Air conditioning control panel program variable simplification method based on deep learning

Abstract

The present invention discloses a method for streamlining program variables of an air-conditioning control panel based on deep learning, comprising the following steps: S1, collecting program variable data and operating environment data of the air-conditioning control panel, preprocessing and constructing an original data set; S2, constructing a dual-stream adaptive analysis architecture to generate an optimized weight distribution of program variables; S3, using a dynamic variant mechanism to nonlinearly adjust the optimized weight distribution of program variables using a variant factor module to generate a set of program variables; S4, using an Actor-Critic structure-based policy network to generate actions for removing or retaining program variables to generate a filtered set of program variables; S5, constructing a dependency topology graph, using an improved minimum spanning tree algorithm to cut redundant paths, and generating a final streamlined set of program variables; S6, automatically generating a new control logic program and deploying it to the air-conditioning control panel. The present invention utilizes deep learning, a dual-stream adaptive analysis architecture and a reinforcement learning strategy to achieve the streamlining of program variables in the air-conditioning control panel.

Description

Technical Field

[0001] The present invention relates to the field of air conditioning control technology, and in particular to a method for streamlining program variables of an air conditioning control panel based on deep learning. Background Art

[0002] With the advancement of modern air conditioning technology, air conditioning control boards, as core components of air conditioning equipment, carry out multiple functions, including operating status monitoring, data processing, and control instruction execution. Embedded programs in air conditioning control boards typically contain a large number of variables, and the design of these variables directly impacts the control board program's storage efficiency, computational speed, and environmental adaptability. However, in the development of existing air conditioning control board programs, a lack of comprehensive analysis of the importance of variables and their dynamic characteristics leads to widespread variable redundancy. This not only consumes a large amount of storage space but also significantly reduces the program's operational efficiency and dynamic environmental adaptability, becoming a significant constraint on further improving air conditioning performance.

[0003] Currently, some existing technologies have proposed variable simplification and optimization methods to address the problem of program variable redundancy. For example, traditional static analysis methods simply rank the importance of variables based on program call relationships and variable usage frequency, and remove low-frequency variables based on threshold rules. However, such methods rely too much on manual experience and lack comprehensive consideration of the global importance of variables and their adaptability to dynamic environments. In addition, static analysis methods are unable to handle the complex correlations between variables and the external operating environment, resulting in unstable simplification results under diverse environmental conditions and even the possibility of mistakenly deleting key variables, thereby affecting the core control functions of the air conditioner.

[0004] Another category of techniques uses heuristic algorithms to optimize program variables. For example, genetic algorithms and ant colony algorithms perform global searches for optimization targets. However, these algorithms typically require significant computational resources and are prone to becoming trapped in local optimal solutions. Furthermore, these heuristic methods often focus on modeling static relationships between variables, ignoring the weight changes of variables in dynamic operating environments and optimizing their adaptability to control logic. This is particularly true in air conditioning operations involving multiple dynamic environmental factors, such as temperature, humidity, and load.

[0005] In recent years, the rise of deep learning technology has provided new insights into program variable optimization. Some studies have attempted to leverage deep neural networks for feature extraction and optimization analysis of program variables. However, existing deep learning-based methods generally employ a single-stream, global analysis architecture and lack fine-grained modeling of dynamic environments, resulting in optimization results that lack robustness in the face of complex environmental changes. Furthermore, these methods often focus on specific tasks and fail to form a closed-loop simplification and optimization loop for the entire lifecycle of program variables.

[0006] To optimize program variable dependencies, existing technologies typically use the minimum spanning tree algorithm or its variants to perform structured modeling and redundant path pruning between variables. However, these algorithms have significant flaws in practical applications. On the one hand, traditional minimum spanning tree algorithms only consider static weights when constructing dependencies and fail to dynamically adjust the pruning threshold to adapt to changes in the importance of program variables. On the other hand, these algorithms' pruning strategies for low-weight paths are too simple and cannot distinguish the key dependencies that may be contained in low-weight paths, resulting in a lack of global coordination in the optimization results. In addition, these methods are prone to under-pruning or over-pruning in complex dependency topologies, which in turn affects the logical integrity and operational efficiency of the air conditioning control panel.

[0007] Furthermore, when generating and verifying optimized program logic, existing technologies mostly rely on manually configured code generation or simple template-based approaches. This results in a lack of flexibility and automation in the code generation process, and the performance of the optimized program cannot be fully verified. Embedded systems operate in complex environments with limited resources. The performance of optimized programs in terms of performance metrics such as storage efficiency, operating speed, and response latency is crucial for the practical application of control panels. However, existing technologies generally lack systematic verification methods for these metrics, making it impossible to fully measure the optimization effect. This further limits the application value of program optimization methods in the development of actual air conditioning control panels.

[0008] Therefore, how to provide a method for streamlining air-conditioning control panel program variables based on deep learning is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0009] One purpose of the present invention is to propose a method for streamlining program variables of air-conditioning control panels based on deep learning. The present invention combines deep learning technology, dynamic variant mechanism, reinforcement learning strategy network and improved minimum spanning tree algorithm to streamline and optimize program variables of air-conditioning control panels, significantly improving program storage efficiency, running speed and dynamic environment adaptability. Through dual-stream adaptive analysis architecture and multi-dimensional weight optimization, variable importance is accurately assessed; reinforcement learning dynamic adjustment strategy is used to intelligently remove redundant variables; through dependency topology modeling, redundant paths are cut, variable logical structure is optimized, and automatic generation and verification of control logic program is realized, ensuring the efficiency, reliability and adaptability of the program.

