Intelligent laundry room environment monitoring and control method

Abstract

The invention discloses an intelligent laundry room environment monitoring and control method, which comprises the following steps of: acquiring multi-source data such as equipment state, temperature and humidity, air quality and operation behavior through multiple types of sensors, performing normalization and abnormal value elimination, and generating a standardized data set; constructing a multi-dimensional knowledge graph and introducing a multi-path reasoning and dynamic path weight distribution mechanism by combining a device entity, an environment attribute and a historical abnormal event; based on a current environment state and historical feedback, models such as a graph neural network are adopted to predict an abnormal state, a feedforward control instruction is output according to confidence, responses such as ventilation, dehumidification and equipment pause are realized, reinforcement learning optimization path selection and knowledge graph dynamic updating are supported, the robustness and the adaptive ability in a complex environment are improved, and the reliability of the system is improved. The intelligent level and the exception handling efficiency of laundry room environment monitoring can be remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent industrial environmental monitoring technology, and in particular to an intelligent method for environmental monitoring and control in a laundry room. Background Technology

[0002] With the continuous improvement of industrial intelligence, industrial environments such as laundries are placing higher demands on the monitoring and intelligent control of the operating environment. Currently, environmental monitoring and anomaly prediction mainly rely on real-time collection of equipment operating status, temperature and humidity, air quality, and operational behavior data from multiple types of sensors. This data is then used through traditional rule-driven analysis or knowledge graph methods based on a single reasoning path to identify and respond to abnormal states. Knowledge graphs, as a high-level semantic information organization and reasoning tool, have been widely applied in recent years in fields such as equipment management and industrial process monitoring. Through the structured expression of entity-relation triples, they enable the accumulation of knowledge and logical reasoning about complex industrial environments. In the field of laundry room environmental monitoring and control, a typical technical approach is as follows: First, multi-source monitoring data is collected and normalized to construct a knowledge graph based on key equipment, environmental parameters, and anomaly patterns. Then, based on preset reasoning rules, causal chains are searched from the knowledge graph to complete the reasoning and proactive early warning of abnormal events. Some advanced systems also integrate machine learning or deep learning to assist in the judgment of environmental trends and abnormal risks, thereby improving prediction accuracy. Publicly available literature and patents generally adopt a single-path reasoning model, that is, the environmental state is analyzed and decided step by step according to a fixed chain of knowledge rules, combined with an alarm strategy based on data thresholds to achieve anomaly response. However, existing knowledge graph-based reasoning and decision-making methods generally suffer from the following shortcomings: (1) The reasoning path is highly dependent on the initially set fixed rule links, making it difficult to dynamically adjust the reasoning process for complex environments and multi-factor coupling scenarios that change in real time, resulting in limited adaptability of the system to environmental changes. (2) When complex anomalies that are not standardized and not fully covered occur in the industrial environment, a single reasoning path cannot fully reflect the nonlinear interaction between multiple sources of information, which restricts the comprehensiveness of anomaly identification and the flexibility of decision-making. (3) The reasoning decision lacks a multi-path selection mechanism. When the main reasoning path fails or the response is delayed, the system is unable to switch to the backup reasoning path in a timely manner or form a variety of coping strategies, which affects the robustness and security of the system. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides an intelligent laundry room environmental monitoring and control method.

[0004] The technical solution of this invention is implemented as follows: A method for monitoring and controlling the environment of an intelligent laundry room, comprising: S1: Collect operating status parameters, temperature and humidity, ventilation status and operation behavior data of multiple devices in the laundry room environment, and record the spatial location label and equipment type identifier of each data collection point; S2: Normalize the collected raw data and remove outliers to eliminate differences in the dimensions of different sensors and data drift, and generate a standardized environmental monitoring dataset. S3: Based on standardized environmental monitoring datasets and historical abnormal event records, a multi-dimensional knowledge graph is constructed, which includes equipment entities, environmental attributes, operational behaviors, and abnormal patterns. Relationships between entities are expressed through semantic triples. S4: Introduce a path diversity evaluation mechanism to dynamically assign weights to multiple potential reasoning paths in the knowledge graph and calculate the path selection priority based on the current environmental state and the execution probability of historical paths. S5: Based on the current environmental state and the selected inference path, perform multi-path feedforward inference to generate prediction results of possible abnormal states in the future, and calculate the confidence index of each prediction result. S6: Determine whether there are any abnormal states in the prediction results with a confidence level higher than the preset threshold. If so, generate the corresponding feedforward control command sequence, including ventilation adjustment, dehumidification start-up or equipment pause operation. S7: Send the generated control commands to the laundry room control execution module and record the environmental status feedback data after execution for subsequent inference path effect evaluation and optimization; S8: The path selection strategy is iteratively updated using a reinforcement learning algorithm. The path weights are adjusted based on the execution feedback of historical inference paths to improve the system's adaptability to complex coupled environments. S9: Regularly check the timeliness of entities and relationships in the knowledge graph. If data updates are found to be lagging or relationships are invalid, trigger the knowledge graph incremental update mechanism to ensure the real-time and accuracy of the reasoning basis. S10: When a sudden change in environmental state is detected or the control command does not achieve the expected effect, switch to the backup inference path and generate an emergency intervention command to improve the system's robust response capability to abnormal scenarios.

[0005] The intelligent laundry room environmental monitoring and control method provided in this application has the following beneficial effects: (1) By constructing a multi-dimensional knowledge graph and a dynamic multi-path reasoning mechanism, and combining strategies such as path diversity evaluation, dynamic weight allocation and path memory / context awareness, this invention can dynamically select multiple reasoning paths according to different environmental situations, effectively overcome the limitations caused by a single path, and improve the adaptability and processing flexibility to complex and variable abnormal states. (2) This invention adopts multi-path parallel feedforward reasoning and introduces deep learning methods such as graph attention network and LSTM, which can perform multi-view correlation analysis on environmental monitoring data and historical anomaly patterns, significantly expand the predictable space of anomaly states, and effectively improve the forward-looking prediction ability of unknown combination anomalies through multi-path confidence induction, greatly improving the timeliness and precision of environmental regulation. (3) The present invention integrates reinforcement learning algorithm, dynamically adjusts path selection weight and priority based on the execution feedback of historical reasoning path, realizes adaptive evolution of reasoning model, and can automatically discover and favor new paths that have been verified in history as the scheme runs for a long period of time, and gradually eliminates failed paths, greatly improving the robustness and continuous adaptability in the face of new anomalies and environmental evolution. (4) This invention constructs a highly reliable and consistent data and knowledge base throughout the entire process, from the acquisition, normalization, standardization, and anomaly removal of raw sensor data to the fusion of knowledge entities and incremental maintenance of relationships. Through anomaly detection, moving average filtering, graph topology optimization, and other means, it effectively reduces the probability of false alarms and missed alarms caused by equipment aging, data packet loss, and sensor drift. Attached Figure Description

