A probabilistic model driven large-scale sparse multi-objective optimization method and system

Abstract

The application discloses a kind of probability model driven large-scale sparse multi-objective optimization method and system, belong to artificial intelligence field.First, the signal to be transmitted of tension machine is abstracted as multidimensional decision variable matrix, is combined with feature mask to construct detection solution, calculates importance score, generates main and auxiliary importance vector;Second, based on K-means clustering result calculation activation probability and establish initial global probability model, and then generate initial elite population, by the weighted fusion of activation frequency vector and initial model obtains refined global probability model, completes first variable grouping;Finally, by grouping coevolution algorithm obtains Pareto optimal compression transmission scheme set, and by field control terminal according to real-time bandwidth automatically selects the most matching scheme to execute data sending.The method focuses on key signal channel by dynamic probability modeling and closed-loop feedback mechanism, significantly reduces the transmission bandwidth overhead and reconstruction error of tension machine CAN data.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, and in particular relates to a probabilistic model-driven method and system for large-scale sparse multi-objective optimization. Background Technology

[0002] With the continuous development of scientific research and engineering practice, large-scale sparse multi-objective optimization problems (LSMOs) are widely found in many fields, such as data compression, signal processing, and resource scheduling. These problems are characterized by a huge number of decision variables (ranging from hundreds to thousands of dimensions), the need to simultaneously optimize multiple conflicting objectives, a large decision space, and a significantly sparse solution set.

[0003] During the operation of power grid tensioning machines, a large amount of high-dimensional CAN signal data is generated. These signals are typically characterized by high dimensionality, sparse effective information, and complex coupling between parameters. Achieving efficient compression and real-time transmission while ensuring signal integrity and accuracy becomes a typical large-scale sparse multi-objective optimization problem. Specifically, this problem requires simultaneously optimizing multiple conflicting objectives such as compression ratio, signal recovery accuracy, and real-time processing capability. Traditional optimization algorithms struggle to directly address the characteristics of high-dimensional, sparse, and dynamically changing signals.

[0004] Existing technologies primarily employ two approaches to address large-scale sparse optimization problems: The first is the Co-evolutionary (CC) method, which reduces search space complexity by decomposing high-dimensional decision variables into several low-dimensional subgroups and optimizing them sequentially. However, in the CAN signal compression scenario of a tensioning machine, static grouping cannot adapt to the dynamic characteristics of signal parameters' importance changing with time and state, potentially leading to key parameters being assigned to different subgroups, thus affecting the efficiency of the compression strategy. The second approach is the SparseEA sparse optimization method, which decomposes the optimization problem into two independent tasks: binary sparse masking and real-valued decision variables, thereby improving the ability to identify sparse solutions. However, this method fails to feed back the learned parameter contributions for dynamic grouping or compression strategy adjustment, limiting its application effectiveness in real-time CAN signal processing systems.

[0005] Therefore, existing technologies for CAN signal compression in tensioning machines suffer from problems such as the inability to dynamically identify key parameters, low utilization of computing resources, insufficient compression efficiency, and signal information loss. This indicates an urgent need for an optimization method and application system capable of real-time learning of signal characteristics and dynamic adjustment of parameter grouping and compression strategies to meet the practical needs of high-dimensional sparse CAN signal processing in power grid tensioning machines. Summary of the Invention

[0006] The purpose of this invention is to address the problems existing in the prior art and to provide a probabilistic model-driven method and system for large-scale sparse multi-objective optimization. This invention can decompose high-dimensional optimization problems into low-dimensional sub-problems and accurately identify key variables, concentrating computational resources on important variables, thereby efficiently solving the dual challenges brought by large scale and sparsity.

[0007] To achieve the above-mentioned objectives, the present invention specifically adopts the following technical solution:

[0008] In a first aspect, the present invention provides a probabilistic model-driven method for large-scale sparse multi-objective optimization, comprising the following steps:

[0009] S1: Using the various sensor signals of the tensioner as decision variables, a temporary population is formed by generating probe solutions that only activate each decision variable through masking operations. The importance score of the decision variable activated by each probe solution is calculated with the goal of minimizing transmission bandwidth overhead and reconstruction error.

[0010] S2: Cluster decision variables according to importance scores, and assign each decision variable an activation probability that is negatively correlated with the size of its cluster to obtain a global probability model; use a double sparse masking mechanism to generate an initial population and merge it with a temporary population to form an elite population; calculate the activation frequency of each decision variable in the elite population and update the global probability model with weights; group the decision variables sequentially with the updated activation probabilities.

[0011] S3: Based on the grouping results, perform grouped co-evolution on the elite population and the global probability model to obtain the optimal compressed transmission scheme for compressed transmission signals.

[0012] Based on the above scheme, each step can be implemented in the following preferred manner.

[0013] As a preferred embodiment of the first aspect, the temporary population is generated as follows: For the multi-source sensor time-domain signals that need to be fully transmitted in the CAN bus of the tensioner, a multi-dimensional decision variable matrix is ​​constructed by sampling and assigning values ​​within their respective value ranges. Simultaneously, for each decision variable representing a sensor signal type, a binary feature mask is generated, and the multi-dimensional decision variable matrix is ​​masked to obtain a probe solution that retains only the corresponding decision variable signal value while compressing the signal values ​​of the remaining decision variables to zero. All probe solutions constitute the temporary population. As a preferred embodiment of the first aspect, the importance score of each decision variable is calculated as follows: The number of bits occupied by all activated decision variables during transmission is taken as the transmission bandwidth overhead; the weighted deviation between the original signal data before compression and the reconstructed signal data after compression is taken as the reconstruction error; a multi-objective performance evaluation function that simultaneously minimizes the transmission bandwidth overhead and the reconstruction error is constructed; the ranking level of each probe solution is calculated using a non-dominated ranking algorithm; and its level value represents the importance score of the activated decision variable.

[0014] As a preferred embodiment of the first aspect, the initial construction method of the global probability model is as follows: all decision variables are clustered into two clusters according to their importance scores, forming a high-importance cluster and a low-importance cluster. The activation probability of each decision variable within each cluster is obtained by normalizing the inverse of its cluster size. After all decision variables obtain the initial values ​​of their activation probabilities, the initial global probability model is formed.

