A multi-agent collaborative scheduling method based on whole-house intelligence

Abstract

The application provides a multi-agent collaborative scheduling method based on whole-house intelligence, and belongs to the technical field of whole-house intelligence. The method comprises the following steps: an edge side receives original sensing data sent by a sensing type terminal side device, and pre-processes the original sensing data to obtain pre-processed sensing data; the edge side performs double-track processing on the pre-processed sensing data to obtain structured numerical values and environmental semantic labels; the edge side receives an intention instruction sent by a user, extracts elements from the intention instruction to obtain semantic elements; a skill combination tree is generated through a context skill routing graph, a group of associated skills matched with the user intention can be automatically screened out, and the execution sequence among devices is determined according to physical causal logic, and finally an ordered device execution instruction sequence containing device identification, operation type, operation parameter and execution time point is output by disassembling the skill combination tree, thereby realizing automatic collaborative scheduling and ordered execution among whole-house multiple devices.

Description

Technical Field

[0001] This invention belongs to the field of whole-house smart technology, and specifically relates to a multi-agent collaborative scheduling method based on whole-house smart technology. Background Technology

[0002] Whole-house smart home refers to a system that uses technologies such as the Internet of Things (IoT) and artificial intelligence to connect various smart terminal devices (such as air conditioners and lighting equipment) in the home to a unified network, enabling interconnection and centralized, intelligent management of these devices. As users' demands for living comfort and convenience continue to rise, whole-house smart home technology is evolving from early single-device intelligent control to multi-device collaborative operation, aiming to create a more comfortable and energy-efficient living environment for users through automated scheduling.

[0003] Currently, most mainstream whole-house smart control systems employ keyword matching or simple rule engines to understand user intent. Specifically, when the system receives a user's voice control command, it typically extracts the device name and operation verb from the command and directly calls the corresponding device's control interface to execute a single operation. For example, when a user issues the command "turn on the living room air conditioner," the system only issues a start command to the air conditioner; when the user wants to improve indoor ventilation, they need to say two separate commands: "open the window" and "start the fresh air system."

[0004] The aforementioned existing technologies have significant limitations: the system can only trigger the operation of one device individually for each user command, and cannot automatically coordinate and schedule multiple related devices (such as simultaneously closing windows and turning on the air conditioner) based on the user's comprehensive needs (such as "make the living room cooler"). Furthermore, it cannot determine a reasonable execution order for multiple related devices (e.g., closing the sunshade curtains first to reduce heat radiation, then turning on the air conditioner to improve cooling efficiency). To achieve a coordinated effect throughout the entire house, users must issue multiple independent commands or perform multiple manual operations. The system's low level of intelligence makes it difficult to meet users' needs for a convenient and comfortable living experience. Summary of the Invention

[0005] To address the problems in the background technology, this invention proposes a multi-agent collaborative scheduling method based on whole-house intelligence.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A multi-agent collaborative scheduling method based on whole-house intelligence includes: The edge side receives raw sensing data sent by sensing-type end-side devices and preprocesses the raw sensing data to obtain preprocessed sensing data. The preprocessed perception data is processed in a dual-track manner at the edge to obtain structured numerical values ​​and environmental semantic labels; The edge receives the intent command sent by the user, extracts elements from the intent command to obtain semantic elements, and determines whether the confidence level of the extracted elements is greater than or equal to a preset confidence threshold. If so, the semantic elements are mapped to user demand semantic tags. If not, the intent command and environmental semantic tags are sent to the cloud so that the cloud can generate user demand semantic tags based on the intent command. The edge side processes the user demand semantic tags, environment semantic tags, and structured numerical values ​​based on the pre-constructed contextual skill routing graph to obtain a skill combination tree; The skill combination tree is decomposed at the edge to obtain the device execution instruction sequence.

[0007] Preferably, the raw sensing data is preprocessed, specifically including: For a number of raw sensing data sent by sensing edge devices within a preset time window, calculate the arithmetic mean of the values ​​taken within the window; For two adjacent raw sensing data continuously sent by the same sensing-type end-side device, calculate the absolute value of the difference between the two raw sensing data values. If the absolute value is less than or equal to a preset state change threshold, it is determined to be duplicate data and is removed. Remove abnormal data whose values ​​exceed the preset reasonable range or whose deviation from the values ​​reported by the same type of sensing end-side devices in the same space exceeds the preset deviation threshold.

[0008] Preferably, the dual-track processing includes numerical track processing and semantic track processing; wherein: The numerical trajectory processing transforms the continuous data in the preprocessed sensing data into structured numerical objects with spatial identifiers and timestamps. The semantic track processing transforms each piece of data in the preprocessed perception data into an environmental semantic label based on a preset threshold rule, and binds a corresponding spatial identifier and timestamp to each environmental semantic label.

[0009] Preferably, the semantic elements include: spatial scope elements, core requirement elements, and constraint elements; The expression for calculating the intention confidence score is:

[0010] In the formula, Confidence scores are extracted from elements representing the current intent instruction; Sub-scores representing the identification confidence of spatial extent features; Sub-scores representing the confidence level of identifying core requirement elements; The confidence sub-score represents the identification confidence level of the constraint element; , , These are the preset weight coefficients corresponding to spatial scope elements, core demand elements, and constraint elements, respectively.

[0011] Preferably, the contextual skill routing graph includes: a contextual node layer, a skill node layer, and mapping rules between the contextual node layer and the skill nodes; wherein, the contextual node layer contains a plurality of contextual nodes, and the skill node layer contains a plurality of skill nodes; Constructing a contextual skills routing graph, specifically including: Each context node is assigned a context identifier, a context description, and a set of trigger tags; wherein, the set of trigger tags includes user requirement semantic tags and environment semantic tags required to match the corresponding context node; Assign a skill identifier, skill function description, associated device set, prerequisite condition set, default parameter set, and operation mode field to each skill node; Establish mapping rules between context nodes and skill nodes; each mapping rule includes four categories: association weight, triggering condition, dependency relationship and failure condition.