[0006] Figure 1 This is a flowchart of an intelligent laundry room environmental monitoring and control method according to the present invention; Figure 2 This is a sub-flowchart of an intelligent laundry room environmental monitoring and control method according to the present invention; Figure 3 This is another sub-flowchart of the intelligent laundry room environmental monitoring and control method of the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, the present invention provides an intelligent laundry room environmental monitoring and control method, specifically including: S1: Collect operating status parameters, temperature and humidity, ventilation status and operation behavior data of multiple devices in the laundry room environment, and record the spatial location label and equipment type identifier of each data collection point; S2: Normalize the collected raw data and remove outliers to eliminate differences in the dimensions of different sensors and data drift, and generate a standardized environmental monitoring dataset. S3: Based on standardized environmental monitoring datasets and historical abnormal event records, a multi-dimensional knowledge graph is constructed, which includes equipment entities, environmental attributes, operational behaviors, and abnormal patterns. Relationships between entities are expressed through semantic triples. S4: Introduce a path diversity evaluation mechanism to dynamically assign weights to multiple potential reasoning paths in the knowledge graph and calculate the path selection priority based on the current environmental state and the execution probability of historical paths. S5: Based on the current environmental state and the selected inference path, perform multi-path feedforward inference to generate prediction results of possible abnormal states in the future, and calculate the confidence index of each prediction result. S6: Determine whether there are any abnormal states in the prediction results with a confidence level higher than the preset threshold. If so, generate the corresponding feedforward control command sequence, including ventilation adjustment, dehumidification start-up or equipment pause operation. S7: Send the generated control commands to the laundry room control execution module and record the environmental status feedback data after execution for subsequent inference path effect evaluation and optimization; S8: The path selection strategy is iteratively updated using a reinforcement learning algorithm. The path weights are adjusted based on the execution feedback of historical inference paths to improve the system's adaptability to complex coupled environments. S9: Regularly check the timeliness of entities and relationships in the knowledge graph. If data updates are found to be lagging or relationships are invalid, trigger the knowledge graph incremental update mechanism to ensure the real-time and accuracy of the reasoning basis. S10: When a sudden change in environmental state is detected or the control command does not achieve the expected effect, switch to the backup inference path and generate an emergency intervention command to improve the system's robust response capability to abnormal scenarios.

[0010] Step S1: Collect operating status parameters, temperature, humidity, ventilation status, and operational behavior data of multiple devices in the laundry room environment, and record the spatial location labels and device type identifiers of each data collection point. Specifically, this includes: S1.1: Acquire output signals from multiple types of sensors deployed in the laundry room area, including but not limited to equipment operation status sensors, temperature and humidity sensors, air quality sensors and infrared human presence sensors, to collect multi-source real-time status data in the laundry room environment; For the various types of sensors deployed in the laundry room area, a multi-channel asynchronous acquisition strategy is adopted (parameter: sampling frequency 1Hz~10Hz, dynamically adjusted according to the differences in sensor type) to simultaneously acquire the raw output signals of equipment operation status sensors, temperature and humidity sensors, air quality sensors and infrared human presence sensors. Furthermore, by identifying and driving the sensor interface protocol (parameters: driver library version and communication port number), the initialization of various sensor hardware and the allocation of data buffers are realized, and data frames containing timestamps, raw values ​​and channel identifiers are obtained. Furthermore, a multi-threaded signal listening method is adopted (parameter: number of threads = number of sensor types) to achieve independent reception and real-time temporary storage of signal streams from different sensors, thereby reducing the acquisition delay caused by single-thread blocking and generating a thread-safe queue-style data cache structure. Furthermore, by using a data integrity verification algorithm (parameter: CRC16 check polynomial), error detection and packet loss determination are achieved for each sensor data frame, data frames with verification errors are removed, and a valid data record set is obtained. Furthermore, a time synchronization and alignment algorithm (parameter: NTP server synchronization period 30s) is adopted to achieve precise time alignment of data records from different sensors, and the results of the previous step are transformed into multi-source real-time status data with dual spatial location and time annotation, so as to achieve a high consistency input effect of the collected data in the subsequent signal analysis and normalization processing stage. For example, in an industrial laundry layout, four equipment status sensors (model DSS-10A), three temperature and humidity sensors (model THS-205), two air quality sensors (model AQS-310), and five infrared human presence sensors (model IRS-430) are deployed. When using a multi-channel asynchronous acquisition strategy, the sampling frequency is set to 10Hz for the equipment status sensors, 2Hz for the temperature and humidity sensors, 1Hz for the air quality sensors, and 5Hz for the infrared human presence sensors. Through interface protocol identification and driver loading processes, Modbus, I2C, analog signal ADC, and ZigBee communication drivers are loaded on the acquisition host, and independent buffer spaces are allocated for each type of sensor. During the signal monitoring phase, four threads are set to correspond to the equipment status, temperature and humidity, air quality, and human presence data streams respectively, ensuring that the acquisition latency is less than 100ms. CRC16 algorithm (polynomial 0xA001) verification shows an error rate of 0.02%, and erroneous data is directly discarded. In the time synchronization stage, the local NTP service is synchronized with the sensor RTC clock at 30-second intervals to ultimately form multi-source real-time status data with millisecond-level recording accuracy. Subsequently, in S1.2~S1.6, it is parsed into equipment operating status feature vectors, standardized temperature and humidity parameters, and air quality indicators, ensuring that the data foundation for the knowledge graph construction stage is stable and highly reliable. S1.2: Based on the Modbus RTU protocol, the output signals of the equipment operation status sensors are analyzed to extract equipment operation status parameters such as equipment start / stop status, operating power, fault codes and cumulative running time, so as to form equipment operation status feature vector; The input signals come from the equipment operation status sensors deployed in the laundry area. The physical quantities carried by the signals include start / stop status, operating power, fault codes, and cumulative operating time, and are transmitted to the data acquisition unit via a serial bus with Modbus RTU protocol. The Modbus RTU communication parsing method (parameters: 9600 bps baud rate, 8 data bits, 1 stop bit, no parity bit, CRC-16 check) is used to decode the output signal of the device operation status sensor frame by frame. Furthermore, the function code parsing method (parameters: 0x03 read holding register, 0x04 read input register) is used to distinguish and extract data from the register fields of device start / stop status, operating power, fault code and cumulative running time, and obtain the corresponding raw numerical data group. Furthermore, a register mapping conversion method (parameters: register address table, data type mapping rules) is adopted to realize the conversion of the original value to the physical quantity. The operating power field is converted to kilowatt value through a proportional coefficient, and the cumulative operating time field is converted to hours through an integer value. Furthermore, a feature vector construction algorithm (parameters: field sorting rules, normalization coefficient mapping table) is used to sequentially encode the start / stop status, operating power, fault code, and cumulative operating time into a device operating status feature vector of length 4. Through the above communication parsing and data conversion processing methods, the Modbus RTU raw signal from the previous step is transformed into a standardized device operating status feature vector that can be directly used in knowledge graph construction, thereby realizing the structured expression of device status and accurate input for subsequent semantic modeling. For example, in the implementation of an industrial laundry, the equipment operation status sensor is connected to an RS-485-enabled acquisition module. The Modbus RTU communication parameters are configured as follows: baud rate 9600 bps, 8 data bits, 1 stop bit, no parity check, and CRC-16 check polynomial 0xA001. The start / stop status register is located at address 40001, with a value of 1 indicating operation and 0 indicating stop. The operating power register is located at address 40002, with a unit of 0.1 kW; a value of 250 needs to be multiplied by 0.1 to get 25.0 kW. The fault code register is located at address 40003, with a value of 1001 indicating an over-temperature fault. The cumulative operating time register is located at address 40004, with a unit of hours; a value of 1234 indicates a cumulative operating time of 1234 hours. The communication parsing module reads the values ​​of registers 40001 to 40004, converts them into physical quantities according to a preset mapping table, and encodes them into a feature vector [1, 25.0, 1001, 1234]. The operating power conversion uses the following formula: in Operating power (kW) The original value of the register. This is the scaling factor (0.1). In this example, =250, =0.1, calculated as follows =25.0 kW. This result was fed into the knowledge graph modeling module as part of the feature vector. It was verified that, after 48 hours of continuous operation, the feature vector decoding achieved 100% consistency with the actual device state, and the communication error rate was less than 0.1%, providing accurate state input for subsequent multi-path reasoning. S1.3: The I2C interface protocol is used to read and convert the output data of the temperature and humidity sensor, converting the raw digital signal into Celsius and relative humidity percentage values ​​to obtain standardized temperature and humidity environmental parameters. S1.4: Filter the analog voltage signal output by the air quality sensor and convert it into PM2.5 concentration and CO2 concentration values ​​through a preset linear fitting equation to obtain air quality indicators related to ventilation status; S1.5: Based on the ZigBee wireless communication protocol, receive status reporting events from infrared human presence sensors, identify operational behavior signals indicating whether an operator is present, and generate operational behavior time-series data for subsequent abnormal behavior correlation analysis. S1.6: Add spatial location labels and device type identifiers to the collected sensor data. The spatial location labels are obtained based on a preset coordinate mapping table, and the device type identifiers are obtained by mapping the device model and communication address, so as to realize dual labeling of the physical space and device attributes of the data collection point.