[0015] As a preferred option for the first aspect mentioned above, the method of generating the initial population using a dual sparse masking mechanism is as follows: First, based on the importance scores of the decision variables, a random number of decision variables are forcibly activated and used as seed variables through a tournament selection method; then, for the remaining decision variables that are not seed variables, they are independently activated according to the activation probabilities in the initial global probability model; finally, a binary sparse mask vector is generated based on the activation status of the decision variables, and the multidimensional decision variable matrix is ​​masked so that the signal values ​​of the activated decision variables are preserved while the signal values ​​of the rest are compressed to zero, resulting in an individual in the initial population; by repeatedly performing decision variable activation and masking operations, an initial population that meets the population size is generated.

[0016] As a preferred option of the first aspect mentioned above, the method for weighted updating of the global probability model is as follows: First, select all individuals located at the first non-dominant frontier from the initial elite population as elite individuals. Then, extract the sparse mask vectors of all elite individuals, count the activation frequency of each decision variable in all elite individuals, and weight and update the activation frequency of each decision variable to the activation probability of the corresponding decision variable in the global probability model to obtain the refined global probability model.

[0017] As a preferred approach to the first aspect mentioned above, the iteration process of each round of grouped co-evolution is as follows: Parent individuals are selected from the population generated in the previous round of iteration. Under the guidance of the global probability model, sparse mask evolution is performed on the selected parent individuals, and decision variable evolution is performed on the currently selected variable group. Offspring individuals are re-masked using the evolved sparse mask vector and the multidimensional decision variable matrix. A new population for the next round of iteration is generated based on the offspring individuals and parent individuals. All individuals in the new population located at the first non-dominant frontier are taken as new elite individuals. The activation frequency of each decision variable is recalculated based on the sparse mask vector of the new elite individuals. The activation frequency of each decision variable is then weighted and updated to the corresponding activation probability of the decision variable in the global probability model. Simultaneously, the next variable group is selected sequentially for the next round of iteration.

[0018] As a preferred embodiment of the first aspect mentioned above, in the sparse mask evolution, for each sparse mask of the parent individual, its crossover probability is determined by the global probability model, and its mutation probability is a preset value, thereby generating the sparse mask vector of the offspring individual; in the decision variable evolution, firstly, the decision variable vector of the parent individual is subjected to a crossover operation to generate the basis of the decision variable vector of the offspring individual, and then, according to the current set of variables to be optimized, a polynomial mutation is performed on the basis of the crossover decision variable vector of the offspring individual, which only takes effect at the position of the activated decision variable in the set of variables.

[0019] As a preferred option in the first aspect mentioned above, the selected variable group is continuously rotated from beginning to end during the iteration process, and when the number of rotations reaches a preset value, a regrouping is triggered, and the decision variables are regrouped in descending order based on the activation probability of each decision variable in the latest global probability model.

[0020] As a preferred option in the first aspect mentioned above, for the third Basis of decision variable vectors Polynomial mutation is performed according to the following formula if and only if the decision variable is in the current set of variables to be optimized and the decision variable is activated:

[0021]

[0022] in, The base after polynomial mutation; and All are disturbances, determined by the distribution index. Decide; and These are the lower and upper bounds of the decision variable, respectively.

[0023] Secondly, the present invention provides a probabilistic model-driven large-scale sparse multi-objective optimization system for implementing the probabilistic model-driven large-scale sparse multi-objective optimization method described in any of the solutions of the first aspect above. Specifically, the system includes three modules, namely:

[0024] The initialization and knowledge acquisition module is used to generate a temporary population of probe solutions that only activate each decision variable by using various sensor signals of the tensioner as decision variables through masking operations. The module calculates the importance score of the decision variables activated by each probe solution with the goal of minimizing transmission bandwidth overhead and reconstruction error.

[0025] The initial grouping module is used to cluster decision variables according to importance scores and assign an activation probability to each decision variable that is negatively correlated with the size of its cluster to obtain a global probability model. The initial population is generated by a dual sparse masking mechanism and merged with the temporary population to form an elite population. The activation frequency of each decision variable in the elite population is calculated and the global probability model is updated in a weighted manner. The decision variables are then grouped sequentially with the updated activation probabilities.

[0026] The co-evolution module is used to perform grouped co-evolution on the elite population and the global probability model based on the grouping results, so as to obtain the optimal compressed transmission scheme for compressed transmission signals.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] First, unlike traditional random or static variable partitioning methods, this invention, for tensioner CAN data, uses the dynamic learning results of a global probability model as the sole basis for variable grouping. This ensures that the grouping process is based on a deep understanding of the interrelationships and structural characteristics of the signals to be compressed. By periodically identifying and aggregating sensor signal channels of higher importance, this method can direct computational resources to the key variable subspace, significantly reducing redundant searches caused by irrelevant channels, thereby improving the efficiency and stability of the tensioner signal compression strategy.

[0029] Secondly, this invention constructs a closed-loop feedback mechanism for CAN data from a tensioning machine, which involves the coordinated efforts of elite solution learning, probabilistic model updating, variable sparsity mining, and dynamic grouping. This mechanism can adaptively adjust the problem decomposition and search strategy according to changes in the search process, ensuring that the optimization process maintains strong robustness and adaptability when facing multi-source heterogeneous signals, strongly correlated signals, and constrained compression requirements at the tensioning machine site. Based on this tightly coupled feedback structure, this invention can continuously drive the population towards a better compression ratio and lower reconstruction error, achieving efficient, multi-objective compression optimization based on real-time CAN signals. Attached Figure Description

[0030] Figure 1This is a flowchart of the steps of the method of the present invention;

[0031] Figure 2 This is a flowchart of the initial grouping process based on variable importance in this invention;

[0032] Figure 3 This is a flowchart of the probabilistic model-driven cooperative evolution of the present invention;

[0033] Figure 4 This is a system block diagram of the present invention;

[0034] Figure 5 This is a schematic diagram of a computer electronic device provided by the present invention. Detailed Implementation

[0035] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0036] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.