[0012] Preferably, the edge side processes the user demand semantic tags, environment semantic tags, and structured numerical values ​​based on a pre-constructed contextual skill routing graph to obtain a skill combination tree, specifically including: Based on the user demand semantic tags and environment semantic tags, the weighted matching degree of each context node in the context node layer is calculated, and the context node with the highest weighted matching degree and greater than the preset matching threshold is selected as the target context node. Based on the mapping rules, all skill nodes associated with the target context node are selected to form a candidate skill node set; By combining environmental semantic tags and structured numerical values, the set of prerequisite conditions for each skill node in the candidate skill node set is verified one by one, and skill nodes that meet all prerequisite conditions are retained to form an effective skill node set. Based on the dependencies recorded in the mapping rules, the set of effective skill nodes is hierarchically sorted to construct a tree-like hierarchical structure; By combining the constraint semantic tags in the user demand semantic tags, corresponding running parameters are matched for each effective skill node to obtain a skill combination tree.

[0013] Preferably, the formula for calculating the weighted matching degree is:

[0014] In the formula, This represents the weighted matching degree value of the i-th context node; This represents the subset of tags in the trigger tag set of the i-th context node that match the input tag; This represents the total set of tags that represent the input user demand semantic tags and environment semantic tags; This represents the set of trigger tags bound to the i-th context node; This represents the weight attribute value of label k.

[0015] Preferably, the edge side decomposes the skill combination tree to obtain a device execution instruction sequence, specifically including: Extract all device operation requests corresponding to skill nodes to form a set of device operation requests to be executed; Using a household total power threshold as a hard constraint, power requests exceeding the available power are pruned. For multi-device operation, parameter coordination calibration and oscillation suppression are performed with a preset comfort window as a constraint; Generate an ordered sequence of device execution instructions based on priority and dependencies.

[0016] Preferably, the formula for determining oscillation suppression is:

[0017] In the formula, This indicates the target temperature value corresponding to the equipment operation; This indicates the minimum adjustment step size for temperature-related parameters. This indicates the current ambient temperature value.

[0018] Preferably, the method further includes: If the command execution status of the edge monitoring equipment is not confirmed within a timeout or if an execution failure status is received, a retry mechanism will be triggered. If the retry still fails, the resource arbitrator is notified to re-resolve the conflict and generate an updated instruction sequence. If an operation that has already been performed fails subsequently, resulting in an inconsistent state, a compensation operation is triggered, and a recovery command is issued to restore the system state to consistency.

[0019] The beneficial effects of this invention are: The method of this invention generates a skill combination tree through a contextual skill routing graph, which can automatically filter out a set of associated skills that match the user's intent, determine the execution order between devices according to physical causal logic, and finally output an ordered sequence of device execution instructions containing device identifiers, operation types, operating parameters and execution time points by decomposing the skill combination tree, thereby realizing automated collaborative scheduling and orderly execution among multiple devices in the whole house.

[0020] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description and the drawings. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart of a multi-agent collaborative scheduling method based on whole-house intelligence according to the present invention is shown. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Reference Figure 1 As shown, a multi-agent collaborative scheduling method based on whole-house intelligence includes the following steps: S10. The edge side receives the raw sensing data sent by the sensing end-side device and preprocesses the raw sensing data to obtain preprocessed sensing data. Specifically, the sensing-type edge devices collect raw sensing data within the corresponding space according to a preset collection frequency, and send the raw sensing data to the edge side in real time. The raw sensing data includes: temperature data, humidity data, light intensity data, door and window opening and closing status data, execution-type edge device operating status data, and total household power data.

[0025] Due to the influence of the sensor's own electrical characteristics, environmental electromagnetic interference, and dynamic changes in the acquisition environment during long-term operation, the raw sensing data collected and transmitted by sensing edge devices inevitably contains high-frequency random noise, small fluctuations and repetitive data between adjacent samples, and outlier data caused by sensor malfunctions or communication anomalies. Directly processing this unprocessed raw sensing data will affect the accuracy of the subsequent results.

[0026] Therefore, edge-deployed nodes need to perform preprocessing operations on the received raw sensing data. Specifically, preprocessing methods include noise reduction, deduplication, and outlier removal, which are executed by the edge context manager built into the edge-deployed nodes.

[0027] During noise reduction, the edge context manager calculates the arithmetic mean of multiple raw sensing data values ​​reported by the same sensing-type edge device within a preset time window, based on the sliding window filtering method. This arithmetic mean is then used to replace the real-time value in the current raw sensing data to suppress high-frequency random noise. The sliding window filtering formula is:

[0028] In the formula, This represents the value of the sensed data at time t after noise reduction processing. This represents the number of sampling points included within the preset time window, and is a positive integer. This represents the original sensor data value of the i-th sampling point before time t; The value of the original sensing data at time t; This indicates the preset sampling interval for sensing edge devices.

[0029] During deduplication, the edge context manager calculates the absolute value of the difference between the value of the second data and the value of the first data for two consecutive adjacent sensing data reported by the same sensing-type edge device. If the absolute value is less than or equal to a preset state change threshold, the second data is determined to be duplicate data and discarded; if it is greater than the preset state change threshold, it is determined to be non-duplicate data and retained. The deduplication operation determination formula is:

[0030] In the formula, This represents the value of the raw sensing data reported at the current moment; This represents the raw sensing data value reported at the adjacent previous time point. The time difference between the two times is equal to the sampling interval of the sensing edge device. ; This indicates the preset state change threshold for the corresponding data type.