[0011] Step S2 involves normalizing the collected raw data and removing outliers to eliminate differences in the dimensions of different sensors and data drift, thereby generating a standardized environmental monitoring dataset. Specifically, this includes: S2.1: Standardize the format of the collected laundry equipment operating status parameters, temperature and humidity data, ventilation status information and operation behavior records to convert the raw data output by different sensors into a unified timestamp and structured field format to obtain a structured raw dataset. The collected operating status parameters, temperature and humidity data, ventilation status information and operation behavior records of laundry room equipment are converted to a unified time reference using a unified time reference conversion method (parameter: Coordinated Universal Time UTC). This ensures that the time synchronization accuracy between multiple data sources meets the millisecond-level requirements. Furthermore, through field mapping and format conversion algorithms (parameters: data type mapping table, field naming rule base), the field names of various types of raw data are unified and the data types are standardized, resulting in a structured set of fields containing unified field identifiers, data types, and value ranges; Furthermore, a data grouping and indexing construction method (parameters: spatial location label index, device type identifier index) is adopted to realize data fragmentation and fast retrieval based on the location of the collection point and device attributes, and to generate an index tree structure that supports multi-dimensional conditional retrieval; Furthermore, by using a time series resampling algorithm (parameters: sampling interval Δt = 1s, interpolation method = linear interpolation), time alignment and data point interpolation filling of data from devices with different sampling frequencies are achieved, and a complete time series structured data matrix is ​​generated. Furthermore, a unit standardization conversion method (parameters: temperature unit °C, relative humidity unit %, gas concentration units μg / m³ and ppm) is adopted to achieve unified conversion of the original units of different sensors to the preset standard units, and to generate a structured data set in which each field conforms to the system's preset unit specifications; Through the above-mentioned time base conversion, field unification, index construction, time series alignment and unit conversion processing, the multi-source heterogeneous raw data in the previous step is transformed into a raw dataset with a unified timestamp and structured field format, realizing the seamless integration of multi-source data in the subsequent normalization and outlier removal process. For example, in an industrial laundry environmental monitoring system, the equipment operation status sensor uses a 100ms sampling period, the temperature and humidity sensor uses a 1s sampling period, the air quality sensor uses a 2s sampling period, and the infrared human presence sensor reports based on event triggers. The raw timestamps output by each device are based on the local RTC, resulting in a ±500ms difference. A unified time base conversion method is used to convert all data timestamps to UTC millisecond values, and linear interpolation is used to align the temperature, humidity, and air quality data to a 1s interval. During field mapping, the equipment operation status fields "power_on" and "state" are unified as "device_status," the temperature field is uniformly named "temperature_C" with a float type, the humidity field is named "humidity_RH" with a float type, and the air quality fields "pm25" and "co2" are unified as "pm25_ug_m3" and "co2_ppm," respectively. During index construction, a collection point location index is generated based on a preset coordinate mapping table, and a type index is generated based on the device type mapped by the communication address. During unit conversion, temperature was converted from its original °F to °C, humidity from its original numerical value to a percentage, and the corresponding air quality value was converted from the original ADC voltage to μg / m³ and ppm according to the calibration curve. After the above processing, the final structured raw dataset contains fields such as timestamp UTC_ms, device_status, temperature_C, humidity_RH, pm25_ug_m3, CO2_ppm, position_id, and device_type, with a data integrity rate of 99.8%. This provides a high-quality input with consistent timing and standardized units for the subsequent S2.2 sliding window mean filtering and denoising process. S2.2: The sliding window mean filtering algorithm is used to denoise the temperature and humidity sensor data in the structured original dataset to eliminate short-term abnormal fluctuations caused by environmental fluctuations or sensor noise, and to obtain smoothed temperature and humidity time series data. S2.3: Based on the Z-score normalization method, the smoothed temperature and humidity time series data, equipment operating parameters and ventilation status indicators are normalized to eliminate the differences in the dimensions of different sensors and obtain a standardized numerical dataset within a unified dimension range. S2.4: Apply the Isolation Forest anomaly detection algorithm to identify outliers in each dimension of the standardized numerical dataset, in order to discover and mark data points that deviate from the normal operating mode, and generate an outlier labeling matrix; S2.5: Based on the outlier labeling matrix, outlier removal and linear interpolation compensation operations are performed on the standardized numerical dataset to eliminate the interference of outlier data on the subsequent inference process, and finally generate a high-quality standardized environmental monitoring dataset.