[0010] According to an embodiment of the present invention, a method for streamlining program variables of an air conditioning control panel based on deep learning includes the following steps:

[0011] S1, collect the program variable data and operating environment data of the air conditioning control panel, pre-process them and construct the original data set;

[0012] S2. Build a two-stream adaptive analysis architecture based on the original dataset. The two-stream adaptive analysis architecture includes a main stream and an auxiliary stream. The main stream performs global importance analysis on program variable data based on a variant network and quantifies the importance score of program variable data using a dynamic screening layer. The auxiliary stream analyzes the weight changes of program variables under different operating environment characteristics based on dynamic environment modeling combined with a multi-head attention mechanism to generate an optimized weight distribution matrix for program variables.

[0013] S3. Using a dynamic variant mechanism and a variant factor module, nonlinearly adjust the optimized weight distribution matrix of the program variables to generate a set of program variables;

[0014] S4. Inputting the program variable set into the policy network based on the Actor-Critic structure to generate the action of removing or retaining the program variables, thereby generating a filtered program variable set;

[0015] S5. Construct a dependency topology graph based on the filtered program variable set and dependency relationships, and use an improved minimum spanning tree algorithm to prune redundant paths to generate a final streamlined program variable set.

[0016] S6. Automatically generate a new control logic program based on the final streamlined set of program variables, use embedded simulation tools to verify the storage efficiency, running speed, and response delay performance of the new control logic program, and deploy the new control logic program to the air conditioning control board.

[0017] Optionally, the S2 specifically includes:

[0018] S21. In the mainstream approach, a variant network based on an improved residual network is used to perform global importance analysis on program variable data. The variant network based on the improved residual network is composed of multiple residual modules, each of which enhances variable feature extraction capability through skip connections to extract preliminary program variable feature vectors.

[0019] S22. Use the dynamic screening layer in the mainstream to quantitatively analyze the preliminary program variable feature vectors. The dynamic screening layer calculates the importance scores of program variables in real time through the dynamic filtering module to generate a global importance score matrix:

[0020]

[0021] Among them, M s (i, j) represents the global importance score of the i-th program variable on the feature dimension of the k-th program variable, σ represents the activation function, N represents the number of feature maps in the dynamic screening layer, and λ in represents the weight of the i-th program variable on the n-th feature map, x i represents the eigenvalue of the i-th program variable, α nrepresents the nonlinear factor, ω n represents the frequency parameter on the nth feature map, f k represents the kth program variable feature dimension, θ n represents the phase shift on the nth feature map, exp represents the exponential function, μ n represents the time decay rate, and t represents the time step;

[0022] S23. In the auxiliary flow, the operating environment characteristics are analyzed based on dynamic environment modeling, and a multi-scale association between program variables and operating environment characteristics is established through environmental labels and operating environment characteristic distribution;

[0023] S24. Use the multi-head attention mechanism to analyze the weight changes of program variables under different operating environments and generate a dynamic weight distribution matrix:

[0024]

[0025] Among them, W d (i,j) represents the weight of the i-th program variable on the j-th environmental feature dimension, a ij Represents the program variable z i and environmental characteristics j The static association score, ρ m (z i ,e j ) represents the m-th head in the multi-head attention mechanism for the program variable z i and environmental characteristics j The weight, ρ m (z p ,e j ) represents the m-th head in the multi-head attention mechanism for the program variable z p and environmental characteristics j The weight, γ m Indicates the amplitude adjustment factor of the mth head, a pj represents the weight of the p-th program variable on the j-th environment feature dimension, h represents the number of heads of the multi-head attention mechanism, M represents the total number of program variables, z i represents the i-th program variable, z p represents the pth program variable, e j represents the jth environmental feature;

[0026] S25. Perform weighted fusion on the global importance score matrix generated by the main stream and the dynamic weight distribution matrix generated by the auxiliary stream to calculate the optimized weight distribution matrix of the program variables:

[0027] P o (i,k,j)=β1·ln(1+M s (i,k)2 )+β2·W d (i,j);

[0028] Among them, P o (i, k, j) represents the optimization weight of the i-th program variable in the k-th program variable feature dimension and the j-th environment feature dimension, β1 and β2 represent the weight distribution coefficients

[0029] S26. Output the optimized weight distribution matrix P of the program variables o (i,k,j).

[0030] Optionally, the S3 specifically includes:

[0031] S31, construct a dynamic variant mechanism, the dynamic variant mechanism includes a dynamic perturbation module, a period adjustment module and a nonlinear weight optimization module, and uses the variant factor module to optimize the weight distribution matrix P of the program variables. o (i, k, j) performs dynamic nonlinear adjustment and calculates the optimized program variable weight matrix:

[0032]

[0033] Where Q(i,j) represents the optimization weight of the i-th program variable on the j-th environment feature dimension, K represents the total number of program variable feature dimensions, and η k Represents the weight adjustment coefficient for the kth program variable feature dimension in the variation factor module, ReLU represents the linear correction function, represents the perturbation factor of the kth program variable feature dimension, ω k represents the frequency parameter of the kth program variable feature dimension, θ k represents the phase shift of the kth program variable feature dimension, f k represents the kth program variable feature dimension, μ k Indicates the decay rate of weight as the feature dimension changes, P o (i, k, j) represents the optimization weight of the i-th program variable in the k-th program variable feature dimension and the j-th environment feature dimension, and exp represents the exponential function;