[0031] Outlier removal involves identifying and removing abnormal data samples from the original sensing data reported by the same type of sensing edge devices that exceed a preset reasonable value range or whose deviation from the values ​​reported by the same type of sensing edge devices in the same space exceeds a preset deviation threshold.

[0032] Furthermore, since the raw sensing data contains precise information about the home environment, uploading it to the cloud would pose a risk of privacy leakage. Therefore, the raw sensing data only flows between sensing-type end-side devices and edge-side deployment nodes throughout the entire process, and is not uploaded to cloud deployment nodes. This data isolation mechanism, which confines the raw sensing data within the boundaries of the home's local network, thereby achieving privacy protection.

[0033] S20. The edge side performs dual-track processing on the preprocessed perception data to obtain structured numerical values ​​and environmental semantic labels. Because the preprocessed sensing data needs to be used in subsequent processes for both conditional matching and logical judgment with discrete semantic descriptions, and for arithmetic operations and quantitative derivations with continuous, precise numerical values, converting all preprocessed sensing data into environmental semantic tags would irreversibly result in the loss of precision in the continuous numerical values. This would prevent consumption patterns that rely on precise numerical values ​​from achieving sufficient resolution to perform accurate calculations. Conversely, retaining only the original numerical values ​​of the preprocessed sensing data without generating environmental semantic tags would prevent consumption patterns that rely on discrete semantic descriptions from completing matching and judgments due to a lack of understandable semantic conditions. Therefore, edge-deployed nodes need to perform dual-track processing on the preprocessed sensing data.

[0034] Specifically, dual-track processing includes numerical track processing and semantic track processing, which are executed by the edge context manager built into the edge-side deployment node.

[0035] During numerical trajectory processing, the edge context manager preserves the continuous data in the preprocessed sensing data, including temperature data, humidity data, light intensity data, and total household power data, as structured numerical objects with spatial identifiers and timestamps. These structured numerical objects contain a spatial identifier, a data type, a numerical value, and a timestamp. The spatial identifier is used to locate the physical space to which the data belongs; the data type is used to distinguish data attribute categories; the numerical value provides the core data source for subsequent quantitative calculations; and the timestamp is used for data time-series alignment.

[0036] During semantic track processing, the edge context manager transforms each piece of data in the preprocessed perception data into standardized environmental semantic label data based on preset whole-house smart domain threshold rules, and binds a corresponding spatial identifier and timestamp to each piece of environmental semantic label data.

[0037] For continuous data, the edge context manager matches preset threshold rules and converts the numerical values ​​into corresponding semantic labels.

[0038] For example, when the preprocessed temperature value of the living room area is 28.5 degrees Celsius, the edge context manager matches the preset temperature threshold rule and generates environmental semantic label data [living room temperature is too high]; when the preprocessed light intensity value of the living room area is 800 lux, the edge context manager matches the preset light threshold rule and generates environmental semantic label data [living room light is sufficient]; when the preprocessed total household power value is 3.2 kWh, the edge context manager matches the preset power threshold rule and generates environmental semantic label data [total household power is sufficient].

[0039] For discrete state data, including door and window opening and closing status data and execution-type end-side device operation status data, the edge context manager directly maps them to environmental semantic label data. For example, when the preprocessed living room area window is in the closed state, the edge context manager directly maps and generates environmental semantic label data [living room window closed]; when the preprocessed living room air conditioner is in standby state, the edge context manager directly maps and generates environmental semantic label data [living room air conditioner standby].

[0040] By performing dual-track processing on the preprocessed sensing data, each piece of preprocessed sensing data exists simultaneously in two forms: structured numerical values ​​and environmental semantic labels. The numerical track fully preserves the true numerical values, providing a data foundation for subsequent calculations; the semantic track maps continuous numerical values ​​into discrete semantic descriptions, enabling the expression of environmental state changes in a labeled manner.

[0041] S30. The edge side receives the intent command sent by the user, extracts elements from the intent command to obtain semantic elements, and determines whether the confidence level of the extracted elements is greater than or equal to a preset confidence threshold. If so, the semantic elements are mapped to user demand semantic tags. If not, the intent command and environmental semantic tags are sent to the cloud so that the cloud can generate user demand semantic tags based on the intent command. Because user expressions are flexible and diverse, making device control decisions directly based on user language can easily lead to misunderstandings of intent. Therefore, it is necessary to extract and determine the confidence level of the intent commands sent by the user, transforming unstructured natural language into standardized semantic elements. These semantic elements include: spatial scope elements, core requirement elements, and constraint elements.

[0042] Specifically, edge-deployed nodes receive intent commands sent by users. These nodes then use a language intent recognition model to extract key elements from these commands, including spatial scope, core requirements, and constraints. Among these: Spatial scope feature extraction refers to the process by which edge-deployed nodes identify spatial terms in the intent commands sent by users and map them to spatial scope semantic labels, such as [living room] and [master bedroom].

[0043] Core requirement element extraction refers to the process by which edge-deployed nodes identify the core requirement verbs and adjectives in the intent commands sent by users and map them into semantic tags for requirement types, such as [cooling] and [ventilation].

[0044] Constraint element extraction refers to the process by which edge-deployed nodes identify constraint class descriptions in user intent commands and map them to constraint semantic labels, such as [power saving] and [quiet].