[0012] Step S3: Based on standardized environmental monitoring datasets and historical anomaly event records, construct a multi-dimensional knowledge graph containing equipment entities, environmental attributes, operational behaviors, and anomaly patterns, where relationships between entities are expressed through semantic triples. For example... Figure 2 As shown, it specifically includes: S3.1: Perform entity recognition processing on the equipment operating status parameters, temperature and humidity, ventilation status and operation behavior data in the standardized environmental monitoring dataset to extract basic entity nodes of laundry equipment, environmental attributes and operator behavior; S3.2: Based on the timestamp, device identifier, anomaly type and handling measures fields in the historical anomaly event records, the anomaly events are classified, labeled and semantically abstracted to generate anomaly pattern entity nodes, which serve as key reasoning basis nodes in the knowledge graph; S3.3: The semantic role labeling algorithm is used to extract the semantic relationships between entities. Based on the co-occurrence frequency of entities and the semantic association strength of the context, a set of triplet relationships such as device-environment, environment-operation, and operation-abnormality is generated. Based on the basic entity nodes and abnormal pattern entity nodes obtained from S3.1 and S3.2, a semantic role labeling algorithm (parameters: window size 7, part-of-speech tagging model is CRF) is used to assign roles to the text descriptions between entities in the standardized environmental monitoring data, so as to realize the semantic function labeling mapping of each entity in the context statement. Furthermore, by using a statistical analysis method based on entity co-occurrence frequency (parameters: time window 1 hour, co-occurrence threshold ≥ 3 times), the synchronous occurrence pattern recognition of equipment entities and environmental attribute entities in time-series data is realized, and a co-occurrence frequency matrix reflecting the correlation between the two is generated. Furthermore, a contextual semantic association strength calculation method (algorithm: point mutual information PMI, parameter: smoothing coefficient 0.1) is adopted to fuse the input co-occurrence frequency matrix and text semantic vector to obtain semantic association strength indices between entity pairs such as device-environment, environment-operation, and operation-anomaly. Furthermore, based on the ranking results of the semantic association strength index, a relation extraction generation algorithm (rule: the top 10% of entity pairs in terms of association strength) is used to generate a set of semantic triples for three types of relations: device-environment, environment-operation, and operation-anomaly. Each triple consists of a subject entity, a relation predicate, and an object entity. The set of triples output by relation extraction results transforms the semantic role labeling and co-occurrence frequency analysis results from the previous step into structured, multi-class relation data that can be used for graph construction, achieving high semantic accuracy in entity relation modeling. For example, for a set of monitoring records including a washing machine WM-100, a temperature and humidity sensor TH-200, and an operator ID-OP35, the input data includes: device start / stop logs, temperature and humidity sampling values ​​(24°C, 78% relative humidity), operation behavior records (opening the lid to retrieve clothes), and abnormal event logs (humidity exceeding the limit alarm). Using a semantic role labeling algorithm, "WM-100 opens the lid to retrieve clothes by OP35 when humidity is 78%" is parsed as: subject device entity WM-100, predicate "under humidity conditions", object environmental attribute humidity 78%, and operation action "opening the lid to retrieve clothes" associated with operation entity OP35. Based on co-occurrence frequency analysis, WM-100 co-occurred 5 times with records of humidity >75% within a 1-hour window, exceeding the threshold; the semantic association strength is calculated using PMI. The result was 2.35, which is higher than the set threshold of 2.0. Therefore, it was extracted as a relation triple (WM-100, under humidity conditions / operation-abnormal, humidity > 75% / humidity exceeds the standard). Finally, the device-environment triple (WM-100, operatesUnderHumidity,Humidity78%), the environment-operation triple (Humidity78%, triggersAction, OP35_openLid), and the operation-abnormal triple (OP35_openLid, leadsTo, HumidityExceedAlert) were generated. In the application, it was verified that it can correctly trigger the feedforward path of ventilation adjustment and dehumidification control in knowledge graph reasoning. S3.4: Perform knowledge fusion processing on the generated triple set, and use a graph embedding-based similarity matching algorithm to eliminate synonymous entities and redundant relationships in order to form a multi-dimensional knowledge graph topology structure under a unified namespace; Based on the set of triplet relationships such as device-environment, environment-operation, and operation-anomaly obtained from S3.3, a graph embedding similarity matching algorithm (parameters: node embedding dimension d=128, number of training iterations=500) is used to realize the quantitative calculation of entity semantic similarity. Furthermore, a low-dimensional vectorized representation of entities and relations in triples is achieved through a node embedding model (method: TransE, parameters: loss function margin=1.0, learning rate 0.01), and the cosine similarity score matrix of each entity pair is obtained. Furthermore, by filtering with a similarity threshold (parameter: threshold τ=0.92), synonym entity recognition is achieved, and a synonym entity merging mapping table is generated for disambiguation of node identifiers under a unified namespace; Furthermore, redundant relations are detected by similarity matching of relation embedding vectors (method: ComplEx complex embedding model, parameter: dimension d=256), and the set of semantically repetitive relation edges is marked. By merging the synonym entity mapping table and the redundant relation tag set, the triple set of the previous step is transformed into a knowledge graph relation set that is de-redundant and semantically consistent, thereby realizing a multi-dimensional knowledge graph topology structure under a unified namespace. For example, in an industrial laundry environmental monitoring system, TransE embedding training was performed on the collected triplet set. The embedding dimension was set to 128, the learning rate to 0.01, and the margin parameter to 1.0. After 500 training iterations, the average cosine similarity on the validation set reached 0.94. A similarity filtering threshold τ was set to 0.92, merging nodes with highly similar vector representations such as "dryer" and "drying machine," as well as "dehumidification equipment" and "drying and dehumidification system." In the relation deduplication operation, a ComplEx model was used to train the relation embedding with a dimension of 256. Semantic duplicate edges such as "temperature and humidity increase → dehumidification equipment starts" and "high humidity → dehumidification equipment is activated" were removed through similarity detection. In the processed knowledge graph topology, the total number of nodes decreased from 153 to 147, the number of relation edges decreased from 824 to 790, and the average query response time decreased from 35ms to 27ms, supporting efficient querying and dynamic updates of the path diversity assessment module. S3.5: Based on the graph database Neo4j, perform persistent storage operations on the knowledge graph, optimize the index and expand the attributes of entity nodes and relation edges in the graph to support efficient querying and dynamic updates of the subsequent multi-path reasoning module.

[0013] Step S4: Introducing a path diversity evaluation mechanism to dynamically assign weights to multiple potential reasoning paths in the knowledge graph, and calculating path selection priority based on the current environmental state and historical path execution probabilities. For example... Figure 3 As shown, it specifically includes: S4.1: Perform topological analysis on the entity relationship network in the knowledge graph, and use a graph traversal algorithm to identify the set of potential reasoning paths associated with the current environment state in order to construct a candidate reasoning path library; S4.2: Based on the execution feedback data of historical reasoning paths, Bayesian statistical methods are used to calculate the execution probability of each path under similar environmental conditions in order to evaluate its historical effectiveness; S4.3: Perform similarity matching between the current environment state feature vector and the historical path context features, and use the cosine similarity algorithm to calculate the path suitability index to quantify the applicability of the path in the current context; S4.4: Based on the path execution probability and path suitability index, a weighted summation algorithm is used to generate dynamic weight values ​​for each candidate path to reflect its priority in the current inference task. Based on the path execution probability data obtained in S4.2 and the path fit index calculated in S4.3, the input objects are two sets of quantized values ​​corresponding to each path in the candidate inference path library; A weighted summation algorithm is used (parameter: path execution probability). Path adaptability Execution probability weighting coefficient Fit weight coefficient This allows for the merging of two types of indicators into a single priority score. Furthermore, by defining the weighted summation formula: Implement the calculation of dynamic path weights, where The final dynamic weight value of the candidate path; Furthermore, the weight parameter setting module automatically adjusts the settings based on the system's preset strategy or environmental characteristics. and The proportion of these factors is used to enhance the ability to prioritize high-confidence historical paths or highly fit paths. Furthermore, the candidate paths are processed through matrix transformation. and Values ​​are assembled into a two-dimensional index matrix according to path indices, and parallel vector weighting operations are called to accelerate all paths. The process of calculating the value; Output by weighted summation algorithm The value vector transforms the execution probability and fitness results of the previous step into a priority score reflecting each candidate path in the current inference task, realizing the importance ranking of different inference path selections and providing a quantitative basis for multi-path sampling and feedforward inference. For example, in an industrial laundry scenario, the candidate inference path library contains 5 paths and their historical execution probabilities. Adaptability index Weighting coefficient settings , .

[0014] Using the weighted summation formula: The values ​​of each path were calculated separately. Value is .