[0034] The dynamic perturbation module is used to adjust the feature dimension f of the kth program variable k Introducing the perturbation factor of the kth program variable feature dimension Used to simulate the randomness of environmental changes during weight adjustment;

[0035] The periodic adjustment module uses the periodic adjustment function sin(ω k f k +θ k ) captures the periodic fluctuations of the characteristic dimensions of program variables;

[0036] The nonlinear weight optimization module uses the linear correction function ReLU and the dynamic attenuation function exp(-μ k f k ) realize nonlinear adjustment of weight value;

[0037] S32. The dynamic variant mechanism dynamically extracts a set of program variables by combining the optimized program variable weight matrix with the environment characteristics:

[0038]

[0039] in, represents the characteristic value of the i-th program variable in the extracted program variable set, x i represents the initial characteristic value of the i-th program variable, and τ(j) represents the dynamic screening threshold of the j-th environmental feature dimension:

[0040]

[0041] Where M represents the total number of program variables, δ represents the dynamic screening sensitivity parameter, and Q mean Represents the mean of the optimized weights on the j-th environmental feature dimension.

[0042] Optionally, the S4 specifically includes:

[0043] S41, inputting the generated program variable set into a policy network based on an Actor-Critic structure, wherein the policy network includes an action generation module and a value evaluation module;

[0044] S42. In the action generation module, a corresponding action is generated according to the current state of the program variable set, wherein the action includes a decision on whether to remove or retain each program variable;

[0045] S43. In the value evaluation module, the action generated by the action generation module is evaluated for value, and a comprehensive analysis is performed based on the state of the program variable set and the impact of the action on storage efficiency and program performance to generate a value evaluation result.

[0046] S44, optimizing the policy network using a policy optimization algorithm based on the actions generated by the action generation module and the value evaluation results of the value evaluation module;

[0047] S45. Further screening the program variable set according to the optimized policy network, removing unnecessary program variables and retaining key program variables;

[0048] S46. Finally, a filtered program variable set is generated.

[0049] Optionally, the S5 specifically includes:

[0050] S51, receiving the filtered program variable set, constructing an initial variable dependency graph with the program variables as nodes, where the edges of the initial variable dependency graph represent functional dependencies between the program variables;

[0051] S52. Calculate the weights of the edges in the initial variable dependency graph, and use a dynamic correlation analysis method to combine the mutual information of the variables and the call frequency to quantify the dependency weights:

[0052]

[0053] Among them, E(p,q) represents the program variable x p and the program variable x q The dependency weight between them, ζ represents the balance factor, I(x p ,x q ) represents the program variable x p and the program variable x q The mutual information value, F(x p ,x q ) represents the program variable x p and the program variable x q The number of joint calls of , T represents the total number of calls of the program variable;

[0054] S53. Generate a variable dependency topology graph based on the calculated dependency weights, where the edge weights in the variable dependency topology graph represent functional correlations between program variables, and model the structured relationship between program variables from a global perspective.

[0055] S54. Based on the generated variable dependency topology graph, prune redundant paths using an improved minimum spanning tree algorithm. The improvements of the improved minimum spanning tree algorithm include: prioritizing high-weight edges during the calculation process to ensure that the core functional relationships between program variables are preserved; introducing a hierarchical screening mechanism to group dependency weights according to a dynamic range and prioritize pruning redundant paths in low-weight groups; and dynamically adjusting the pruning threshold in each iteration.

[0056] S55. Optimize the variable dependency topology graph pruned by the improved minimum spanning tree algorithm to generate a final streamlined set of program variables.

[0057] Optionally, the S6 specifically includes:

[0058] S61, receiving the final streamlined program variable set, mapping it to the control logic template, and automatically generating a new control logic program based on the functional dependencies and operation logic of the program variables;

[0059] S62, inputting the new control logic program into an embedded simulation tool for simulation verification, where the simulation tool includes a storage efficiency module, an operation speed module, and a response delay module;

[0060] S63. Calculate the storage occupancy rate of the new control logic program using the storage efficiency module, compare the storage usage before and after optimization, and record the storage efficiency improvement ratio.

[0061] S64, using the running speed module to measure the average execution time of the new control logic program under different workloads, and analyzing whether the processing performance of the program meets the operating requirements of the air conditioning control board;

[0062] S65, using the response delay module to measure the time interval from the input signal to the output control signal generation of the new control logic program, to verify whether the real-time performance of the program meets the expected performance indicators;

[0063] S66. Deploy the new control logic program that has been verified by embedded simulation and meets the requirements of storage efficiency, operation speed, and response delay to the air conditioning control board, and perform functional testing in the actual operating environment.

[0064] The beneficial effects of the present invention are:

[0065] First, this invention utilizes a dual-stream adaptive analysis architecture, combining global importance analysis in the primary stream with dynamic environment modeling in the auxiliary stream, to achieve a comprehensive, multi-dimensional evaluation of program variables under different operating environments. The primary stream employs an improved residual network to quantitatively analyze the global importance of variables, while the auxiliary stream uses a multi-head attention mechanism to dynamically model the weight changes of variables under different environmental characteristics, generating an optimized weight distribution matrix. This effectively addresses the existing single-stream architecture's lack of dynamic environment modeling capabilities, ensuring the stability and robustness of variable reduction results.

[0066] Secondly, the dynamic variant mechanism of the present invention realizes the dynamic adjustment of program variable weights by introducing a dynamic perturbation module, a periodic adjustment module, and a nonlinear weight optimization module. This mechanism can not only adapt to complex and changeable operating environments, but also perform fine-grained adjustment of variable weights through nonlinear optimization, thereby improving the dynamic adaptability of the program variable set. Compared with the existing method of relying on static rules to simplify variables, the present invention can more flexibly adjust the importance weights of variables, avoid the accidental deletion of key variables, and remove inefficient and redundant variables, thereby achieving precise optimization of the program variable set.