[0045] Furthermore, due to limitations in the local computing and storage resources of edge-deployed nodes, the built-in language intent recognition model is only optimized for common short commands and typical life scenarios in the smart home field. The model's semantic understanding depth and complex command processing capabilities have limitations, and it cannot cover all forms of user intent commands. Therefore, after completing the semantic elements, it is also necessary to calculate the intent confidence score extracted in this element extraction, the expression of which is:

[0046] In the formula, Confidence scores are extracted from elements representing the current intent instruction; Sub-scores representing the identification confidence of spatial extent features; Sub-scores representing the confidence level of identifying core requirement elements; The confidence sub-score represents the identification confidence level of the constraint element; , , These are preset weight coefficients for spatial scope elements, core requirement elements, and constraint elements, respectively, and their sum is 1. Among them, the weight coefficient corresponding to the core requirement element has the highest value to ensure the priority of identifying core scheduling requirements.

[0047] Next, the edge-side deployment node compares the intent confidence score with a preset confidence threshold. If the intent confidence score is greater than or equal to the preset confidence threshold, the edge-side deployment node determines that the natural language intent recognition model has sufficient processing power. The edge-side deployment node directly maps the extracted semantic elements into user demand semantic tags, which include spatial range semantic tags, demand type semantic tags, and constraint condition semantic tags. If the intent confidence score is less than the preset confidence threshold, the edge-side deployment node determines that the complexity of the intent command sent by the user exceeds the processing power of the natural language intent recognition model. The edge-side deployment node sends the user's intent command and environmental semantic tags to the cloud-side deployment node. The cloud-side deployment node uses a full language model to extract semantic tags and sends the extracted user demand semantic tag data to the edge-side deployment node. In this step, the environmental semantic tags enable the full language model in the cloud to accurately determine the user's true intent based on the current environmental state (e.g., a high living room temperature), avoiding misunderstandings.

[0048] It should be further noted that both the natural language intent recognition model and the full language model are well-known and mature technologies in this field. This article does not provide a detailed description of the specific network architecture, training methods, model parameters, and internal structure of the natural language intent recognition model and the full language model. Those skilled in the art can select the appropriate model for deployment based on the actual application scenario requirements.

[0049] S40. The edge side processes the user demand semantic tags, environment semantic tags, and structured values ​​based on the pre-built contextual skill routing map to obtain a skill combination tree; Since user demand semantic tags can only reflect users' subjective needs and cannot directly correspond to multi-device collaborative scheduling schemes, directly triggering a single device operation through semantic tags makes it difficult to determine the device's action based on the current environmental state, nor can it determine the execution priority between devices. Therefore, it is necessary to rely on a pre-built contextual skill routing graph, with context matching as the core, to complete the selection, arrangement, and parameter adaptation of skills, generating a structured skill combination tree.

[0050] Constructing a contextual skills routing graph, specifically: First, a context node layer is constructed: a context node is an abstract definition of a typical smart home scenario. During the construction process, a unique scenario identifier and scenario description are configured for each context node, and a preset set of trigger tags is bound to each context node. The set of trigger tags contains user requirement semantic tags and environmental semantic tags required to match the corresponding context node. At the same time, a weight attribute is configured for each tag in the set of trigger tags. The weight attribute is divided into required tags with a weight of 1.0 and optional tags with a weight between 0.0 and 1.0, which are used to reflect the degree of contribution of the tag to the context matching.

[0051] After completing the construction of the context node layer, the skill node layer is constructed: the skill node is an abstract definition of reusable atomic operation capabilities. During the construction process, each skill node is configured with a unique skill identifier, skill function description, associated device set, precondition set, default parameter set and operation mode field. The operation mode field can be either passive adjustment class or active energy consumption class, which can be directly read and used in the subsequent resource arbitration process.

[0052] After defining the context nodes and skill nodes, configure mapping rules to establish the association between the two types of nodes: the mapping rules define the association between the context nodes and skill nodes. Each mapping rule contains four types of content: association weight, triggering conditions, dependency relationships, and failure conditions. These are used to limit the calling priority of the corresponding skill node in its context, the environmental conditions required for it to take effect, the prerequisite skill nodes required for execution, and the environmental state rules for dynamic exit.

[0053] Specifically, edge-deployed nodes use a pre-built contextual skill routing graph, taking user need semantic tags, environment semantic tags, and structured numerical values ​​as input. First, a contextual node matching operation is performed. The edge-deployed nodes use both user need semantic tags and environment semantic tags as input tags to calculate the weighted matching degree for each contextual node. The formula for calculating the weighted matching degree is:

[0054] In the formula, This represents the weighted matching degree value of the i-th context node; This represents the subset of tags in the trigger tag set of the i-th context node that match the input tag; This represents the total set of tags that represent the input user demand semantic tags and environment semantic tags; This represents the set of trigger tags bound to the i-th context node; This represents the weight attribute value of tag k, where the value is 1.0 for required tags and a decimal between 0.0 and 1.0 for optional tags; This represents the miss penalty coefficient, which is used to apply partial penalty to uncovered trigger tags to avoid obtaining artificially high scores when the set of trigger tags is large but the actual number of matches is very small.

[0055] Required tags refer to the core triggering conditions that must be met for a scenario node to be established. For example, "[temperature is too high]" is a required tag for a summer cooling scenario. If the current temperature does not reach the threshold of being too high, the scenario should not be activated.

[0056] Optional labels are auxiliary enhancement conditions that make a scenario node valid. Their absence will not prevent scenario matching but will lower the score. For example, "[Outer window closed]" is an optional label for a cooling scenario. Cooling can still occur when the window is open, but the cooling efficiency is higher when the window is closed.

[0057] After the calculation is completed, the edge-deployed nodes select the scenario node with the highest weighted matching degree and a value greater than the preset matching threshold as the target scenario node; if there are multiple scenario nodes with the same weighted matching degree, the scenario node with the higher preset priority is selected as the target scenario node.

[0058] After the target context node is selected, the edge-side deployment node selects all skill nodes associated with the target context node based on the mapping rules in the context skill routing graph, forming a candidate skill node set.