[0037] In power grid construction operations, tensioning machines (tensioning and wire laying equipment) output 25-40 types of operating status signals in real time via a CAN bus at a sampling frequency of 50-100Hz. These signals include, but are not limited to, conductor tension, hydraulic pressure, main pump current, hydraulic system temperature, laying speed, braking status, oil pump pressure, engine speed, and slip coefficient. If all signals are transmitted losslessly at each sampling moment, the raw data volume of a single 10-second data packet can reach several MB, far exceeding the economic bandwidth capacity of existing 4G / 5G private networks or fiber optic links, while also significantly increasing the burden on terminal storage and processing. Actual engineering needs show that only a very small number of critical safety signals (such as tension and braking status) require high-precision transmission at specific moments (such as sudden tension changes or emergency braking). At other times, non-critical signals can be transmitted sparsely with acceptable distortion through zero-order hold, linear prediction, or large-amplitude quantization. Therefore, while ensuring lossless or high-precision preservation of critical safety signals, implementing lossy compression within a controllable error range for a large number of non-critical or low-sensitivity signals is a necessary means to improve the data transmission efficiency of tensioning machines and the overall performance of the system. However, different CAN signals vary significantly in their importance to construction safety, equipment control performance, and operational condition sensing value. Traditional fixed compression strategies cannot simultaneously address the fidelity requirements of critical data and the compression needs of non-critical data. Therefore, an adaptive optimization method is needed that can automatically identify the importance of variables and generate differentiated, multi-level compression strategies accordingly.

[0038] In view of the above background, in a preferred embodiment of the present invention, a probabilistic model-driven large-scale sparse multi-objective optimization method is provided, and the method is applied to the optimization scenario of high-frequency CAN bus data sparse transmission of tensioning machines in power grid construction operations, so as to reduce the communication bandwidth pressure of remote monitoring systems and ensure high-fidelity real-time transmission of key safety signals.

[0039] like Figure 1 As shown, in a preferred embodiment of the present invention, the above-mentioned probability model-driven large-scale sparse multi-objective optimization method generally includes three steps: initialization and knowledge acquisition, initial grouping based on variable importance, and probability model-driven co-evolution, which correspond to steps S1 to S3 below. The specific implementation process of each step will be described in detail below.

[0040] I. Initialization and Knowledge Acquisition

[0041] S1: Using the various sensor signals of the tensioner as decision variables, a temporary population is formed by generating probe solutions that only activate each decision variable through masking operations. The importance score of the decision variable activated by each probe solution is calculated with the goal of minimizing transmission bandwidth overhead and reconstruction error.

[0042] It should be noted that step S1 of this invention aims to identify the differences in the contribution of different sampling points of the tensioner's CAN signal in the time and frequency domains to the overall compression performance, lossless or lossy reconstruction accuracy, and anomaly detection capability, and to form the initial knowledge required for subsequent probabilistic models and variable grouping based on this. By performing univariate activation perturbation, independence evaluation, and multi-objective performance analysis on the 500-dimensional decision variables, this step constructs an importance vector reflecting the operating characteristics of the tensioner, providing a foundation for the automated generation of compression strategies.

[0043] It should be noted that in step S1 of the present invention, the temporary population is generated as follows: for the time-domain signals of the multi-source sensors that need to be fully transmitted in the CAN bus of the tensioner, a multi-dimensional decision variable matrix is ​​constructed by sampling and assigning values ​​within their respective value ranges. At the same time, for each decision variable representing a sensor signal type, a binary feature mask is generated, and a masking operation is performed on the multi-dimensional decision variable matrix to obtain a detection solution that retains only the corresponding decision variable signal value while compressing the signal values ​​of the other decision variables to zero. All detection solutions form a temporary population.

[0044] It should be noted that in step S1 of this invention, the importance score of each decision variable is calculated as follows: the number of bits occupied by all activated decision variables during transmission is taken as the transmission bandwidth overhead, and the weighted deviation between the original signal data before compression and the reconstructed signal data after compression is taken as the reconstruction error. A multi-objective performance evaluation function that minimizes both transmission bandwidth overhead and reconstruction error is constructed. The ranking level of each probe solution is calculated by a non-dominated ranking algorithm, and its level value represents the importance score of the activated decision variable.

[0045] It should be noted that in step S1 of this invention, each decision variable in the multidimensional decision variable matrix corresponds to a type of time-domain signal that needs to be transmitted via the tensioning machine's CAN bus. There are a total of [number] signal types. The sampling parameters include, but are not limited to, tension, hydraulic pressure, main pump current, hydraulic system temperature, wire release speed, braking status, oil pump pressure, engine speed, and slip coefficient. The sampling frequency is [missing information]. Hz, then within the time window The total number of data points (i.e., the dimensions of decision variables) is .

[0046] In step S1 of this embodiment, 2 seconds of operational data within a typical operating cycle of the tensioning machine are selected as the optimization object, including 25 types of sensor signals, a sampling frequency of 50Hz, and a total of 2500 raw sampling points. To balance computational efficiency and representativeness, a key sampling strategy combined with uniform sampling is adopted, selecting 250 key sampling moments for optimization, resulting in a decision variable dimension of 500. Specifically, in this embodiment, a... Decision variable matrix This matrix is ​​used to simulate the possible numerical combinations of signals under different operating conditions, thereby ensuring the coverage and stability of the detection results. Each row in the decision variable matrix... This represents a sample vector used for a single detection. Each element is a decision variable, and its specific value can be generated using a fully random sample generation strategy, i.e., based on the physical range allowed by the field equipment and the sensor's range within its corresponding value range. Internal randomization is used to simulate a wide range of possible operating conditions for the tensioning machine under different construction environments, thereby ensuring comprehensive coverage of variable importance assessment. This represents the upper bound of the decision variable's values. This serves as the lower bound for the values ​​of the decision variables. The specific values ​​of the decision variables can also be generated using a benchmark perturbation strategy. This involves applying a small-range perturbation to each real-time signal based on the normal operating conditions of the tensioning machine (such as steady-state tension, normal hydraulic pressure, and typical speed) to simulate minute fluctuations and abnormal precursor states in real-world operating scenarios, thereby enhancing the sensitivity of the scoring to key signals.

[0047] Furthermore, to assess the independent influence of each CAN bus's real-time signal on the compression task, this invention also constructs a feature mask for detection for each decision variable. For the first... Given a set of decision variables, generate a feature mask. , This is the index for the decision variable. The mask is a unit vector, only the first... A position set to 1 indicates that the data point is "activated," meaning it is used as critical data for high-precision or lossless transmission. The remaining positions are... When all bits are 0, it indicates that the data point is "sparsed," meaning it is either discarded during transmission (the receiver maintains the value from the previous moment) or only the low-bit prediction value is transmitted. This physically simulates the extreme sparsity strategy of "transmitting only the current data point and discarding all other data points," aiming to quantify the "marginal contribution rate" of this data point to the overall recovery through extreme testing. This mechanism allows each probe to isolate the impact of a single signal dimension on the performance of multiple targets, thereby obtaining an interpretable measure of independent importance.

