to ensure a streamlined and efficient structure.

[0099] The conditional information gain is calculated as follows.

[0100] For nodes and , and the established edge set , the conditional information gain is expressed as: ; in For a known edge set Under these conditions, the node For Node Information gain; For a known edge set under conditions The conditional entropy of For the known and under conditions The conditional entropy of Is a conditional symbol.

[0101] when When and Add an edge between them; when the existing edge When , prune the edge, where is the preset threshold for adding edges, The preset threshold for pruning edges.

[0102] The structure learning algorithm adopts a scoring-based approach and uses the Bayesian Information Criterion as the scoring function: ; in Scoring the Bayesian Information Criterion; For the network structure; is the observation dataset; is the logarithm operator; For the network structure and parameter estimates Download data The likelihood function value of ; is the parameter estimate; is the number of model parameters; is the data sample size.

[0103] The BIC criterion balances model fit and complexity to prevent overfitting.

[0104] The system regularly executes the above process to continuously optimize the network structure. When a change in operating conditions is detected, the system triggers an additional structural learning process to quickly adapt to the new conditions. This adaptive adjustment mechanism enables the Bayesian network to continuously evolve with environmental changes and maintain optimal fusion effects.

[0105] In the specific application of guide rail accuracy assessment, the system may find that the correlation between certain sensors and accuracy parameters increases under specific operating conditions (such as the increased importance of acceleration sensors during high-speed operation), and automatically adjust the network structure to strengthen these associations; while weakening those connections whose correlation decreases under the current operating conditions (such as the reduced importance of temperature sensors during stable operation).

[0106] It should be understood that in some embodiments, the system may use an incremental structural learning algorithm to perform online structural adjustments on newly received data, thereby avoiding full retraining and improving system response speed.

[0107] Step 4.3: Fuse heterogeneous sensor data for accuracy evaluation; According to another embodiment of the present application, the system utilizes a constructed dynamic Bayesian network model to fuse data from different types of sensors and infer the guide rail accuracy parameters. The specific fusion process includes: Data preprocessing: normalize, denoise and extract features of each sensor data; Evidence input: The processed sensor data is input into the Bayesian network as evidence; Probabilistic inference: Apply algorithms such as junctiontree or variational inference to calculate the posterior probability distribution of the accuracy parameter; Result integration: Based on the posterior distribution, output the point estimate and confidence interval of the accuracy parameter.

[0108] The probabilistic inference framework of the Bayesian network can effectively handle the uncertainty of sensor data. When some sensor data is missing or noisy, the system can still make reasonable inferences based on other sensor data. At the same time, the framework can also provide confidence in the evaluation results, providing a reliability reference for decision-making.

[0109] In addition, the system also adopts a dynamic weight adjustment strategy to dynamically adjust the weight of the sensor in the fusion process according to the reliability of the sensor under different working conditions, further improving the fusion accuracy; for example, when the temperature changes drastically, the system will reduce the weight of the temperature-sensitive sensor data and increase the weight of the temperature compensation sensor.

[0110] Step 5: Based on the adaptive adjustment results, a self-evolving evaluation strategy system is constructed to continuously improve the evaluation strategy through historical performance data. Construct a self-evolving evaluation strategy system so that the accuracy evaluation method can be continuously optimized and evolved through experience accumulation; this step includes the following sub-steps.

[0111] Step 5.1, establish a policy library and performance evaluation mechanism; According to one embodiment of the present application, the system constructs an evaluation strategy library to store evaluation strategies for different scenarios, guide rail types and accuracy requirements; each strategy contains information such as sensor configuration, resource allocation plan, fusion model parameters, etc.; at the same time, a performance evaluation mechanism is established to quantitatively evaluate the effectiveness of the strategy from multiple dimensions such as evaluation accuracy, resource consumption, and time efficiency.

[0112] The system records the actual performance of the strategy during each evaluation process, including the deviation between the evaluation result and the true value (by comparing it with high-precision offline measurements), resource usage, completion time and other indicators, calculates the overall performance score, and associates the score with the strategy for storage for subsequent strategy selection and optimization.