[0067] Furthermore, the reinforcement learning strategy network of the present invention utilizes the synergistic effect of the action generation module and the value assessment module, utilizing an optimization method based on the actor-critic architecture, to make intelligent decisions about whether to remove or retain program variables. Through the design of an adaptive reward function, this network balances the relationship between storage efficiency, program performance, and environmental adaptability. This addresses the existing problem of relying solely on static analysis or heuristic algorithms and lacking intelligent dynamic optimization, significantly improving the automation and efficiency of variable reduction.

[0068] Regarding variable dependency optimization, the present invention constructs a dependency topology based on dynamic correlation analysis and utilizes an improved minimum spanning tree algorithm to prune redundant paths. Through a hierarchical screening mechanism and a dynamic pruning threshold adjustment strategy, the improved algorithm not only preserves the core functional relationships within the variable set but also effectively removes low-weight redundant paths, optimizing the global structure of the variable set. This approach overcomes the inability of the traditional minimum spanning tree algorithm to dynamically adapt to changes in variable weights, ensuring that the streamlined variable set possesses greater functional integrity and logical coordination.

[0069] Finally, during the generation and verification phase of the control logic program, the present invention achieves rapid generation of optimized control logic programs through an automated templated code generator, while utilizing embedded simulation tools to comprehensively verify the program's performance indicators, such as storage efficiency, operating speed, and response delay. The simulation verification results provide a reliable basis for the actual deployment of the program, and further ensure the program's stability and adaptability through actual operation tests in a hardware environment. Compared to the problem of the lack of comprehensive verification methods in the prior art, the present invention achieves efficient evaluation and performance assurance of the optimized program, ensuring the program's reliability and superior performance in a variety of complex operating environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0071] Figure 1 This is the overall flow chart of the method for streamlining program variables of an air-conditioning control panel based on deep learning proposed by the present invention;

[0072] Figure 2 This is a schematic diagram of the dual-stream adaptive analysis architecture structure of the air-conditioning control panel program variable simplification method based on deep learning proposed in the present invention. DETAILED DESCRIPTION

[0073] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0074] refer to Figure 1 and Figure 2 ,The air conditioning control panel program variable simplification method based on deep learning includes the following steps:

[0075] S1, collect the program variable data and operating environment data of the air conditioning control panel, pre-process them and construct the original data set;

[0076] S2. Build a two-stream adaptive analysis architecture based on the original dataset. The two-stream adaptive analysis architecture includes a main stream and an auxiliary stream. The main stream performs global importance analysis on program variable data based on a variant network and quantifies the importance score of program variable data using a dynamic screening layer. The auxiliary stream analyzes the weight changes of program variables under different operating environment characteristics based on dynamic environment modeling combined with a multi-head attention mechanism to generate an optimized weight distribution matrix for program variables.

[0077] S3. Using a dynamic variant mechanism and a variant factor module, nonlinearly adjust the optimized weight distribution matrix of the program variables to generate a set of program variables;

[0078] S4. Inputting the program variable set into the policy network based on the Actor-Critic structure to generate the action of removing or retaining the program variables, thereby generating a filtered program variable set;

[0079] S5. Construct a dependency topology graph based on the filtered program variable set and dependency relationships, and use an improved minimum spanning tree algorithm to prune redundant paths to generate a final streamlined program variable set.

[0080] S6. Automatically generate a new control logic program based on the final streamlined set of program variables, use embedded simulation tools to verify the storage efficiency, running speed, and response delay performance of the new control logic program, and deploy the new control logic program to the air conditioning control board.

[0081] In this embodiment, S2 specifically includes:

[0082] S21. In the mainstream approach, a variant network based on an improved residual network is used to perform global importance analysis on program variable data. The variant network based on the improved residual network is composed of multiple residual modules, each of which enhances variable feature extraction capability through skip connections to extract preliminary program variable feature vectors.

[0083] S22. Use the dynamic screening layer in the mainstream to quantitatively analyze the preliminary program variable feature vectors. The dynamic screening layer calculates the importance scores of program variables in real time through the dynamic filtering module to generate a global importance score matrix:

[0084]

[0085] Among them, M s (i, j) represents the global importance score of the i-th program variable on the feature dimension of the k-th program variable, σ represents the activation function, N represents the number of feature maps in the dynamic screening layer, and λ in represents the weight of the i-th program variable on the n-th feature map, x i represents the eigenvalue of the i-th program variable, α n represents the nonlinear factor, ω n represents the frequency parameter on the nth feature map, f k represents the kth program variable feature dimension, θ n represents the phase shift on the nth feature map, exp represents the exponential function, μ n represents the time decay rate, and t represents the time step;

[0086] S23. In the auxiliary flow, the operating environment characteristics are analyzed based on dynamic environment modeling, and a multi-scale association between program variables and operating environment characteristics is established through environmental labels and operating environment characteristic distribution;

[0087] S24. Use the multi-head attention mechanism to analyze the weight changes of program variables under different operating environments and generate a dynamic weight distribution matrix:

[0088]