[0048] After constructing the above feature mask, this invention extends the feature mask using a broadcast mechanism to... Then multiply it element-by-element by the decision variable matrix to obtain The detection solution. Among them, the first... Each detection solution Represented as:

[0049]

[0050] in, This represents element-wise multiplication. Used to indicate only the first Compression parameter configuration when one data point is active and the rest are disabled.

[0051] In this embodiment, two conflicting objective functions are designed. The first objective function is to minimize the transmission bandwidth overhead, and the second objective function is to minimize the reconstruction error, which are expressed as follows:

[0052]

[0053]

[0054] in, To minimize; For transmission bandwidth overhead; This represents the reconstruction error; For transmission of the first The number of bits used when there are multiple decision variables; The importance weights of the decision variables are preset; This refers to the original signal data before compression. This refers to the reconstructed signal data after compression. Here, "original signal data before compression" refers to the data that has not yet been transmitted or compressed at the data sending end; "reconstructed signal data after compression" does not refer to reconstructing the actual received signal data at the data receiving end, but rather to simulating the reconstructed signal data by replacing the value of one decision variable in the original signal data with its value from the previous moment, while keeping the values ​​of the other decision variables unchanged.

[0055] Furthermore, in this embodiment, considering the safety of power grid operations, different weights are assigned to different decision variables. For key safety variables such as tension and braking state, specific weights are set. For auxiliary variables such as engine speed and oil pump pressure, In calculating the weighted reconstruction error, this embodiment processes a portion of the decision variables in the original data. Specifically, it replaces these data with the data transmitted in the previous time step, thereby generating reconstructed data.

[0056] Then, this embodiment begins to evaluate the objective function of these D probe solutions, calculating the objective function value for each probe solution. and The ranking of each decision variable is calculated using an existing non-dominated sorting algorithm. This ranking value is then used as the importance score for the corresponding decision variable; a lower ranking (1 being optimal) indicates higher importance. Specifically, the importance score for each decision variable... Defined as:

[0057]

[0058] in, It is a non-dominated sorting algorithm; This represents the multi-objective performance evaluation function mentioned above. This function comprehensively considers two performance metrics: transmission bandwidth overhead and weighted reconstruction error. It can be used to evaluate the actual impact of the compression strategy on the CAN bus performance of the tensioner when only a single signal dimension is activated.

[0059] This embodiment, under different decision variable generation strategies, obtained several sets of different importance score vectors after multiple objective function evaluations. These importance score vectors were then aggregated to form the principal importance vectors used for subsequent S2 clustering and probabilistic model construction. and the auxiliary importance vector used for seed selection .

[0060] II. Initial grouping based on variable importance

[0061] S2: Cluster decision variables according to importance scores and assign each decision variable an activation probability negatively correlated with the size of its cluster to obtain a global probability model; use a double sparse masking mechanism to generate an initial population and merge it with a temporary population to form an elite population; calculate the activation frequency of each decision variable in the elite population and update the global probability model with weights; and group the decision variables sequentially with the updated activation probabilities.

[0062] It should be noted that step S2 of this invention aims to transform the discrete importance scores obtained in S1 into a continuous prior probability distribution, thereby guiding the algorithm to automatically identify "critical safety variables" and "ordinary operational variables." Through unsupervised clustering, the algorithm can adaptively define what constitutes "importance" without manually setting a fixed threshold. Specifically, this step provides a reliable prior structure for the subsequent co-evolutionary main loop by constructing a guide population, refining the probability model, and performing variable grouping for the first time. Its purpose is to establish an initial variable distribution that better fits the problem structure before evolution begins, giving the algorithm stronger convergence guidance capabilities. This variable grouping result guides subsequent probability model updates and the co-evolutionary process, enabling the compression strategy to be optimized in stages according to the principle of "prioritizing important signals and delaying secondary signals," which better suits the structural characteristics and compression requirements of the tensioner CAN data.

[0063] It should be noted that in step S2 of the present invention, the initial construction method of the global probability model is as follows: all decision variables are clustered into two clusters according to their importance scores, forming a high importance cluster and a low importance cluster. The activation probability of each decision variable within the cluster is obtained by normalizing the inverse of its cluster size. After all decision variables obtain the initial values ​​of their activation probabilities, the initial global probability model is formed.

[0064] It should be noted that in step S2 of this invention, the method of generating the initial population using a dual sparse masking mechanism is as follows: First, based on the importance scores of the decision variables, a random number of decision variables are forcibly activated and used as seed variables through a tournament selection method; then, for the remaining decision variables that are not seed variables, they are independently activated according to the activation probabilities in the initial global probability model; finally, a binary sparse mask vector is generated based on the activation status of the decision variables, and the multidimensional decision variable matrix is ​​masked so that the signal values ​​of the activated decision variables are preserved while the remaining signal values ​​are compressed to zero, thus obtaining an individual in the initial population; by continuously repeating the decision variable activation and masking operations, an initial population that meets the population size is generated.

[0065] It should be noted that in step S2 of this invention, the method for weighted updating of the global probability model is as follows: First, all individuals located at the first non-dominant frontier are selected from the initial elite population as elite individuals. Then, the sparse mask vectors of all elite individuals are extracted, and the activation frequency of each decision variable in all elite individuals is statistically analyzed. The activation frequency of each decision variable is then weighted and updated to the activation probability of the corresponding decision variable in the global probability model to obtain the refined global probability model. In step S2 of this embodiment, as follows... Figure 2 As shown, firstly, the 1-row D-dimensional primacy vector obtained in stage S1 is... Input the K-Means Clustering algorithm, set the number of clusters K=2, and thus merge all... The decision variables are divided into high-importance clusters. and low importance clusters The high-importance cluster contains decision variables that can significantly reduce reconstruction errors in univariate detection. In the tensioner scenario, these decision variables correspond to sampling points at moments of tension abrupt change, brake lock-up, or hydraulic peak. This cluster is typically small in size. The low-importance cluster contains decision variables that have a smaller impact on system performance. These correspond to auxiliary signals during stable operation, such as stable oil temperature and idle speed. This cluster is larger in size.