[0059] Subsequently, the edge-deployed nodes verify the set of prerequisites for each skill node in the candidate skill node set one by one. The verification process combines two types of data: environmental semantic tags and structured numerical data. The prerequisites for discrete states are matched and judged by environmental semantic tags, while the prerequisites for continuous numerical values ​​are matched and judged by structured numerical data. Only skill nodes that meet all prerequisites are retained, and skill nodes that do not meet the prerequisites are removed, thus forming the final set of effective skill nodes.

[0060] Preconditions refer to the environmental constraints that must be met before a skill node can be executed, ensuring that the skill is feasible and reasonable under the current physical conditions. For example, the preconditions for the cooling and temperature control skill can be set as "the indoor temperature is greater than 26 degrees Celsius" (verified by structured numerical verification) and "the air conditioner is in standby or off state" (verified by environmental semantic tags). Cooling can only be started when both conditions are met.

[0061] The edge-side deployment nodes are based on the dependencies in the mapping rules. The skill nodes in the effective skill node set are sorted hierarchically. Skill nodes without prerequisite dependencies are set as first-level skill nodes, skill nodes that depend on first-level skill nodes are set as second-level skill nodes, and so on to build a tree-like hierarchical structure.

[0062] "No prerequisites" means that the execution of a skill node does not require waiting for any other skill node to complete first, and can be scheduled immediately. For example, the sunshade and heat insulation skill can directly start the operation of closing the sunshade as soon as sufficient light is detected, without waiting for any other equipment to act first. Therefore, it is set as a first-level skill node in the tree hierarchy, and can respond immediately.

[0063] Pre-dependency refers to the requirement that the execution of certain skill nodes depends on the prior completion of another skill node. This relationship stems from the causal logic of the physical world. For example, the cooling and temperature control skill depends on the sunshade and heat insulation skill, because closing the sunshade curtains first can reduce the solar radiation heat load, making the air conditioner more efficient and energy-saving; if the order is reversed, turning on the air conditioner first and then closing the curtains will cause a large amount of cooling energy to be lost through the windows, resulting in energy waste and reduced control effect.

[0064] After the hierarchical structure is completed, the edge-side deployment nodes combine the constraint semantic tags in the user requirement semantic tags to match the corresponding running parameters for each effective skill node.

[0065] Matching corresponding operating parameters for each active skill node means selecting the parameter combination that best meets the constraint objective from the preset parameter set of the skill node based on the constraint semantic tags in the user demand semantic tags. For example, when the constraint semantic tag includes "[energy saving]", the edge-side deployment node matches a higher set temperature (such as 27 degrees Celsius) and a lower fan speed (such as low fan speed) for the cooling and temperature control skill, so as to meet the user's cooling needs while keeping energy consumption at a low level.

[0066] Finally, the edge-side deployment nodes output a structured skill combination tree. The top layer of the skill combination tree is the target context node, the middle layer is the hierarchical skill nodes, and the bottom layer is the running parameters corresponding to each skill node.

[0067] S50. The edge side decomposes the skill combination tree to obtain the device execution instruction sequence.

[0068] Because the skill combination tree is a logical skill hierarchy structure generated based on context matching, it only contains the functional definitions, operating parameters, and dependencies of skill nodes. It cannot be directly used as the execution instructions for edge devices. Furthermore, if multiple skill nodes' corresponding device operations are executed concurrently, issues may arise such as total power exceeding household electricity safety thresholds, environmental parameters deviating from comfortable ranges due to coupling of multiple device operations within the same comfort domain, and device execution order not conforming to pre-dependency logic. Directly decomposing the skill combination tree and sending it to terminal devices for execution could easily lead to adverse consequences such as electricity safety risks, frequent oscillations in environmental parameters, and scheduling logic execution failures. Therefore, it is necessary for the resource arbitrator built into the edge-side deployed nodes to decompose the skill combination tree and resolve multi-level conflicts, generating a standardized and orderly sequence of device execution instructions to ensure the safety, rationality, and stability of the scheduling process.

[0069] The resource arbitrator first decomposes the received skill combination tree, extracting the execution class end-side device operation set and operating parameters corresponding to all skill nodes to form a set of device operation requests to be executed. Each device operation request is then labeled with four attributes: comfort domain type, rated power value, priority value, and operation mode field. Specifically: the comfort domain type is divided into three categories: temperature control comfort domain, lighting comfort domain, and ventilation comfort domain; the priority value is preset according to the rule that safety-related operations are higher than comfort-related operations, and passive energy-saving operations are higher than active energy-consuming operations; the operation mode field is directly read from the skill node definition, and its value is either passive adjustment type or active energy-consuming type.

[0070] Next, the resource arbitrator performs a first-level power budget constraint check, using the total household power threshold as a hard constraint upper limit, to verify whether the total power of the operations to be performed by the devices is within the currently available power range. Specifically, the resource arbitrator first calculates the sum of the rated power of all operations in the set of device operation requests to be performed, obtaining the total requested power value. Then, the resource arbitrator reads the total household power data from the structured data and calculates the current remaining available power value. The formula for calculating the current remaining available power value is:

[0071] In the formula, This indicates the current remaining available power value; Indicates the total household power threshold; This indicates the current total household power consumption.

[0072] The resource arbitrator compares the total requested power value with the current remaining available power value. If the total requested power value is less than or equal to the current remaining available power value, the power constraint verification is passed, and all pending device operation requests proceed to the next level of judgment. If the judgment condition is not met, a power pruning operation is performed. In the power pruning operation, the resource arbitrator first identifies operations marked as necessary in the set of pending device operation requests. These operations originate from the predefined attributes of skill nodes and cannot be pruned. Subsequently, the resource arbitrator sorts the non-necessary pending device operation requests according to their priority values ​​from low to high, and sequentially removes the lowest priority non-necessary pending device operation requests. The total power value of the remaining pending device operation requests is calculated in real time until the remaining total power value is less than or equal to the current remaining available power value. The pruned pending device operation requests are then retained and proceed to the next level of judgment.