[0089] Among them, W d (i,j) represents the weight of the i-th program variable on the j-th environmental feature dimension, a ij Represents the program variable z i and environmental characteristics j The static association score, ρ m (z i ,e j ) represents the m-th head in the multi-head attention mechanism for the program variable z i and environmental characteristics j The weight, ρ m (z p ,e j ) represents the m-th head in the multi-head attention mechanism for the program variable z p and environmental characteristics j The weight, γ m Indicates the amplitude adjustment factor of the mth head, apj represents the weight of the p-th program variable on the j-th environment feature dimension, h represents the number of heads of the multi-head attention mechanism, M represents the total number of program variables, z i represents the i-th program variable, z p represents the pth program variable, e j represents the jth environmental feature;

[0090] S25. Perform weighted fusion on the global importance score matrix generated by the main stream and the dynamic weight distribution matrix generated by the auxiliary stream to calculate the optimized weight distribution matrix of the program variables:

[0091] P o (i,k,j)=β1·ln(1+M s (i,k) 2 )+β2·W d (i,j);

[0092] Among them, P o (i, k, j) represents the optimization weight of the i-th program variable in the k-th program variable feature dimension and the j-th environment feature dimension, β1 and β2 represent the weight distribution coefficients

[0093] S26. Output the optimized weight distribution matrix P of the program variables o (i,k,j).

[0094] In this embodiment, S3 specifically includes:

[0095] S31, construct a dynamic variant mechanism, the dynamic variant mechanism includes a dynamic perturbation module, a period adjustment module and a nonlinear weight optimization module, and uses the variant factor module to optimize the weight distribution matrix P of the program variables. o (i, k, j) performs dynamic nonlinear adjustment and calculates the optimized program variable weight matrix:

[0096]

[0097] Where Q(i,j) represents the optimization weight of the i-th program variable on the j-th environment feature dimension, K represents the total number of program variable feature dimensions, and η k Represents the weight adjustment coefficient for the kth program variable feature dimension in the variation factor module, ReLU represents the linear correction function, represents the perturbation factor of the kth program variable feature dimension, ω k represents the frequency parameter of the kth program variable feature dimension, θ k represents the phase shift of the kth program variable feature dimension, f k represents the kth program variable feature dimension, μ kIndicates the decay rate of weight as the feature dimension changes, P o (i, k, j) represents the optimization weight of the i-th program variable in the k-th program variable feature dimension and the j-th environment feature dimension, and exp represents the exponential function;

[0098] The dynamic perturbation module is used to adjust the feature dimension f of the kth program variable k Introducing the perturbation factor of the kth program variable feature dimension Used to simulate the randomness of environmental changes during weight adjustment;

[0099] The periodic adjustment module uses the periodic adjustment function sin(ω k f k +θ k ) captures the periodic fluctuations of the characteristic dimensions of program variables;

[0100] The nonlinear weight optimization module uses the linear correction function ReLU and the dynamic attenuation function exp(-μ k f k ) realize nonlinear adjustment of weight value;

[0101] S32. The dynamic variant mechanism dynamically extracts a set of program variables by combining the optimized program variable weight matrix with the environment characteristics:

[0102]

[0103] in, represents the characteristic value of the i-th program variable in the extracted program variable set, x i represents the initial characteristic value of the i-th program variable, and τ(j) represents the dynamic screening threshold of the j-th environmental feature dimension:

[0104]

[0105] Where M represents the total number of program variables, δ represents the dynamic screening sensitivity parameter, and Q mean Represents the mean of the optimized weights on the j-th environmental feature dimension.

[0106] In this embodiment, the S4 specifically includes:

[0107] S41, inputting the generated program variable set into a policy network based on an Actor-Critic structure, wherein the policy network includes an action generation module and a value evaluation module;

[0108] S42. In the action generation module, a corresponding action is generated according to the current state of the program variable set, wherein the action includes a decision on whether to remove or retain each program variable;

[0109] S43. In the value evaluation module, the action generated by the action generation module is evaluated for value, and a comprehensive analysis is performed based on the state of the program variable set and the impact of the action on storage efficiency and program performance to generate a value evaluation result.

[0110] S44, optimizing the policy network using a policy optimization algorithm based on the actions generated by the action generation module and the value evaluation results of the value evaluation module;

[0111] S45. Further screening the program variable set according to the optimized policy network, removing unnecessary program variables and retaining key program variables;

[0112] S46. Finally, a filtered program variable set is generated.

[0113] In this embodiment, the S5 specifically includes:

[0114] S51, receiving the filtered program variable set, constructing an initial variable dependency graph with the program variables as nodes, where the edges of the initial variable dependency graph represent functional dependencies between the program variables;

[0115] S52. Calculate the weights of the edges in the initial variable dependency graph, and use a dynamic correlation analysis method to combine the mutual information of the variables and the call frequency to quantify the dependency weights:

[0116]

[0117] Among them, E(p,q) represents the program variable x p and the program variable x q The dependency weight between them, ζ represents the balance factor, I(x p ,x q ) represents the program variable x p and the program variable x q The mutual information value, F(x p ,x q ) represents the program variable x p and the program variable x q The number of joint calls of , T represents the total number of calls of the program variable;

[0118] S53. Generate a variable dependency topology graph based on the calculated dependency weights, where the edge weights in the variable dependency topology graph represent functional correlations between program variables, and model the structured relationship between program variables from a global perspective.

[0119] S54. Based on the generated variable dependency topology graph, prune redundant paths using an improved minimum spanning tree algorithm. The improvements of the improved minimum spanning tree algorithm include: prioritizing high-weight edges during the calculation process to ensure that the core functional relationships between program variables are preserved; introducing a hierarchical screening mechanism to group dependency weights according to a dynamic range and prioritize pruning redundant paths in low-weight groups; and dynamically adjusting the pruning threshold in each iteration.