[0066] In step S2 of this embodiment, the activation probability of each cluster is negatively correlated with the cluster size. When constructing the initial global probability model based on the activation probability, if a decision variable belongs to a high-importance cluster, its initial global probability is set to the activation probability of the high-importance cluster. If the decision variable belongs to a low-importance cluster, its initial global probability is set to the activation probability of the low-importance cluster.

[0067] Specifically, this embodiment is based on the size of the high-importance cluster. and the size of low importance clusters Calculate their respective activation probabilities and The calculation formula is as follows:

[0068]

[0069]

[0070] After calculating the activation probabilities mentioned above, this invention constructs an initial global probability model. If the decision variable Then its initial global probability Conversely, the same applies. Specifically defined as:

[0071]

[0072] This model is used to guide the generation of sparse masks in the initial population, making it easier to activate highly important decision variables, thus making it easier to enter the next stage of elite selection.

[0073] In this embodiment, the initial population contains a total of Individual, each individual From the decision variable vector and sparse mask matrix Obtained by element-wise multiplication, represented as ;in, For the number of individuals.

[0074] In this embodiment, for the first The remaining variables, and their corresponding sparse mask vectors Represented as:

[0075]

[0076] in, That is, a column in the sparse mask matrix; For the first The activation probabilities of the remaining variables in the initial global probability model; To activate a decision variable with a certain probability, when a decision variable is forcibly activated, its corresponding value in the sparse mask matrix is ​​1. When a decision variable is not forcibly activated, the activation probability in the initial global probability model is referenced. Specifically, if the activation probability in the initial global probability model is 0.2, it means that the decision variable will be activated with a probability of 0.2. Finally, the output of the entire initialization process is a... The decision variable matrix, a A sparse mask matrix and a The initial global probability model.

[0077] In step S2 of this embodiment, the generated initial population ( Individuals) and the temporary population generated in S1 ( Each individual (i.e., a probe solution) is merged to form a co-existing solution. From a mixed population of individuals, a non-dominated sorting algorithm and crowding distance calculation are used to select N individuals with the best overall performance, forming the initial elite population. Each individual in the initial elite population actually represents a candidate compression strategy.

[0078] In step S2 of this embodiment, the method for calculating the activation frequency vector is as follows: First, starting from the... From the initial elite population of individuals, all individuals located at the first non-dominant front (FrontNo=1) are selected as elite individuals. Then, the sparse mask vectors of all elite individuals are extracted to form a... Elite Mask Matrix , The total number of elite individuals is given; then, the elite mask matrix is ​​summed column-wise to obtain a matrix with dimension 1. The sum vector is then used, where each element represents the total number of times the corresponding decision variable is activated (masked with a value of 1) among all elite individuals. Subsequently, this sum vector is divided element-wise by the total number of elite individuals. The dimension is obtained as activation frequency vector .

[0079] Specifically, each element in the above activation frequency vector This can represent the tensioning machine. The frequency with which each decision variable dimension is activated in the elite compression strategy reflects its structural contribution to maintaining reconstruction accuracy, sensitivity to abnormal operating conditions, and overall compression performance. This frequency guides subsequent variable importance modeling and grouping strategy adjustment. The specific calculation method is as follows:

[0080]

[0081] in, For the first The decision variable is at the _th ... The number of times an elite individual is activated.

[0082] After obtaining the activation frequency vector, this embodiment further calculates the new activation frequency vector. The model is weighted and fused with the initial global probability model to generate a refined global probability model. Its update formula is as follows:

[0083]

[0084] in, It is a preset, fixed learning rate used in the initial stage. Through this step, the global probabilistic model not only incorporates prior knowledge from the initial exploration but also incorporates empirical data from the best individuals in the initial population, greatly enhancing its reliability.

[0085] In step S2 of this embodiment, the specific method for initially grouping all decision variables according to the refined global probability model is as follows: First, all decision variables are sorted in descending order according to their probability values ​​in the refined global probability model. Then, the sorted decision variables are equally divided into a preset number of variables. In the variable group, an initial global grouping scheme is formed. , For one of the variable groups, For variable group index, This refers to the number of variable groups. In this embodiment, the 500-dimensional decision variables are divided into 20 groups, meaning each group contains 25 decision variables.

[0086] After completing S2, we obtain the three core structures necessary to enter the co-evolutionary cycle: the initial elite population (a high-quality search starting point), the refined global probability model, and the structured initial variable grouping. These results together form the basis of the method in S3 of this invention, making subsequent compression strategy optimization more efficient and robust.

[0087] III. Co-evolution driven by probabilistic models

[0088] S3: Based on the grouping results, perform grouped co-evolution on the elite population and the global probability model to obtain the optimal compressed transmission scheme for compressed transmission signals.

[0089] It should be noted that after completing the initial elite population construction, global probability model refinement, and variable grouping in S2, the algorithm enters the core optimization stage. S3 is the main part of the entire tensioner CAN data compression algorithm. Its goal is to optimize the compression transmission scheme in stages, groups, and progressively strengthens by utilizing a multi-group cooperative evolution mechanism, guided by the prior knowledge provided by the initial variable grouping structure in S2. Essentially, it prioritizes the optimization of highly important variables and progressively optimizes less important variables group by group, ensuring that the method of this invention both preserves the fidelity of key signals and improves compression efficiency.

[0090] It should be noted that in step S3 of this invention, the iteration process of each round of group co-evolution is as follows: for the population generated in the previous round of iteration... Parent selection is performed. Guided by the global probability model, sparse mask evolution is carried out on the selected parent individuals, and decision variable evolution is performed on the currently selected variable group. The evolved sparse mask vector and multidimensional decision variable matrix are used to re-mask and generate offspring individuals. A new population for the next round of iteration is generated based on the offspring individuals and parent individuals. All individuals in the new population located at the first non-dominated frontier are taken as new elite individuals. The activation frequency of each decision variable is recalculated based on the sparse mask vector of the new elite individuals. The activation frequency of each decision variable is then weighted and updated to the activation probability of the corresponding decision variable in the global probability model. At the same time, the next variable group is selected in sequence for the next round of iteration.

[0091] In embodiments of the present invention, each iteration of grouped co-evolution is described in detail in the form of sub-steps, such as... Figure 3 As shown, its sub-steps are as follows:

[0092] S31: First, process the data generated in the previous iteration, which includes... The new population of each individual is considered as the current generation population. And from the current generation population, a binary tournament selection method is used to select the parent generation to generate Each parent individual.