[0113] It should be noted that the performance evaluation adopts a multi-objective evaluation framework, which dynamically adjusts the weights of indicators in each dimension according to the different needs of the application scenarios. For example, in real-time monitoring scenarios, the weight of time efficiency will be increased; and in high-precision detection scenarios, the weight of evaluation accuracy will be higher.

[0114] Step 5.2, implement strategy optimization and evolution mechanism; This application designs a strategy optimization and evolution mechanism to continuously improve the evaluation strategy through historical performance data; this mechanism mainly includes the following algorithms.

[0115] Strategy extraction: extracting high-performance strategies and their applicable conditions from historical records; Strategy generalization: Analyze the commonalities of high-performance strategies and form a universal strategy template; Strategy mutation: random small mutations to the existing strategy to explore potential improvement space; Strategy crossover: combining the advantages of different strategies to generate new strategies; Strategy verification: Verify the effectiveness of new strategies through simulation or small-scale testing.

[0116] The system regularly executes the above process to continuously update and expand the strategy library; as the system operation time increases, the experience and knowledge accumulated in the strategy library will also increase, and the system performance will continue to improve.

[0117] Optionally, in some embodiments, the system may use a genetic algorithm framework to manage the strategy library evolution process; each strategy is represented as an "individual", high-performance strategies are retained through the principle of survival of the fittest, and new strategies are generated through mutation and crossover operations to achieve optimized evolution of the strategy library.

[0118] Step 5.3: Implement adaptive strategy selection based on scene similarity; When the system faces a new evaluation task, it needs to select or generate the evaluation strategy that best suits the current scenario from the strategy library; to this end, according to another embodiment of the present application, the system designs a strategy adaptive selection algorithm based on scenario similarity.

[0119] Feature extraction: Extract the feature vector of the current scene (including guide rail type, working parameters, accuracy requirements, etc.); Similarity calculation: Calculate the similarity between the current scenario and each historical scenario in the strategy library; Strategy selection: Based on similarity and historical performance, select the most suitable basic strategy for the current scenario; Strategy adjustment: Fine-tune the basic strategy based on scenario differences to generate the final execution strategy.

[0120] This similarity-based strategy selection mechanism can make full use of historical experience to quickly generate high-quality strategies for new scenarios, while avoiding strategy search from scratch and improving system response speed.

[0121] In addition, the system also implements a strategy recommendation function. When the user manually configures the system, it can recommend the most suitable strategy parameters based on the current scenario characteristics to assist the user in decision-making. In actual applications, the system can integrate user choices and feedback into the strategy optimization process through interactive learning to achieve human-machine collaborative evolution.

[0122] like Figure 6As shown, the radar chart comprehensively compares the performance of the multimodal information entropy collaborative optimization method proposed in this application with the traditional method in five key performance indicators. It can be clearly seen that the optimization system of this application has improved resource utilization efficiency, system adaptability and autonomous learning ability compared with traditional methods, while reducing management complexity, which is completely consistent with the above description of technical effects. This figure intuitively shows the comprehensive advantages of the method of this application compared with traditional methods in multiple dimensions.

[0123] A guide rail accuracy evaluation system based on multi-sensor fusion is used to execute the above-mentioned guide rail accuracy evaluation method based on multi-sensor fusion, comprising: Information entropy quantification module, used to build a multimodal sensor value quantification model based on information entropy; Sensor network configuration module, used to build a self-optimizing sensor network configuration system; Resource allocation module, used to implement multi-level dynamic resource allocation; Data fusion module, used to build a dynamic Bayesian network model to fuse heterogeneous sensor data; Evaluation strategy module, used to build a self-evolving evaluation strategy system.

[0124] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A guide rail accuracy evaluation method based on multi-sensor fusion, characterized in that: include: Construct a multimodal sensor value quantification model based on information entropy, calculate the sensor information contribution and construct a multi-sensor mutual information coupling model; Based on the output of the mutual information model, a self-optimizing sensor network configuration system is constructed. The multimodal sensor system is abstracted into a graph structure and a graph reinforcement learning algorithm is applied to generate the optimal sensor combination configuration. Based on the optimal sensor combination configuration, multi-level dynamic resource allocation is achieved, a multi-time scale optimization framework is constructed, and adaptive data compression based on Huffman coding is implemented; Based on compressed and optimized data, a dynamic Bayesian network model is constructed to fuse heterogeneous sensor data and achieve adaptive adjustment of the network structure; Based on the adaptive adjustment results, a self-evolving evaluation strategy system is constructed to continuously improve the evaluation strategy through historical performance data.