[0015] During the calculation process, and The index matrix is ​​assembled into a 5×2 matrix, and weighting and summation are performed in batches using parallel vector operations, with a single calculation latency of less than 2ms. The final generated... The value vector is used for normalization and probability wheel sampling in the subsequent S4.5 step, thereby ensuring that paths with high historical validity and high adaptability are preferentially selected in the multi-path inference sequence. The verification results show that the path selection accuracy of this strategy is improved by more than 12% under different environmental perturbation conditions. S4.5: Normalize the dynamic weight values ​​and perform path sampling based on the probability roulette selection strategy to generate a priority sequence for multi-path inference, which can then be called by the subsequent multi-path feedforward inference module.

[0016] Step S5: Based on the current environmental state and the selected inference path, perform multi-path feedforward inference to generate prediction results of possible abnormal states in the future, and calculate the confidence index of each prediction result. Specifically, this includes: S5.1: Based on the current laundry room environment state vector, including equipment operating parameters, temperature and humidity values, ventilation status and operation behavior identifiers, perform multi-path matching operation in the knowledge graph to obtain a set of multiple reasoning paths semantically associated with the current state; Based on the current laundry room environment state vector, the input data includes equipment operating parameters, temperature and humidity values, ventilation status and operation behavior identifiers, and multi-path matching operation is performed in the knowledge graph. The state vector feature extraction method is adopted (parameters include equipment operating power and temperature). (Including humidity percentage, PM2.5 concentration, CO2 concentration, and operational behavior coding), to achieve standardized mapping of environmental status in the semantic space of the knowledge graph; Furthermore, a path retrieval algorithm based on depth-first search (DFS) (parameters: maximum traversal depth = 5, relation type filtering set = [device-environment, environment-operation, operation-anomaly]) is used to match and search the state vector with the knowledge graph relation network, and obtain an initial set of candidate reasoning paths; Furthermore, through a state semantic matching algorithm (parameter: cosine similarity threshold) (Vector dimension = 128) to calculate the similarity between the current state and the knowledge graph node set, filter out node pairs with high semantic relevance, and generate a subset of associated paths; Furthermore, by using a path structure consistency verification method (parameters: path length range 2~6, relation direction consistency constraint = TRUE), the candidate path topology is compared with the predefined inference specification, and a set of multiple inference paths that conform to the structural constraints is output. Through the above matching and filtering operations, the entity nodes corresponding to the state vector are linked to the highly relevant multi-path set in the knowledge graph, so as to realize the preparation of multi-path reasoning input for the current environmental state. For example, in an industrial laundry environment, the equipment operating parameters are: operating power Cumulative running time minutes; temperature is Temperature: Celsius, Humidity: PM2.5 concentration was CO2 concentration is ppm; the operational behavior is encoded as presence and execution of device loading actions. After being input into the semantic mapping module, the state vector is mapped to a 128-dimensional embedding. Using DFS retrieval with a traversal depth limit of 5, the cosine similarity of path P1 (high humidity operation → humidity exceeding threshold → mold risk) in the device-environment-abnormality relationship chain is [value missing]. The path P2 (high power + poor ventilation → temperature rise → risk of equipment overheating) has a similarity of [value missing]. Both paths satisfy the length and direction consistency checks, forming the final multi-path set {P1, P2} as input for subsequent GAT embedding and feedforward prediction. During the verification phase, the matching algorithm accurately retrieves the associated abnormal pattern paths corresponding to the environmental state, ensuring the diversity and relevance of subsequent predictions. S5.2: For each path in the set of reasoning paths, a knowledge graph embedding method based on graph attention network (GAT) is used to map the entities and relations in the path to a unified low-dimensional semantic space to obtain a vectorized representation of the path. For sub-step S5.2, the input condition is a set of multiple inference paths semantically associated with the current environment state obtained in step S5.1. Each path consists of sequentially connected entity nodes and relational edges. Entities include device nodes, environment attribute nodes, and operation behavior nodes, while relational edges include semantic associations such as operational impact, environmental effect, and behavior trigger. A Graph Attention Network (GAT) embedding method (parameters: number of attention heads h, embedding dimension d, activation function type f) is used to encode the distributed entity features and relational features in the paths, achieving weighted adjacency feature aggregation. The aggregation vector of the target node is calculated using the neighbor node feature weighted aggregation formula. This can be expressed by the following MathML formula: in, The association weight between node i and its neighbor node j is calculated based on the attention mechanism. It is a linear transformation matrix. Let j be the feature vector of node j; Furthermore, the unnormalized scores of attention weights are nonlinearly transformed using the LeakyReLU activation function to enhance the suppression of low-relevance neighbors and the salient effect of high-relevance neighbors, and Softmax normalization is used to ensure that the distribution of weight values ​​within the range of neighbor nodes is 1. Furthermore, multi-head attention fusion processing is performed on the entity features and relation features in the path (parameter: number of multi-heads h=8, fusion method is concatenation Concat), which integrates the diverse interaction patterns between entities and relations from different attention perspectives and obtains the local embedding matrix of the path; Furthermore, the multi-head attention output matrix is ​​input into the fully connected mapping layer. Through the combined action of the linear transformation matrix and the activation function f=ELU, the path embedding is mapped to a unified low-dimensional semantic space (parameter: embedding dimension d=128), generating a global vectorized representation of the path. Through this representation, the entity-relationship combination within the path is transformed into a structured semantic feature vector, providing a unified input format for subsequent temporal evolution modeling. By using the embedding method of graph attention network, the original path structure is transformed into low-dimensional vector data containing multi-view attention weights, which realizes efficient encoding of path semantic features and improves comparability, thus meeting the need for deep capture of path relationship patterns in subsequent reasoning processes. For example, in a laundry room environmental monitoring scenario, the inference path includes four entity nodes {device A, ambient temperature and humidity, ventilation status, operator} and three relational edges {device A's operation affects ambient temperature and humidity, ventilation status adjustment affects ambient temperature and humidity, operator behavior triggers device A's operating mode}. Each node has an initial feature dimension of 64. A multi-head attention mechanism is chosen, with 8 attention heads (h=8) and each attention head outputting 16 dimensions. Unnormalized weights are activated using LeakyReLU (with a negative slope parameter of 0.2), and then normalized using Softmax to obtain the weight distribution matrix between node pairs. This matrix is ​​then concatenated by the eight attention heads to form a 128-dimensional embedding vector for each node. Taking the relational edge (device A's operation affects ambient temperature and humidity) as an example, the average weight value in the multi-head attention calculation is 0.65, significantly higher than other edges, indicating that this relation is more important in the current scenario. Finally, the nodes of the path are embedded sequentially and compressed into a unified 128-dimensional semantic vector through a fully connected mapping layer, which serves as the input for the subsequent LSTM time evolution modeling in step S5.3. It has been verified that the Pearson correlation coefficient between the embedded vector and the historical anomaly prediction label reaches 0.92, which significantly improves the accuracy of the prediction model in identifying the trend of environmental anomalies. S5.3: Based on the vectorized representation of the path and historical path execution feedback data, a Long Short-Term Memory (LSTM) network is used to model the temporal evolution characteristics of the path in order to predict the evolution trend of the path in future time steps and generate the path evolution state vector. S5.4: Perform multilayer perceptron (MLP) classification on the path evolution state vector to identify the anomaly type label corresponding to the path evolution and output the prediction results of the possible anomaly states that may occur in the future under each path. For the path evolution state vectors generated by LSTM, a multilayer perceptron (MLP) classification algorithm (parameters: number of hidden layers h=3, activation function ReLU, learning rate lr=0.001, batch size batch=64) is used to realize the function of determining the abnormal type of the future evolution trend of each path; Furthermore, matrix multiplication is performed between the weight matrix of the fully connected layer and the path evolution state vector, followed by nonlinear activation processing, to obtain the first-stage high-dimensional feature mapping result, which is used to capture the nonlinear correlation pattern between path evolution features and anomaly types. Furthermore, the Dropout random deactivation mechanism (parameter: deactivation probability p=0.5) is used to perform feature sparsification on the feature mapping results of the first stage, which suppresses the risk of overfitting of the model in a small sample high-dimensional feature space and improves the generalization ability of different inference paths. Furthermore, the sparse features are compressed in dimension through a second-stage fully connected layer, and the probability distribution of each anomaly type is calculated using the Softmax normalization function. The Softmax calculation process conforms to the following formula: in, Let be the predicted probability of the i-th type of anomaly. For the linear output value of class i, Let be the linear output value of the j-th class, n be the total number of anomaly types, and e be the natural coefficient. Furthermore, a maximum value indexing operation is performed on the probability distribution vector to select the anomaly type with the highest probability as the future anomaly prediction label for the corresponding inference path, and a set of prediction results containing anomaly types and corresponding probabilities is generated. By using MLP classification and Softmax probability mapping, the path evolution state vector from the previous step is transformed into structured anomaly type prediction data, achieving accurate labeling and classification of the path for possible future anomalies. For example, in an industrial laundry environment, the path evolution state vector has a dimension of 128. The first fully connected layer of the MLP has 256 nodes, the second layer has 128 nodes, and the third layer has 64 nodes. The activation function is ReLU, and Dropout is set to 0.5. The model classifies three types of anomalies, with a Softmax output dimension of 3, corresponding to abnormal temperature and humidity, insufficient ventilation, and equipment malfunction. For a given input state vector... After passing through the first layer of weight matrix After mapping and ReLU activation, 256-dimensional features are obtained. After Dropout, the features are compressed to 128 dimensions in the second layer, and then compressed to 64 dimensions in the third layer. Softmax is then calculated, yielding anomalies in temperature and humidity (0.15), insufficient ventilation (0.75), and equipment malfunction (0.10). The most probable event is insufficient ventilation. The model outputs the label of the possible abnormal state that may occur in the future along this path as "insufficient ventilation," with a probability of 0.75. Validation results show that this processing flow achieves a Top-1 prediction accuracy of 91% on the test set, meeting the system's early warning accuracy requirements. S5.5: Based on the path weight allocation mechanism and the confidence score function of the path evolution state vector, calculate the confidence index of the abnormal state prediction result corresponding to each inference path to evaluate the reliability of the prediction result.