[0093] It should be noted that in step S3 of this invention, for the first iteration, the population currently used... This constitutes the initial elite population. In each subsequent iteration, the population used in the current iteration is generated from the previous iteration. The multi-objective function of tension machine compression includes minimizing transmission bandwidth overhead and minimizing reconstruction error. This invention employs a binary tournament selection method for parent selection to generate... There are several parent individuals. When selecting parents, the selection criteria are: prioritize individuals with higher non-dominant rank (lower rank value); if rank values ​​are the same, prioritize individuals with greater crowding distance, to balance solution convergence and diversity. Because there is replacement selection, the same excellent compaction transport scheme may be repeatedly selected into the mating pool, thus ultimately generating... Each parent individual.

[0094] S32: Perform evolutionary operations on the parent individuals to generate... Each parent individual has several offspring. Evolutionary operations on the parent individual are performed under the guidance of a global probability model. The evolutionary operation includes two parts: sparse mask evolution (used to evolve the sparse mask vector) and decision variable evolution. Since each individual can actually be considered as obtained from the sparse mask vector and the multidimensional decision variable matrix through masking operations, the sparse mask evolution is used to evolve the sparse mask vector, while the decision variable evolution is used to evolve the multidimensional decision variable matrix. The evolved sparse mask vector and multidimensional decision variable matrix can then be used to regenerate offspring individuals through masking operations. It should be noted that because the sparse mask vector and the multidimensional decision variable matrix have different dimensions, in this invention, during masking operations on any individual, the sparse mask vector needs to be broadcast to expand to the same dimension as the multidimensional decision variable matrix. Both the sparse mask evolution and the decision variable evolution include crossover and mutation operations, as detailed below:

[0095] In sparse mask evolution, the sparse mask vectors of two parent individuals are crossed under the guidance of a global probability model. For each sparse mask in the sparse mask vector corresponding to the parent individual, the crossover probability is determined by the activation probability of the decision variable corresponding to that sparse mask in the global probability model. Then, a mutation operation is performed on each sparse mask in the crossed sparse mask vector, with a mutation probability of a preset value (e.g., the mutation probability can be set to 0.5), thereby generating the sparse mask vector of the offspring individual.

[0096] In the evolution of decision variables, the decision variable vectors of the parent individuals are first cross-operated to generate the basis of the decision variable vectors of the offspring individuals. Then, based on the currently selected variable group, a polynomial mutation is performed on the basis of the cross-operated decision variable vectors of the offspring individuals to obtain the decision variable vectors of the offspring individuals. This polynomial mutation only takes effect at the position of the decision variable that is in the variable group and is activated.

[0097] It should be noted that the decision variable vector of the parent individual refers to the vector corresponding to a certain decision variable in the multidimensional decision variable matrix of the parent individual. Since the polynomial mutation of this invention only needs to perform mutation operation on the activated decision variables within the currently selected variable group, the corresponding decision variable vector can be taken from the multidimensional decision variable matrix for decision variable evolution. Then, the evolved decision variable vector is combined with other unevolved decision variable vectors to form the multidimensional decision variable matrix corresponding to the offspring individual. The offspring individual is generated by performing element-wise multiplication of the sparse mask vector of the offspring individual generated by sparse mask evolution into a binary mask.

[0098] In step S3 of this embodiment, the above process is the core evolutionary step, which can be decoupled into two parts: sparse mask evolution and decision variable evolution. During sparse mask evolution, mask 1 indicates that the decision variable should retain high precision during the compression process, while mask 0 indicates that the decision variable can be represented with low precision or compressed according to the algorithm. This process is influenced by the current global probability model. The guidance is as follows. For example, during crossover, each sparse mask determines whether to perform crossover based on the activation probability of the decision variable corresponding to that sparse mask in the global probability model, prioritizing the swapping of "difference bits" composed of high-probability variables; during mutation, it attempts to activate new variables with a preset probability of 0.5 (among the variables that have never been activated, based on...). The value is activated through tournament selection), which can introduce variables that become important in new operating conditions (such as load changes increasing the importance of the current signal); and disable existing variables with a probability of 0.5 (from the already activated variables, based on...). (Selecting the least important value and turning it off) helps improve data compression rate.

[0099] It should be noted that in step S32 of this invention, the basis of the offspring individual decision variable vector... It can be calculated using the following formula:

[0100]

[0101] in, Let be the decision variable vector of the two parent individuals; It is the expansion factor, which is a random number between 0 and 1. and the preset distribution index Decide:

[0102]

[0103] In this embodiment, This is the Simulated Binary Crossover (SBX) commonly used in multi-objective evolution. In a tensioner, the vector includes the compressed quantization level of the tension signal, the filtering parameters of the velocity signal, and the sampling period of the position data.

[0104] For the Basis of decision variable vectors Polynomial mutation is performed according to the following formula only if the decision variable is in the current set of variables to be optimized and the sparse mask bit corresponding to the decision variable is 1:

[0105]

[0106] in, The base after polynomial mutation; and All are disturbances, determined by the distribution index. Decide; and These are the lower and upper bounds of the decision variables, respectively. The main purpose of this step is to perform round-robin optimization of the compression transfer schemes for decision variables assigned to high importance groups, and to postpone the optimization of the compression transfer schemes for decision variables assigned to low importance groups.

[0107] S33: When generated After a certain number of offspring individuals, this embodiment further includes the current... Parent individuals and newly generated individuals The merging of individual offspring forms a group containing A new temporary population of individuals, and selection from the new temporary population. Individuals constitute a new population. This is used for the next iteration. Specifically, the process of generating a new population is as follows: the temporary population is non-dominated and sorted, and then the new population is filled in descending order of superiority (Front 1, Front 2, ...) until the size of the new population reaches or exceeds the target size. If the last selected frontier causes population overcrowding, a minimum distance truncation mechanism is applied to all individuals within that layer. This mechanism calculates the pairwise Euclidean distances between individuals within the layer and iteratively removes the "most crowded" individuals in their neighborhoods (i.e., those with the smallest distance to their nearest neighbors) until the population size is exactly [missing information]. This step is used to ensure that the final compressed transmission scheme balances performance and security.

[0108] S34: With new populations All individuals located at the first non-dominant frontier As new elite individuals, these new elite individuals embody the most valuable variable selection method in the current iteration (i.e., the CAN signal compression transmission scheme); the activation frequency of each decision variable is recalculated based on the sparse mask vector of the new elite individuals, and the activation frequency of each decision variable is weighted and updated to the activation probability of the corresponding decision variable in the global probability model. At the same time, the variable group ranked after the currently selected variable group is used as the variable group selected in the next round of iteration.