[0073] After completing the first-level power verification, the resource arbitrator performs the second-level comfort domain conflict resolution. Addressing the coupling interference problem of multiple device operations within the same comfort domain, it uses a preset comfort window as a constraint to avoid significant fluctuations and repeated adjustments in environmental parameters. Specifically, the resource arbitrator first groups the device operation requests to be executed according to the comfort domain type, resulting in a temperature control comfort domain operation group, a lighting comfort domain operation group, and a ventilation comfort domain operation group. Then, the resource arbitrator processes each comfort domain operation group one by one, calculating the estimated values ​​of environmental parameters after the combined effect of all operations within the group. Taking the temperature control comfort domain as an example, the formula for calculating the estimated values ​​of environmental parameters under the combined effect of multiple devices is:

[0074] In the formula, This represents the estimated environmental parameters for the temperature control comfort zone, i.e., the steady-state estimated temperature after the combined action of multiple devices. The current ambient temperature value is obtained from structured data. This represents the steady-state temperature change under the sole operation of the j-th temperature control device. This value is pre-calibrated through testing in a standard test environment. This indicates the total number of devices operated within the temperature control comfort zone operation group.

[0075] After calculation, the resource arbitrator adds a confidence level label to the estimated value. When the operation group contains antagonistic equipment operations, such as simultaneous cooling and heating operations, the confidence level label is low. The resource arbitrator then adds an additional manual verification or delayed secondary verification step for this group of operations during the sequence generation phase. Subsequently, the resource arbitrator compares the estimated environmental parameter values ​​with the preset comfort window range. Taking the temperature control comfort domain as an example, the verification judgment formula is:

[0076] In the formula, This indicates the lower limit value of the temperature control comfort window; This represents the upper limit of the temperature control comfort window. If the estimated environmental parameter value is within the comfort window range, the corresponding comfort domain is considered conflict-free, and all operations within the group are retained. If it exceeds the comfort window range, a parameter co-calibration operation is performed. During the parameter co-calibration operation, the resource arbitrator reads the operation mode field of each operation. Following the rule that passive adjustment operations take precedence over active energy consumption operations, it prioritizes retaining the parameters of passive adjustment operations and reduces the intensity of active energy consumption operations until the estimated environmental parameter value after the combined effect falls within the comfort window range. After completing the parameter calibration, the resource arbitrator performs oscillation suppression processing, setting a minimum state holding time and a minimum parameter adjustment step size for each device operation. The minimum state holding time is preset to 5 minutes. Taking the temperature control comfort domain as an example, the oscillation suppression judgment formula is:

[0077] In the formula, This indicates the target temperature value corresponding to the equipment operation; This indicates the minimum step size for adjusting temperature-related parameters. When the above formula is satisfied, the corresponding device operation is cancelled to avoid frequent adjustments that could cause environmental parameter oscillations. Device operation requests that have undergone comfort domain conflict resolution proceed to the next level of judgment.

[0078] After completing the first two levels of conflict resolution, the resource arbitrator executes a three-level execution sequence generation algorithm to generate an ordered sequence of device execution instructions. Specifically, the resource arbitrator first sorts the device operation requests to be executed from highest to lowest priority value, then adjusts the execution order based on the dependencies of skill nodes, with operations having prerequisite dependencies executed after their dependent operations. Subsequently, the resource arbitrator sets a preset execution interval of 3 seconds for adjacent operations to avoid instantaneous power surges caused by multiple devices starting simultaneously. Finally, the resource arbitrator binds each operation to a corresponding device identifier, operation type, operating parameters, and execution time point, forming the final sequence of device execution instructions. The sequence of device execution instructions contains multiple device execution instructions, each containing four types of information: target device identifier, operation type, operating parameters, and execution time point.

[0079] After the device execution command sequence is sent to the corresponding execution-type edge device, the edge deployment node monitors the execution status of each command throughout the process. If the edge deployment node does not receive a confirmation response from the execution-type edge device within a preset timeout period (preset to 5 seconds), or receives an execution failure status, the edge deployment node records the anomaly in the audit log and triggers a retry mechanism. The same device execution command with the same device identifier can be retried a maximum of two times, with a retry interval of 3 seconds. If the retry still fails, the edge deployment node notifies the resource arbitrator to remove the device operation from the set of pending device operation requests. The resource arbitrator then re-executes the secondary and tertiary judgments, generates an updated device execution command sequence, and sends it out, ensuring that the operations of other devices are not affected. If a device operation that has already been performed results in an inconsistent system state due to subsequent operation failures, such as a sunshade being closed but the air conditioner failing to start, the edge-side deployment node triggers a compensation operation and sends a recovery command to the device that has already performed the operation, such as reopening the sunshade to its original position. The recovery command also includes four types of information: target device identifier, operation type, operating parameters, and execution time, so that the system state returns to the consistent state before the operation was performed.

[0080] To further understand the above-mentioned multi-agent cooperative scheduling method, this embodiment takes a summer cooling scheduling scenario in a family living room as an example. To further understand the technical solution of this invention, the specific implementation steps of this method are described in detail below, taking the scenario of summer cooling scheduling in a family living room as an example: In this scenario, the temperature and humidity sensor, light intensity sensor, door and window magnetic sensor, and smart meter deployed in the living room collect data at a preset sampling interval of 30 seconds. After preprocessing, the data is reported to the edge gateway to obtain stable structured values: real-time living room temperature 28.7℃, desktop light intensity 860 lux, living room windows are closed, and the current total power of the household is 1200W. Simultaneously, environmental semantic labels such as "living room temperature is too high", "living room lighting is sufficient", and "living room windows are closed" are generated.