[0017] Step S6: Determine whether there are any abnormal states in the prediction results with a confidence level higher than a preset threshold. If so, generate a corresponding feedforward control command sequence, including ventilation adjustment, dehumidification start-up, or equipment pause operation. Specifically, this includes: S6.1: Sort the multiple abnormal state prediction results and their corresponding confidence indices output in S5, and arrange the predicted abnormal events from high to low based on the confidence level to determine the abnormal type that needs to be responded to first. S6.2: Based on a preset confidence threshold, the sorted prediction results are filtered, and the set of abnormal events with a confidence level higher than the threshold is extracted as the triggering basis for generating feedforward control instructions. Based on the abnormal state prediction results and their corresponding confidence ranking sequences output from step S6.1, a confidence threshold filtering algorithm is used (parameter: preset confidence threshold). This enables the automatic identification and extraction of high-confidence anomaly events; Furthermore, by comparing the confidence value of each prediction result with... The numerical value is used to generate a binary screening label matrix for abnormal events, where a matrix element of 1 indicates that the confidence level of the event is greater than the threshold, and a matrix element of 0 indicates that the confidence level of the event is less than or equal to the threshold, thus achieving preliminary screening of abnormal events; Furthermore, a logical index extraction method (parameters: label matrix, list of abnormal events) is adopted to extract the abnormal events marked as 1 from the sorted sequence, generate a set of high-confidence abnormal events, and retain their original prediction attribute information and association inference path identifiers to support the generation of subsequent association control instructions; Furthermore, by using a set deduplication and index rearrangement algorithm (parameters: set of abnormal events, inference path index table), duplicate items in the set of high-confidence abnormal events are deleted and their order is reconstructed, ensuring the uniqueness and priority order of events in the set within the sequence. By using the above-mentioned threshold comparison and index extraction processing method, the sorted prediction results in the previous step are transformed into a set of high-confidence abnormal events, thereby achieving accurate location of triggering conditions in the feedforward control instruction generation stage. For example, in an industrial laundry environment, the current inference output predicts anomalous events in a set of five different event types, with confidence values ​​of 0.92, 0.86, 0.78, 0.95, and 0.64, respectively, and a preset confidence threshold is set. Set the confidence level to 0.85. Using a threshold filtering algorithm, events with a confidence level greater than 0.85 are marked as 1, and a label matrix is ​​generated. Events 1 (0.92), 2 (0.86), and 4 (0.95) were extracted into a high-confidence abnormal event set using a logical indexing method. The set was then reconstructed by sorting the events by their original confidence levels as [Event 4 (0.95), Event 1 (0.92), Event 2 (0.86)]. In the subsequent control command generation stage, the corresponding response strategy templates for this set were dehumidifier startup (Event 4), ventilation adjustment (Event 1), and partial equipment suspension (Event 2), respectively. In actual verification, this screening process reduced the false detection rate of feedforward control trigger conditions to below 5%, shortened the control response delay by approximately 15%, and significantly improved the system's processing accuracy and response speed in high humidity and abnormal air quality scenarios. S6.3: Based on the selected high-confidence abnormal event types, call the predefined abnormal response strategy library in the knowledge graph to obtain the control operation template that matches the abnormal type, including ventilation adjustment parameters, dehumidification equipment start threshold or equipment pause conditions; S6.4: Based on the acquired control operation template, combined with the real-time monitoring data of the current laundry room environment and equipment status, generate a specific feedforward control instruction sequence, including equipment control actions, execution order and parameter settings; Based on the control operation template obtained from step S6.3, the real-time monitoring data of the current laundry room environment, and the equipment status parameters, a parameter matching algorithm (based on the allowable parameter range of the equipment and the recommended value range of the template) is used to determine the compatibility between the template parameters and the real-time environmental parameters, and to generate an initial set of control command parameters. Furthermore, by using a device priority scheduling algorithm (input: exception type, device execution delay, energy consumption weight), the execution order of device control actions is optimized, and a list of device control actions with optimized order is obtained. Furthermore, a parameter correction model (input: real-time environmental state vector and template setpoint) is adopted to achieve parameter fine-tuning based on deviation compensation and generate a corrected equipment control parameter matrix; Furthermore, a control instruction assembly algorithm (based on the IEC industrial control instruction encoding format) is adopted to combine the equipment control action list with the modified equipment control parameter matrix into a complete control instruction sequence, and generate an encoded instruction set that can be parsed by the control execution module; Through the above processing method, the control operation template of the previous step is transformed into a feedforward control instruction sequence containing specific equipment control actions, execution order and parameter setting values, so as to realize the targeted intervention and real-time adjustment function for predicted abnormal states. For example, in an industrial laundry where the humidity is predicted to exceed 85% and the ventilation speed to be less than 0.8 m / s within the next 5 minutes, the "Humidity Exceeds Limit + Insufficient Ventilation" template from the anomaly response strategy library is invoked. The ventilation adjustment target is 2.5 m / s, and the dehumidifier start-up humidity threshold is 80%. Real-time monitoring data shows that the current ventilation speed is only 0.6 m / s, the humidity is 82%, and the equipment status shows that both main fans are available, while one dehumidifier is in standby. The parameter matching algorithm compares the template ventilation target with the real-time ventilation speed, setting the target fan speed to 2800 rpm; the dehumidifier start-up conditions are met, setting the dehumidifier operating power to 90% of its rated value. The scheduling algorithm prioritizes fan start-up based on the fan start-up delay (3 seconds) and dehumidifier start-up delay (5 seconds), then executes dehumidifier start-up. The parameter correction model, considering that the humidity prediction curve will reach 88% within the next 3 minutes, adjusts the dehumidifier power parameter to 95% of its rated value. The control command assembly algorithm, based on the IEC 61131-3 standard, encodes the sequence "Start Fan 1 (2800 rpm) → Start Fan 2 (2800 rpm) → Start Dehumidifier (95% rated power)" into standard control messages transmittable via industrial Ethernet, generating the final feedforward control command sequence. In actual execution, this command increases the ventilation speed to 2.4 m / s within 2 minutes and reduces the humidity to 78% within 4 minutes, achieving early intervention and effective control in anticipation of abnormal conditions. S6.5: Perform a validity check on the generated control command sequence to ensure that the command parameters are within the allowable range of the device and do not conflict with the current system state. If the check passes, mark it as an executable control command and prepare to send it to the control execution module.