2. The guide rail accuracy evaluation method based on multi-sensor fusion according to claim 1 is characterized in that: The step of calculating the sensor information contribution comprises: Obtain the probability distribution of guide rail accuracy parameters in the historical data set and calculate its entropy value; Calculate the conditional entropy of the guide rail accuracy under the condition of known sensor data; The information contribution of the sensor is calculated by the difference between the entropy value and the conditional entropy.

3. The guide rail accuracy evaluation method based on multi-sensor fusion according to claim 1 is characterized in that: The steps of constructing a sensor multi-sensor mutual information coupling model include: Calculate mutual information between sensor pairs and identify redundant sensor groups and complementary sensor groups; Calculate high-order mutual information between sensors to capture complex nonlinear complementary relationships; A resource allocation function is established to associate the contribution of sensor information with resource consumption.

4. The guide rail accuracy evaluation method based on multi-sensor fusion according to claim 1 is characterized in that: The step of applying the graph reinforcement learning algorithm to optimize the sensor configuration includes: The sensor topology network is represented as a graph structure, where nodes represent sensors and edges represent relationships between sensors; Construct a graph neural network as a function approximator to achieve node feature aggregation and node feature update; A deep Q-learning framework is used for reinforcement learning training, and the sensor configuration is optimized by maximizing the balance objective function between information gain and resource consumption.

5. The guide rail accuracy evaluation method based on multi-sensor fusion according to claim 1 is characterized in that: The multi-level dynamic resource allocation includes optimization of the following time scales: Microsecond-level sensor adaptive sampling optimization; Millisecond-level data processing strategy adjustment; Second-level optimization of computing resource allocation; Sensor network topology reconstruction in minutes.

6. The guide rail accuracy evaluation method based on multi-sensor fusion according to claim 1 is characterized in that: The steps of implementing adaptive data compression based on Huffman coding include: Count the occurrence probabilities of each value of sensor data; Calculate the negative logarithm of the probability of occurrence of the data value as its encoding length; Construct a Huffman tree to assign variable length codes to each data value; Dynamically update the Huffman tree as data distribution changes.

7. The guide rail accuracy evaluation method based on multi-sensor fusion according to claim 1 is characterized in that: The steps of constructing the dynamic Bayesian network model include: Construct a three-layer network structure including sensor layer, feature layer and accuracy parameter layer; Establish four dependency relationships: intra-layer same-time dependency, inter-layer same-time dependency, intra-layer time dependency, and inter-layer time dependency; A network structure adaptive adjustment algorithm is designed based on the conditional information gain principle.

8. The guide rail accuracy evaluation method based on multi-sensor fusion according to claim 1 is characterized in that: The steps of constructing the self-evolving evaluation strategy system include: Establish a policy library and performance evaluation mechanism to store evaluation strategies for different scenarios; Implement strategy optimization and evolution mechanisms to continuously improve evaluation strategies through historical performance data; Design a strategy adaptive selection algorithm based on scenario similarity to select the evaluation strategy that best suits the current scenario from the strategy library.

9. The guide rail accuracy evaluation method based on multi-sensor fusion according to claim 3 is characterized in that: The resource allocation function quantifies the resource utilization efficiency of the sensor by calculating the ratio of sensor information contribution to resource consumption, and preferentially configures high-efficiency sensors according to the resource efficiency ratio.

10. A guide rail accuracy evaluation system based on multi-sensor fusion, characterized in that: A guide rail accuracy assessment method based on multi-sensor fusion for executing any one of claims 1 to 9, comprising: Information entropy quantification module, used to build a multimodal sensor value quantification model based on information entropy; Sensor network configuration module, used to build a self-optimizing sensor network configuration system; Resource allocation module, used to implement multi-level dynamic resource allocation; Data fusion module, used to build a dynamic Bayesian network model to fuse heterogeneous sensor data; Evaluation strategy module, used to build a self-evolving evaluation strategy system.

Images