[0081] After a user issues a voice command via the central control screen, "Cool down the living room and keep it quiet," the edge processing unit extracts three semantic elements: spatial scope, core needs, and constraints. An intent confidence score is then calculated using a weighted formula. The preset weights for the three elements are as follows: , , The corresponding confidence sub-scores are as follows: , , Then the confidence score of the intention .

[0082] This value is higher than the preset confidence threshold of 0.75, so the semantic tags for user needs are generated locally on the edge side without the need to upload them to the cloud for processing.

[0083] The edge side performs matching operations based on a pre-built contextual skill routing graph: the input label set contains 4 labels, of which 2 required labels each have a weight of 1.0, and 2 optional labels each have a weight of 0.6; the "Summer Living Room Cooling" context node is bound to 5 trigger labels, and 3 of them are matched this time (2 required labels and 1 optional label), with a penalty coefficient for missing. .

[0084] The weighted matching degree is: The sum of the matching tag weights is 1.0 + 1.0 + 0.6 = 2.6; The total weight of the input labels = 1.0 + 1.0 + 0.6 + 0.6 = 3.2; The total weight of the trigger tag = 1.0 + 1.0 + 0.6 + 0.4 + 0.3 = 3.3; The value is higher than the preset matching threshold of 0.6, so "cooling down the living room in summer" is selected as the target scenario node.

[0085] After precondition verification and hierarchical sorting, a skill combination tree is finally generated, which includes three skills: sunshade and heat insulation, cooling and temperature control, and air purification. The corresponding operating parameters are matched with the "quiet" constraint. The edge-side resource arbitrator decomposes the skill combination tree and obtains three equipment operations to be executed: closing the electric sunshade (rated power 15W), starting the wall-mounted air conditioner (rated power 800W), and running the air purifier silently (rated power 30W).

[0086] Total requested power: 15W + 800W + 30W = 845W; Preset total household power threshold is 3200W; Current remaining available power: 3200W 1200W = 2000W. Since 845W < 2000W, the first-level power budget constraint verification passes directly.

[0087] When performing level 2 comfort domain conflict resolution, the calibrated steady-state temperature reduction for shading and heat insulation is 0.8℃, and the steady-state temperature reduction for cooling and temperature regulation is 2.1℃. The estimated steady-state temperature after the combined action of multiple devices is: This value falls within the preset temperature control comfort window range of 24℃ to 28℃.

[0088] The preset temperature parameter adjustment minimum step is 0.5℃, and the absolute value of the difference between the target temperature and the current temperature is... No operation cancellation is triggered, and there is no risk of frequent adjustment and oscillation.

[0089] Finally, based on priority and dependencies, the system first executes the operation to close the sunshade, then starts the air conditioning after a 3-second interval, and then starts the air purifier after another 3-second interval, generating a sequence of device execution instructions with precise execution timestamps. After the instructions are issued, the edge device monitors the execution status with a 5-second timeout period. If no response is received within the timeout period, a retry is triggered. A single instruction can be retried a maximum of 2 times, with a 3-second interval between retryes.

[0090] Based on the same inventive concept as the above method, this invention also proposes a multi-agent collaborative scheduling system based on whole-house intelligence, comprising: The sensing data preprocessing module is used to receive raw sensing data sent by sensing-type end-side devices at the edge side, and preprocess the raw sensing data to obtain preprocessed sensing data. The dual-track processing module for sensing data is used to perform dual-track processing on the preprocessed sensing data at the edge to obtain structured numerical values ​​and environmental semantic labels. The user intent parsing module is used to receive intent commands sent by users at the edge, extract elements from the intent commands to obtain semantic elements, and determine whether the confidence level of the extracted elements is greater than or equal to a preset confidence threshold. If so, the semantic elements are mapped to user demand semantic tags. If not, the intent commands and environmental semantic tags are sent to the cloud so that the cloud can generate user demand semantic tags based on the intent commands. The contextual skill matching module is used at the edge to process the user demand semantic tags, environment semantic tags, and structured values ​​based on a pre-built contextual skill routing map to obtain a skill combination tree; The device instruction generation module is used to decompose the skill combination tree at the edge to obtain a sequence of device execution instructions.

[0091] Based on the same inventive concept as the above method, the present invention also proposes a device including a memory and a processor, wherein the memory stores computer instructions that can be executed on the processor, and the processor executes the above-described multi-agent collaborative scheduling method based on whole-house intelligence when executing the computer instructions.

[0092] Based on the same inventive concept as the above method, the present invention also proposes a computer-readable storage medium storing computer instructions, which, when executed, can realize the above-mentioned multi-agent collaborative scheduling method based on whole-house intelligence.

[0093] Any references to memory, storage, database, or other media used in the embodiments provided in this invention may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory.

[0094] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0095] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-agent collaborative scheduling method based on whole-house intelligence, characterized in that, include: The edge side receives raw sensing data sent by sensing-type end-side devices and preprocesses the raw sensing data to obtain preprocessed sensing data. The preprocessed perception data is processed in a dual-track manner at the edge to obtain structured numerical values ​​and environmental semantic labels; The edge receives the intent command sent by the user, extracts elements from the intent command to obtain semantic elements, and determines whether the confidence level of the extracted elements is greater than or equal to a preset confidence threshold. If so, the semantic elements are mapped to user demand semantic tags. If not, the intent command and environmental semantic tags are sent to the cloud so that the cloud can generate user demand semantic tags based on the intent command. The edge side processes the user demand semantic tags, environment semantic tags, and structured numerical values ​​based on the pre-constructed contextual skill routing graph to obtain a skill combination tree; The skill combination tree is decomposed at the edge to obtain the device execution instruction sequence.