[0109] Furthermore, during the above iteration process, the selected variable group is continuously rotated from beginning to end, and when the number of rotations reaches a preset value, a regrouping is triggered, and the decision variables are regrouped in descending order based on the activation probability of each decision variable in the latest global probability model.

[0110] In this embodiment, after the new generation of the population is generated, the algorithm immediately enters an adaptive global probability model update and variable set update phase based on the compression characteristics of the tensioner's CAN signal. Specifically, the current global probability model is updated using the following formula:

[0111]

[0112] in, This is the newly calculated activation frequency vector for the current iteration; This is the updated global probability model. Next, index the variable group to be optimized. The variable set is updated by advancing one position sequentially, thus enabling the next generation of evolution. In this embodiment, a modulo operation is specifically used to implement the cyclic rotation, ensuring that all... Each variable group has a fair chance for optimization. Specifically, the update rule for the variable group index is as follows:

[0113]

[0114] in, This is for modulo operations. Furthermore, whenever the number of existing iterations reaches an integer multiple of a fixed interval, a regrouping is triggered, i.e., the following operation is performed: the input is the latest global probability model that has just been updated in S34. The grouping mechanism will re-rank and divide all decision variables based on the latest importance assessment, generating a new global grouping scheme that is better suited to the problem structure of the current optimization stage. This new scheme will replace the old one to guide the subsequent evolution process. The rotation mechanism ensures that key variables with high-frequency evolutionary needs are not left unoptimized, and that low-frequency triggered safety signals are not permanently ignored.

[0115] After the above process, the present invention uses the last generation population obtained by the grouped co-evolutionary algorithm as the Pareto optimal compressed transmission scheme set. Then, the field control terminal automatically selects the best matching compressed transmission scheme from the compressed transmission scheme set according to the real-time bandwidth of the current communication link (that is, within the current bandwidth allowable range, selects the sparse mask and decision variable combination with the smallest error) and executes data transmission, thereby completing the optimization.

[0116] Therefore, this invention enables the compressed transmission scheme to always adapt to the changes in signal characteristics during the operation of the tensioner by continuously learning probabilistically, rotating variable group sequences, and dynamically regrouping based on signal importance, thereby ensuring that the compressed data still has high recovery accuracy and high sensitivity to abnormal operating conditions.

[0117] Example

[0118] The steps in this embodiment are the same as those in the probabilistic model-driven large-scale sparse multi-objective optimization method shown in steps S1 to S3 above, and will not be repeated here. The main focus is on demonstrating the specific dataset, some specific parameter settings, and implementation results of this embodiment. For ease of description, the method shown in steps S1 to S3 will be referred to as the method of this invention below.

[0119] To evaluate the performance of this invention, this embodiment was tested on eight benchmark problems, SMOP1-SMOP8, and its performance was compared with two existing representative large-scale algorithms and three excellent large-scale sparse algorithms, including LMOCSO, SparseEA, PM-MOEA, MSKEA, and SCEA. LMOCSO is a representative large-scale algorithm based on competitive swarm optimization, SparseEA is the most classic algorithm for solving large-scale sparse multi-objective optimization problems, and the latter three (PM-MOEA, MSKEA, and SCEA) are algorithms that perform well in solving large-scale sparse multi-objective optimization problems. All experiments were implemented on the PlatEMO platform. In this embodiment, the population size for each benchmark problem was set to 100, and the number of evaluations was 100D.

[0120] Regarding the evaluation metrics for algorithm performance, this embodiment uses the Inverse Generational Distance (IGD) metric to assess the solution set quality of the benchmark problems. The IGD value is calculated by uniformly sampling 10,000 reference points on the real Pareto front. Each test instance is run 30 times independently, and statistical analysis is performed using the Wilcoxon rank-sum test at a significance level of 0.05. The symbols “+”, “−”, and “=" after parentheses for other algorithms indicate that the results of the comparison algorithms are significantly better than, worse than, or equivalent to the experimental results of the method proposed in this invention, respectively. As shown in Table 1, the method proposed in this invention achieves the best performance on the eight benchmark problems SMOP1-SMOP8.

[0121] Table 1. Comparison of test results of the present invention and various test algorithms on SMOP1-SMOP8 problems.

[0122] It should also be noted that the probabilistic model-driven large-scale sparse multi-objective optimization method in the above embodiments can essentially be executed by a computer program or module. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a probabilistic model-driven large-scale sparse multi-objective optimization system corresponding to the probabilistic model-driven large-scale sparse multi-objective optimization method provided in the above embodiments, such as... Figure 4 As shown, it includes:

[0123] The initialization and knowledge acquisition module is used to generate a temporary population of probe solutions that only activate each decision variable by using various sensor signals of the tensioner as decision variables through masking operations. The module calculates the importance score of the decision variables activated by each probe solution with the goal of minimizing transmission bandwidth overhead and reconstruction error.

[0124] The initial grouping module is used to cluster decision variables according to importance scores and assign an activation probability to each decision variable that is negatively correlated with the size of its cluster to obtain a global probability model. The initial population is generated by a dual sparse masking mechanism and merged with the temporary population to form an elite population. The activation frequency of each decision variable in the elite population is calculated and the global probability model is updated in a weighted manner. The decision variables are then grouped sequentially with the updated activation probabilities.

[0125] The co-evolution module is used to perform grouped co-evolution on the elite population and the global probability model based on the grouping results, so as to obtain the optimal compressed transmission scheme for compressed transmission signals.

[0126] It is understood that the probabilistic model-driven large-scale sparse multi-objective optimization method described in S1-S3 above can essentially be implemented by a computer program. Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer program product corresponding to the probabilistic model-driven large-scale sparse multi-objective optimization method provided in the above embodiments, which includes a computer program / instructions. When executed by a processor, the computer program / instructions can implement the probabilistic model-driven large-scale sparse multi-objective optimization method as described in the above embodiments.

[0127] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the probabilistic model-driven large-scale sparse multi-objective optimization method provided in the above embodiments, such as... Figure 5 As shown, it includes a memory and a processor;

[0128] The memory is used to store computer programs;

[0129] The processor is configured to implement the probabilistic model-driven large-scale sparse multi-objective optimization method in the above embodiments when executing the computer program.