[0018] Step S7: The generated control commands are sent to the laundry room control execution module, and the environmental status feedback data after execution is recorded for subsequent inference path effect evaluation and optimization. Specifically, this includes: S7.1: Based on the feedforward control command sequence, initialize and configure the communication interface of the control execution module to establish a stable communication link with each execution unit in the laundry room, and ensure the real-time transmission and execution of control commands; S7.2: The Modbus-TCP protocol is used to encode control commands into standard control messages, and the commands are sent to the corresponding ventilation and conditioning devices, dehumidification equipment or equipment controllers via industrial Ethernet to ensure accurate triggering of control actions; S7.3: After the control command is issued, start the feedback data acquisition task to monitor the environmental status of the laundry room at multiple points after execution and obtain real-time feedback datasets including temperature and humidity, air flow rate, and equipment operating status. S7.4: Perform timestamp alignment and data format conversion on the collected feedback data to eliminate timing misalignment caused by communication delays and sampling frequency differences, and generate structured feedback data records; S7.5: Record structured feedback data and write it into the system evaluation database, and associate it with the issued control command identifier to provide data support for the evaluation of the subsequent inference path execution effect and the update of reinforcement learning strategy; The structured feedback data, after being timestamped and converted to the correct data format, is used as input. A unique feedback record identifier (parameters: environmental acquisition timestamp, device ID, acquisition point coordinates, sensor type code) is generated using field mapping to achieve unique storage of feedback data records in the system evaluation database. Furthermore, by constructing a hash index algorithm (parameters: SHA-256, index key = feedback record identifier), efficient feedback data retrieval is achieved, and an index association table with the control instruction identifier is obtained; Furthermore, a bidirectional mapping table generation method is adopted (parameters: association key 1 = feedback record identifier, association key 2 = control command identifier) ​​to realize the bidirectional association and binding of feedback data and control commands, and generate a data cross-reference index that can be called for path evaluation; Furthermore, based on the transactional consistency write mechanism (parameters: ACID transaction support, batch commit window = 500 records), consistent storage of feedback data and associated binding records is achieved, ensuring the integrity and no data loss of the evaluation database under high-concurrency write scenarios; Furthermore, by using a data label expansion algorithm (parameters: label set = execution result status, environmental response category, anomaly resolution duration), multi-dimensional evaluation labels are added to the feedback records to support the data requirements of multi-feature input during the subsequent reinforcement learning strategy update process; By optimizing the index and expanding the label, the structured feedback data generated in the previous step is transformed into evaluation database storage records with unique identifiers, two-way binding indexes and multi-dimensional labels, so that the evaluation of the inference path execution effect and the construction of reinforcement learning sample sets can be directly called. For example, in an industrial laundry environmental monitoring system, the ventilation control device executes a high-confidence feedforward control command (control command identifier CTRL-202403080915). After the command is executed, the system obtains a feedback dataset through multi-point temperature and humidity acquisition, including temperature (degrees Celsius) and humidity (relative humidity %) from five sensor acquisition points: the acquisition timestamp range is from 09:15:05 to 09:15:35 on March 8, 2024, with the average temperature dropping to 22.5°C and the average humidity decreasing to 48%. The system generates a structured feedback data matrix based on timestamp alignment and format conversion, and uses a field mapping algorithm to generate a unique record identifier FD-202403080915-AX01. The index key value is calculated using the SHA-256 hash algorithm, and cross-reference entries are constructed in the index association table, bidirectionally binding the feedback record with the CTRL-202403080915 command. Database transactions use a batch commit window of 500 records to ensure the stability of batch writes under high-frequency control scenarios. Finally, a set of labels was added to the feedback record, such as execution result status "completed", environmental response category "rapid humidity drop", and anomaly resolution time = 30 seconds, forming an evaluation database storage record. This record was successfully called in the subsequent construction of the S8 reinforcement learning sample set to train the path selection strategy model and improve the system's ability to optimize responses to similar humidity anomalies.

[0019] Step S8: The path selection strategy is iteratively updated using a reinforcement learning algorithm. Path weights are adjusted based on execution feedback from historical inference paths to improve the system's adaptability to complex coupled environments. Specifically, this includes: S8.1: Based on the execution feedback data of historical inference paths, construct a reinforcement learning training sample set, where each sample contains the current environment state, the selected inference path, the execution result feedback and the corresponding reward value, so as to form a mapping relationship between path selection and system response; S8.2: A deep Q-network (DQN) is used as the path selection strategy model. The model is trained on the constructed training sample set, and the neural network parameters are optimized through the backpropagation algorithm to obtain the mapping strategy from the current environment state to the optimal inference path. S8.3: Introduce an experience replay mechanism to store historical execution experience in the experience pool during model training and randomly select samples in batches for training, so as to break the correlation of data and improve the stability and generalization ability of model training. S8.4: Based on the trained path selection strategy model, calculate the Q value of each inference path under the current environment state, sort the inference paths according to the Q value, and update the path weight allocation strategy in the path diversity evaluation module to achieve dynamic path selection optimization. S8.5: Set up a periodic model update mechanism to trigger the model retraining process based on system runtime or the number of times the inference path is executed. Combine the latest environmental data and execution feedback to continuously optimize the path selection strategy, so as to improve the system's adaptability and robustness to the complex coupled environment of the laundry room.