[0120] S55. Optimize the variable dependency topology graph pruned by the improved minimum spanning tree algorithm to generate a final streamlined set of program variables.

[0121] In this embodiment, S6 specifically includes:

[0122] S61, receiving the final streamlined program variable set, mapping it to the control logic template, and automatically generating a new control logic program based on the functional dependencies and operation logic of the program variables;

[0123] S62, inputting the new control logic program into an embedded simulation tool for simulation verification, where the simulation tool includes a storage efficiency module, an operation speed module, and a response delay module;

[0124] S63. Calculate the storage occupancy rate of the new control logic program using the storage efficiency module, compare the storage usage before and after optimization, and record the storage efficiency improvement ratio.

[0125] S64, using the running speed module to measure the average execution time of the new control logic program under different workloads, and analyzing whether the processing performance of the program meets the operating requirements of the air conditioning control board;

[0126] S65, using the response delay module to measure the time interval from the input signal to the output control signal generation of the new control logic program, to verify whether the real-time performance of the program meets the expected performance indicators;

[0127] S66. Deploy the new control logic program that has been verified by embedded simulation and meets the requirements of storage efficiency, operation speed, and response delay to the air conditioning control board, and perform functional testing in the actual operating environment.

[0128] Example 1:

[0129] To verify the feasibility of this invention, we applied it to a certain brand of household intelligent air conditioning control system as a test platform. This air conditioning control system contains multiple complex embedded programs with a large number of program variables, and the air conditioner needs to maintain efficient operation under different environmental conditions. However, due to variable redundancy, existing control logic has significant deficiencies in storage space usage, operating speed, and environmental adaptability. The experiment aimed to streamline and optimize the program variables of the air conditioning control panel using the method of this invention and verify the improvement effect on the aforementioned key performance indicators.

[0130] In this example, we first collected operational data and environmental information from the air conditioner control panel, including temperature, humidity, operating mode (cooling, heating, dehumidification), and load. The raw data set contained 500 program variables and 10 different environmental characteristics. The program variables recorded the air conditioner's sensor status, control logic instructions, and energy consumption monitoring information. After preprocessing, this data was used to construct a multidimensional raw data set, providing a foundation for subsequent optimization.

[0131] Next, the raw data is analyzed and processed using a two-stream adaptive analysis architecture. The main stream extracts global importance scores for program variables based on an improved residual network, while the auxiliary stream, combined with a multi-head attention mechanism, generates dynamic weight distributions for these variables under different environmental conditions. Through joint modeling in this two-stream architecture, an optimized weight distribution matrix for program variables is generated, laying the foundation for dynamic adjustment and streamlining of program variables.

[0132] During the optimization process using the dynamic variant mechanism, the system leverages the variant factor module to dynamically adjust the optimization weight distribution, performing fine-grained nonlinear optimization targeting the importance and environmental adaptability of program variables. The optimized set of program variables is then fed into a reinforcement learning policy network based on an actor-critic architecture. With the support of an intelligent decision-making mechanism, 100 redundant variables are removed, resulting in a streamlined set of program variables.

[0133] Subsequently, the dependency topology of the streamlined variable set was modeled using an improved minimum spanning tree algorithm. This improved algorithm further optimized the structure of the program variable set by prioritizing high-weight paths, introducing a hierarchical filtering mechanism, and dynamically adjusting the pruning threshold. This ultimately resulted in a streamlined program variable set consisting of 200 core variables, a 60% reduction from the original set.

[0134] Finally, based on the streamlined set of program variables, the system automatically generated a new air conditioning control logic program. The program's performance was fully verified using embedded simulation tools. The results showed that the optimized program significantly improved storage space usage, execution speed, and response latency. Actual operational testing confirmed that the optimized control logic demonstrated strong stability and adaptability under a variety of environmental conditions, including high temperature, high humidity, and low and high loads.

[0135] Table 1 Comparison of key performance indicators before and after air conditioning control panel program optimization

[0136]

[0137]

[0138] It can be seen from the data in Table 1 above that the optimization method of the present invention has shown significant advantages in multiple key performance indicators, fully verifying its actual effect in the optimization of air-conditioning control panel programs.

[0139] First, the number of program variables decreased by 60%. Before optimization, the program contained 500 variables, but after optimization, only 200 remained. This demonstrates that the combined application of the dual-stream adaptive analysis architecture, dynamic variant mechanism, and reinforcement learning strategy successfully removed a large number of redundant variables, greatly simplifying the program structure and laying the foundation for freeing up storage space.

[0140] Secondly, the storage space usage dropped from 1024KB before optimization to 512KB after optimization, a 50% reduction. This result verifies the direct contribution of program variable simplification to improved storage efficiency. It also shows that in embedded systems, this method can effectively reduce the storage burden of programs and is suitable for resource-constrained hardware environments.

[0141] In terms of running speed, the average running time of the optimized program is 9.2 milliseconds, a 41.77% improvement compared to the 15.8 milliseconds before optimization. This improvement is due to the removal of redundant variables and the optimized variable logic structure, which significantly reduces redundant calculations during program execution, thereby accelerating processing speed and improving system responsiveness.

[0142] Response latency, a key performance indicator for air conditioning control systems, has also been significantly improved. The optimized response latency is now 1.3 milliseconds, a 48% reduction from the 2.5 milliseconds before optimization. This improvement demonstrates that the present invention not only improves program efficiency but also significantly enhances the real-time performance of the air conditioning system, enabling it to quickly respond to user operations and environmental changes, improving the user experience.