2. The multi-agent collaborative scheduling method based on whole-house intelligence according to claim 1, characterized in that, The preprocessing of the raw sensory data specifically includes: For a number of raw sensing data sent by sensing edge devices within a preset time window, calculate the arithmetic mean of the values ​​taken within the window; For two adjacent raw sensing data continuously sent by the same sensing-type end-side device, calculate the absolute value of the difference between the two raw sensing data values. If the absolute value is less than or equal to a preset state change threshold, it is determined to be duplicate data and is removed. Remove abnormal data whose values ​​exceed the preset reasonable range or whose deviation from the values ​​reported by the same type of sensing end-side devices in the same space exceeds the preset deviation threshold.

3. The multi-agent collaborative scheduling method based on whole-house intelligence according to claim 1, characterized in that, The dual-track processing includes numerical track processing and semantic track processing; wherein: The numerical trajectory processing transforms the continuous data in the preprocessed sensing data into structured numerical objects with spatial identifiers and timestamps. The semantic track processing transforms each piece of data in the preprocessed perception data into an environmental semantic label based on a preset threshold rule, and binds a corresponding spatial identifier and timestamp to each environmental semantic label.

4. The multi-agent collaborative scheduling method based on whole-house intelligence according to claim 1, characterized in that, The semantic elements include: spatial scope elements, core requirement elements, and constraint elements; The expression for calculating the intention confidence score is: In the formula, Confidence scores are extracted from elements representing the current intent instruction; Sub-scores representing the identification confidence of spatial extent features; Sub-scores representing the confidence level of identifying core requirement elements; The confidence sub-score represents the identification confidence level of the constraint element; , , These are the preset weight coefficients corresponding to spatial scope elements, core demand elements, and constraint elements, respectively.

5. The multi-agent collaborative scheduling method based on whole-house intelligence according to claim 1, characterized in that, The contextual skill routing graph includes: a contextual node layer, a skill node layer, and mapping rules between the contextual node layer and the skill nodes; wherein, the contextual node layer contains a number of contextual nodes, and the skill node layer contains a number of skill nodes; Constructing a contextual skills routing graph, specifically including: Each context node is assigned a context identifier, a context description, and a set of trigger tags; wherein, the set of trigger tags includes user requirement semantic tags and environment semantic tags required to match the corresponding context node; Assign a skill identifier, skill function description, associated device set, prerequisite condition set, default parameter set, and operation mode field to each skill node; Establish mapping rules between context nodes and skill nodes; each mapping rule includes four categories: association weight, triggering condition, dependency relationship and failure condition.

6. The multi-agent collaborative scheduling method based on whole-house intelligence according to claim 5, characterized in that, The edge side processes the user demand semantic tags, environment semantic tags, and structured numerical values ​​based on a pre-constructed contextual skill routing graph to obtain a skill combination tree, specifically including: Based on the user demand semantic tags and environment semantic tags, the weighted matching degree of each context node in the context node layer is calculated, and the context node with the highest weighted matching degree and greater than the preset matching threshold is selected as the target context node. Based on the mapping rules, all skill nodes associated with the target context node are selected to form a candidate skill node set; By combining environmental semantic tags and structured numerical values, the set of prerequisite conditions for each skill node in the candidate skill node set is verified one by one, and skill nodes that meet all prerequisite conditions are retained to form an effective skill node set. Based on the dependencies recorded in the mapping rules, the set of effective skill nodes is hierarchically sorted to construct a tree-like hierarchical structure; By combining the constraint semantic tags in the user demand semantic tags, corresponding running parameters are matched for each effective skill node to obtain a skill combination tree.

7. A multi-agent collaborative scheduling method based on whole-house intelligence according to claim 6, characterized in that, The formula for calculating the weighted matching degree is: In the formula, This represents the weighted matching degree value of the i-th context node; This represents the subset of tags in the trigger tag set of the i-th context node that match the input tag; This represents the total set of tags that represent the input user demand semantic tags and environment semantic tags; This represents the set of trigger tags bound to the i-th context node; This represents the weight attribute value of label k.

8. The multi-agent collaborative scheduling method based on whole-house intelligence according to claim 1, characterized in that, The edge side decomposes the skill combination tree to obtain a device execution instruction sequence, specifically including: Extract all device operation requests corresponding to skill nodes to form a set of device operation requests to be executed; Using a household total power threshold as a hard constraint, power requests exceeding the available power are pruned. For multi-device operation, parameter coordination calibration and oscillation suppression are performed with a preset comfort window as a constraint; Generate an ordered sequence of device execution instructions based on priority and dependencies.

9. A multi-agent collaborative scheduling method based on whole-house intelligence according to claim 8, characterized in that, The formula for determining oscillation suppression is as follows: In the formula, This indicates the target temperature value corresponding to the equipment operation; This indicates the minimum adjustment step size for temperature-related parameters. This indicates the current ambient temperature value.

10. A multi-agent collaborative scheduling method based on whole-house intelligence according to claim 1, characterized in that, After obtaining the device execution instruction sequence, the method further includes: If the command execution status of the edge monitoring equipment is not confirmed within a timeout or if an execution failure status is received, a retry mechanism will be triggered. If the retry still fails, the resource arbitrator is notified to re-resolve the conflict and generate an updated instruction sequence. If an operation that has already been performed fails subsequently, resulting in an inconsistent state, a compensation operation is triggered, and a recovery command is issued to restore the system state to consistency.

Images