[0130] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0131] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the probabilistic model-driven large-scale sparse multi-objective optimization method provided in the above embodiments. The storage medium stores a computer program, which, when executed by a processor, can implement the probabilistic model-driven large-scale sparse multi-objective optimization method in the above embodiments.

[0132] Specifically, in the computer-readable storage medium of the above three embodiments, the stored computer program is executed by a processor, which can perform the aforementioned steps S1 to S3.

[0133] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.

[0134] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0135] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the system and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0136] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A probabilistic model-driven method for large-scale sparse multi-objective optimization, characterized in that, include: S1: Using various sensor signals of the tensioner as decision variables, a temporary population is formed by generating probe solutions that only activate each decision variable through masking operations. The importance score of the decision variable activated by each probe solution is calculated with the goal of minimizing transmission bandwidth overhead and reconstruction error. The probe solution is an individual generated by masking operations that only activates a single decision variable. S2: Cluster decision variables based on importance scores and assign each decision variable an activation probability negatively correlated with the size of its cluster to obtain a global probability model; use a dual sparse masking mechanism to generate an initial population and merge it with a temporary population to form an elite population; calculate the activation frequency of each decision variable in the elite population and update the global probability model with weights; group the decision variables sequentially with the updated activation probabilities; use a dual-level operation of forcibly activating seed variables and activating non-seed variables by probability. S3: Based on the grouping results, perform grouped co-evolution on the elite population and the global probability model to obtain the optimal compressed transmission scheme for compressed transmission signals; The importance score of each decision variable is calculated as follows: the number of bits occupied by all activated decision variables during transmission is the transmission bandwidth overhead, and the weighted deviation between the original signal data before compression and the reconstructed signal data after compression is the reconstruction error. A multi-objective performance evaluation function that minimizes both transmission bandwidth overhead and reconstruction error is constructed. The ranking level of each probe solution is calculated by a non-dominated ranking algorithm, and its level value represents the importance score of the activated decision variable. The iteration process of each round of group co-evolution is as follows: Parent individuals are selected from the population generated in the previous round. Under the guidance of the global probability model, sparse mask evolution is performed on the selected parent individuals, and decision variable evolution is performed on the currently selected variable group. Offspring individuals are re-masked using the evolved sparse mask vector and the multidimensional decision variable matrix. A new population for the next round of iteration is generated based on the offspring individuals and parent individuals. All individuals in the new population located at the first non-dominant frontier are taken as new elite individuals. The activation frequency of each decision variable is recalculated based on the sparse mask vector of the new elite individuals. The activation frequency of each decision variable is then weighted and updated to the corresponding activation probability of the decision variable in the global probability model. Simultaneously, the next variable group is selected in sequence for the next round of iteration. In the evolution of sparse masks, for each sparse mask of the parent individual, its crossover probability is determined by the global probability model. Each sparse mask determines whether to perform crossover based on the activation probability of the decision variable corresponding to the sparse mask in the global probability model. Its mutation probability is a preset value, thereby generating the sparse mask vector of the offspring individual. In the evolution of decision variables, the decision variable vectors of the parent individuals are first cross-operated to generate the basis of the decision variable vectors of the offspring individuals. Then, based on the current set of variables to be optimized, a polynomial mutation is performed on the basis of the cross-operated decision variable vectors of the offspring individuals. This polynomial mutation only takes effect at the position of the decision variable that is in the set of variables and is activated.

2. The probabilistic model-driven large-scale sparse multi-objective optimization method as described in claim 1, characterized in that, The temporary population is generated as follows: For the multi-source sensor time-domain signals that need to be fully transmitted in the tensioner CAN bus, a multi-dimensional decision variable matrix is ​​constructed by sampling and assigning values ​​within their respective value ranges. Each decision variable in the multi-dimensional decision variable matrix corresponds to a type of time-domain signal that needs to be transmitted through the tensioner CAN bus. At the same time, for each decision variable representing a sensor signal type, an independent binary feature mask is generated for each decision variable, and a masking operation is performed on the multi-dimensional decision variable matrix to obtain a detection solution that retains only the corresponding decision variable signal value while compressing the signal values ​​of the other decision variables to zero. All detection solutions form a temporary population.

3. The probabilistic model-driven large-scale sparse multi-objective optimization method as described in claim 1, characterized in that, The initial construction method of the global probability model is as follows: all decision variables are clustered into two clusters according to their importance scores, forming a high importance cluster and a low importance cluster. The activation probability of each decision variable within each cluster is obtained by normalizing the inverse of its cluster size. After all decision variables obtain the initial values ​​of their activation probabilities, the initial global probability model is formed.

4. The probabilistic model-driven large-scale sparse multi-objective optimization method as described in claim 2, characterized in that, The method for generating the initial population using a dual sparse masking mechanism is as follows: First, based on the importance scores of the decision variables, a random number of decision variables are forcibly activated using a tournament selection method and used as seed variables. Then, for the remaining decision variables that are not seed variables, they are independently activated according to the activation probabilities in the initial global probability model. Finally, a binary sparse mask vector is generated based on the activation status of the decision variables, and the multidimensional decision variable matrix is ​​masked so that the signal values ​​of the activated decision variables are preserved while the signal values ​​of the rest are compressed to zero, resulting in an individual in the initial population. By continuously repeating the decision variable activation and masking operations, an initial population that meets the population size is generated.

5. The probabilistic model-driven large-scale sparse multi-objective optimization method as described in claim 4, characterized in that, The method for weighted updating of the global probability model is as follows: First, select all individuals located at the first non-dominant frontier from the initial elite population as elite individuals. Then, extract the sparse mask vector of all elite individuals, count the activation frequency of each decision variable in all elite individuals, and weight the activation frequency of each decision variable to update the activation probability of the corresponding decision variable in the global probability model to obtain the refined global probability model.

6. The probabilistic model-driven large-scale sparse multi-objective optimization method as described in claim 1, characterized in that, During the iteration process, the selected variable group is continuously rotated from beginning to end, and when the number of rotations reaches a preset value, it is triggered to regroup, and the decision variables are regrouped in descending order based on the activation probability of each decision variable in the latest global probability model.

7. A large-scale sparse multi-objective optimization system driven by a probabilistic model, characterized in that, This method is used to implement the probabilistic model-driven large-scale sparse multi-objective optimization method as described in any one of claims 1 to 6.

Images