[0143] The environmental adaptability score increased from 6.8 before optimization to 9.2 after optimization, a 35.29% improvement. This data demonstrates that the multi-head attention mechanism and dynamic weight adjustment method enable the program to better adapt to changes in dynamic environmental characteristics such as temperature, humidity, and load. The optimized program maintains efficient operation under diverse operating conditions, demonstrating its strong environmental adaptability and robustness.

[0144] Finally, average power consumption dropped from 85 watts before optimization to 78 watts, a reduction of 8.24%. This result shows that variable reduction and logic optimization not only reduce storage and computing resource consumption, but also reduce the overall power consumption of the system, making it more energy-efficient and environmentally friendly, and providing support for the long-term and efficient operation of embedded devices.

[0145] This example verifies the feasibility and effectiveness of the present invention in real-world scenarios by applying it to a certain brand of household intelligent air conditioning control system. Optimization results demonstrate that the present invention significantly reduces the number of program variables and storage space occupied, improves program execution speed and responsiveness, and enhances environmental adaptability and energy savings. Data validates the efficient and stable performance of the optimized program under diverse operating conditions, comprehensively resolving issues such as low storage efficiency, poor operational performance, and insufficient environmental adaptability caused by variable redundancy in traditional methods. This provides a new technical approach and practical basis for optimizing smart homes and embedded systems.

[0146] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for streamlining program variables of air-conditioning control panels based on deep learning, characterized in that: The steps include: S1, collect the program variable data and operating environment data of the air conditioning control panel, pre-process them and construct the original data set; S2. Build a two-stream adaptive analysis architecture based on the original dataset. The two-stream adaptive analysis architecture includes a main stream and an auxiliary stream. The main stream performs global importance analysis on program variable data based on a variant network and quantifies the importance score of program variable data using a dynamic screening layer. The auxiliary stream analyzes the weight changes of program variables under different operating environment characteristics based on dynamic environment modeling combined with a multi-head attention mechanism to generate an optimized weight distribution matrix for program variables. S3. Using a dynamic variant mechanism and a variant factor module, nonlinearly adjust the optimized weight distribution matrix of the program variables to generate a set of program variables; S4. Inputting the program variable set into the policy network based on the Actor-Critic structure to generate the action of removing or retaining the program variables, thereby generating a filtered program variable set; S5. Construct a dependency topology graph based on the filtered program variable set and dependency relationships, and use an improved minimum spanning tree algorithm to prune redundant paths to generate a final streamlined program variable set. S6. Automatically generate a new control logic program based on the final streamlined set of program variables, use embedded simulation tools to verify the storage efficiency, running speed, and response delay performance of the new control logic program, and deploy the new control logic program to the air conditioning control board.

2. The method for streamlining program variables of an air-conditioning control panel based on deep learning according to claim 1, characterized in that: The S2 specifically includes: S21. In the mainstream approach, a variant network based on an improved residual network is used to perform global importance analysis on program variable data. The variant network based on the improved residual network is composed of multiple residual modules, each of which enhances variable feature extraction capability through skip connections to extract preliminary program variable feature vectors. S22. Use the dynamic screening layer in the mainstream to quantitatively analyze the preliminary program variable feature vectors. The dynamic screening layer calculates the importance scores of program variables in real time through the dynamic filtering module to generate a global importance score matrix: Among them, M s (i,k) represents the global importance score of the i-th program variable on the feature dimension of the k-th program variable, σ represents the activation function, N represents the number of feature maps in the dynamic screening layer, and λ in represents the weight of the i-th program variable on the n-th feature map, x i represents the eigenvalue of the i-th program variable, α n represents the nonlinear factor, ω n represents the frequency parameter on the nth feature map, f k represents the kth program variable feature dimension, θ n represents the phase shift on the nth feature map, exp represents the exponential function, μ n represents the time decay rate, and t represents the time step; S23. In the auxiliary flow, the operating environment characteristics are analyzed based on dynamic environment modeling, and a multi-scale association between program variables and operating environment characteristics is established through environmental labels and operating environment characteristic distribution; S24. Use the multi-head attention mechanism to analyze the weight changes of program variables under different operating environments and generate a dynamic weight distribution matrix: Among them, W d (i,j) represents the weight of the i-th program variable on the j-th environmental feature dimension, a ij Represents the program variable z i and environmental characteristics j The static association score, ρ m (z i ,e j ) represents the m-th head in the multi-head attention mechanism for the program variable z i and environmental characteristics j The weight, ρ m (z p ,e j ) represents the m-th head in the multi-head attention mechanism for the program variable z p and environmental characteristics j The weight, γ m Indicates the amplitude adjustment factor of the mth head, a pj represents the weight of the p-th program variable on the j-th environment feature dimension, h represents the number of heads of the multi-head attention mechanism, M represents the total number of program variables, z i represents the i-th program variable, z p represents the pth program variable, e j represents the jth environmental feature; S25. Perform weighted fusion on the global importance score matrix generated by the main stream and the dynamic weight distribution matrix generated by the auxiliary stream to calculate the optimized weight distribution matrix of the program variables: P o (i,k,j)=β1·ln(1+M s (i,k) 2 )+β2·W d (i,j); Among them, P o (i, k, j) represents the optimization weight of the i-th program variable in the k-th program variable feature dimension and the j-th environment feature dimension, β1 and β2 represent the weight distribution coefficients S26. Output the optimized weight distribution matrix P of the program variables o (i,k,j).

3. The method for streamlining program variables of an air-conditioning control panel based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31, construct a dynamic variant mechanism, the dynamic variant mechanism includes a dynamic perturbation module, a period adjustment module and a nonlinear weight optimization module, and uses the variant factor module to optimize the weight distribution matrix P of the program variables. o (i, k, j) performs dynamic nonlinear adjustment and calculates the optimized program variable weight matrix: Where Q(i,j) represents the optimization weight of the i-th program variable on the j-th environment feature dimension, K represents the total number of program variable feature dimensions, and η k Represents the weight adjustment coefficient for the kth program variable feature dimension in the variation factor module, ReLU represents the linear correction function, represents the perturbation factor of the kth program variable feature dimension, ω k represents the frequency parameter of the kth program variable feature dimension, θ k represents the phase shift of the kth program variable feature dimension, f k represents the kth program variable feature dimension, μ k Indicates the decay rate of weight as the feature dimension changes, P o (i, k, j) represents the optimization weight of the i-th program variable in the k-th program variable feature dimension and the j-th environment feature dimension, and exp represents the exponential function; The dynamic perturbation module is used to adjust the feature dimension f of the kth program variable k Introducing the perturbation factor of the kth program variable feature dimension Used to simulate the randomness of environmental changes during weight adjustment; The periodic adjustment module uses the periodic adjustment function sin(ω k f k +θ k ) captures the periodic fluctuations of the characteristic dimensions of program variables; The nonlinear weight optimization module uses the linear correction function ReLU and the dynamic attenuation function exp(-μ k f k ) realize nonlinear adjustment of weight value; S32. The dynamic variant mechanism dynamically extracts a set of program variables by combining the optimized program variable weight matrix with the environment characteristics: in, represents the characteristic value of the i-th program variable in the extracted program variable set, x i represents the initial characteristic value of the i-th program variable, and τ(j) represents the dynamic screening threshold of the j-th environmental feature dimension: Where M represents the total number of program variables, δ represents the dynamic screening sensitivity parameter, and Q mean Represents the mean of the optimized weights on the j-th environmental feature dimension.

4. The method for streamlining program variables of an air-conditioning control panel based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41, inputting the generated program variable set into a policy network based on an Actor-Critic structure, wherein the policy network includes an action generation module and a value evaluation module; S42. In the action generation module, a corresponding action is generated according to the current state of the program variable set, wherein the action includes a decision on whether to remove or retain each program variable; S43. In the value evaluation module, the action generated by the action generation module is evaluated for value, and a comprehensive analysis is performed based on the state of the program variable set and the impact of the action on storage efficiency and program performance to generate a value evaluation result. S44, optimizing the policy network using a policy optimization algorithm based on the actions generated by the action generation module and the value evaluation results of the value evaluation module; S45. Further screening the program variable set according to the optimized policy network, removing unnecessary program variables and retaining key program variables; S46. Finally, a filtered program variable set is generated.

5. The method for streamlining program variables of an air-conditioning control panel based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51, receiving the filtered program variable set, constructing an initial variable dependency graph with the program variables as nodes, where the edges of the initial variable dependency graph represent functional dependencies between the program variables; S52. Calculate the weights of the edges in the initial variable dependency graph, and use a dynamic correlation analysis method to combine the mutual information of the variables and the call frequency to quantify the dependency weights: Among them, E(p,q) represents the program variable x p and the program variable x q The dependency weight between them, ζ represents the balance factor, I(x p ,x q ) represents the program variable x p and the program variable x q The mutual information value, F(x p ,x q ) represents the program variable x p and the program variable x q The number of joint calls of , T represents the total number of calls of the program variable; S53. Generate a variable dependency topology graph based on the calculated dependency weights, where the edge weights in the variable dependency topology graph represent functional correlations between program variables, and model the structured relationship between program variables from a global perspective. S54. Based on the generated variable dependency topology graph, prune redundant paths using an improved minimum spanning tree algorithm. The improvements of the improved minimum spanning tree algorithm include: prioritizing high-weight edges during the calculation process to ensure that the core functional relationships between program variables are preserved; introducing a hierarchical screening mechanism to group dependency weights according to a dynamic range and prioritize pruning redundant paths in low-weight groups; and dynamically adjusting the pruning threshold in each iteration. S55. Optimize the variable dependency topology graph pruned by the improved minimum spanning tree algorithm to generate a final streamlined set of program variables.

6. The method for streamlining program variables of an air-conditioning control panel based on deep learning according to claim 1, characterized in that: The S6 specifically includes: S61, receiving the final streamlined program variable set, mapping it to the control logic template, and automatically generating a new control logic program based on the functional dependencies and operation logic of the program variables; S62, inputting the new control logic program into an embedded simulation tool for simulation verification, where the simulation tool includes a storage efficiency module, an operation speed module, and a response delay module; S63. Calculate the storage occupancy rate of the new control logic program using the storage efficiency module, compare the storage usage before and after optimization, and record the storage efficiency improvement ratio. S64, using the running speed module to measure the average execution time of the new control logic program under different workloads, and analyzing whether the processing performance of the program meets the operating requirements of the air conditioning control board; S65, using the response delay module to measure the time interval from the input signal to the output control signal generation of the new control logic program, to verify whether the real-time performance of the program meets the expected performance indicators; S66. Deploy the new control logic program that has been verified by embedded simulation and meets the requirements of storage efficiency, operation speed, and response delay to the air conditioning control board, and perform functional testing in the actual operating environment.

Images