[0020] Step S9: Periodically check the timeliness of entities and relationships in the knowledge graph. If data updates are delayed or relationships are invalid, trigger the knowledge graph incremental update mechanism to ensure the real-time performance and accuracy of the reasoning foundation. Specifically, this includes: S9.1: Based on the equipment status, environmental parameters and operational behavior data collected from the industrial laundry environmental monitoring system, a knowledge graph entity update detection module is constructed to identify the update delay or abnormal drift of entity attribute values. S9.2: Perform semantic consistency verification on the semantic triple relationships between entities in the knowledge graph. Use graph embedding algorithm to calculate the similarity between the current relationship vector and the historical relationship vector. If the similarity is lower than the set threshold, mark it as a potentially invalid relationship. S9.3: Based on the timestamp mechanism, the last update time of each entity and relation in the knowledge graph is recorded and compared with the current system time. If the time difference exceeds the preset update cycle threshold, it is determined that the data update is delayed. S9.4: Perform incremental update operations on the detected lagging entities and failed relationships. Use the incremental update interface of the graph database to fuse the newly collected environmental monitoring data with the historical knowledge graph to generate an updated set of semantic triples. S9.5: Based on the updated set of semantic triples, perform topological optimization of the knowledge graph and use graph compression algorithm to merge and simplify redundant paths to improve the efficiency and accuracy of path selection in subsequent reasoning processes; S9.6: Submit the optimized knowledge graph version to the inference engine module and generate a version update log, including the update time, the list of updated entities / relationships, and the update method, for subsequent inference path evaluation and system auditing.

[0021] Step S10: When a sudden change in environmental state is detected or the control command fails to achieve the expected effect, switch to the backup inference path and generate an emergency intervention command to improve the system's robust response capability to abnormal scenarios. Specifically, this includes: S10.1: Real-time comparison of current laundry room environmental monitoring data with historical control command execution feedback to identify whether the control command execution deviation exceeds the preset threshold, in order to determine whether there is a risk of control failure; S10.2: Based on the environmental state mutation detection algorithm, perform mutation analysis on temperature, humidity, ventilation status and equipment operating parameters. If a mutation signal exceeding the normal fluctuation range is detected, it is marked as an abnormal evolution acceleration event. S10.3: Based on the set of alternative reasoning paths in the knowledge graph, call the path switching evaluation model, and generate a path switching priority ranking list by combining the current environment state and the execution effect of historical paths; S10.4: Based on the selected backup reasoning path, perform emergency reasoning operations on the current environmental state, generate multiple sets of candidate intervention instruction sequences, and perform feasibility screening based on system safety constraints to obtain the optimal emergency intervention plan; S10.5: The generated emergency intervention command is sent to the laundry room control execution module, and the environmental status and system response data before and after the intervention are recorded for subsequent path evaluation and reinforcement learning model training.

[0022] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0023] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring and controlling the environment of an intelligent laundry room, characterized in that, Includes the following steps: S1: Collect operating status parameters, temperature and humidity, ventilation status and operation behavior data of multiple devices in the laundry room environment, and record the spatial location label and equipment type identifier of each data collection point; S2: Normalize the collected raw data and remove outliers to generate a standardized environmental monitoring dataset; S3: Based on the standardized environmental monitoring dataset and historical abnormal event records, construct a multi-dimensional knowledge graph that includes equipment entities, environmental attributes, operational behaviors, and abnormal patterns, where the relationships between entities are expressed through semantic triples; S4: Introduce a path diversity evaluation mechanism to dynamically assign weights to multiple potential reasoning paths in the multi-dimensional knowledge graph and calculate the path selection priority based on the current environmental state and the historical path execution probability. S5: Based on the current environmental state and the selected inference path, perform multi-path feedforward inference to generate prediction results of possible abnormal states in the future, and calculate the confidence index of each prediction result. S6: Determine whether there is an abnormal state in the prediction result with a confidence level higher than a preset threshold. If so, generate a corresponding feedforward control command sequence, including ventilation adjustment, dehumidification start-up, or equipment pause operation. S7: Send the generated control commands to the laundry room control execution module and record the environmental status feedback data after execution.

2. The intelligent laundry room environmental monitoring and control method according to claim 1, characterized in that, Following step S7, the following is also included: S8: The path selection strategy is iteratively updated using a reinforcement learning algorithm, and the path weights are adjusted based on the execution feedback of historical inference paths; S9: Periodically check the timeliness of entities and relationships in the knowledge graph. If data updates are found to be lagging or relationships are invalid, trigger the knowledge graph incremental update mechanism. S10: When a sudden change in environmental state is detected or the control command does not achieve the expected effect, switch to the backup inference path and generate an emergency intervention command.

3. The intelligent laundry room environmental monitoring and control method according to claim 1, characterized in that, Step S1 specifically includes: Acquire output signals from multiple types of sensors deployed within the laundry room area, and collect multi-source real-time status data of the laundry room environment; The output signals of the equipment operation status sensors are analyzed to extract equipment operation status parameters and form equipment operation status feature vectors. The output data of the temperature and humidity sensor is read and processed by unit conversion, and the raw digital signal is converted into Celsius and relative humidity percentage values ​​to obtain standardized temperature and humidity environmental parameters. The analog voltage signal output by the air quality sensor is filtered and converted into PM2.5 concentration and CO2 concentration values ​​through a preset linear fitting equation to obtain air quality indicators related to ventilation status. Receive status reports from infrared human presence sensors, identify operation behavior signals indicating whether an operator is present, and generate operation behavior timing data. Spatial location tags and device type identifiers are attached to the collected sensor data. The spatial location tags are obtained based on a preset coordinate mapping table, and the device type identifiers are obtained by mapping the device model and communication address.

4. The intelligent laundry room environmental monitoring and control method according to claim 3, characterized in that, In step S1, the acquisition of output signals from multiple types of sensors adopts a multi-channel asynchronous acquisition strategy. The sampling frequency of different sensors is set in the range of 1Hz to 10Hz, and the data is acquired in real time and stably through automatic identification of sensor protocols, driver loading, and CRC16 verification.

5. The intelligent laundry room environmental monitoring and control method according to claim 1, characterized in that, Step S2 specifically includes: The collected operating status parameters, temperature and humidity data, ventilation status information and operation behavior records of the laundry room equipment are processed to unify the format, and the raw data output by different sensors are converted into a unified timestamp and structured field format to obtain a structured raw dataset. The temperature and humidity sensor data in the structured original dataset are denoised using a sliding window mean filtering algorithm to obtain smoothed temperature and humidity time series data. The smoothed temperature and humidity time series data, the equipment operating parameters, and the ventilation status information are normalized using a standardization method to obtain a standardized numerical dataset within a unified dimension. The isolated forest anomaly detection algorithm is applied to identify outliers in each dimension of the standardized numerical dataset, generating an outlier label matrix. Based on the outlier labeling matrix, outlier removal and linear interpolation compensation operations are performed on the standardized numerical dataset to generate a standardized environmental monitoring dataset.

6. The intelligent laundry room environmental monitoring and control method according to claim 1, characterized in that, Step S3 specifically includes: Entity recognition processing is performed on the equipment operating status parameters, temperature and humidity, ventilation status and operation behavior data in the standardized environmental monitoring dataset to extract basic entity nodes of laundry equipment, environmental attributes and operator behavior. Based on the timestamp, device identifier, anomaly type and handling measures fields in the historical anomaly event records, the anomaly events are classified, labeled and semantically abstracted to generate anomaly pattern entity nodes. A semantic role labeling algorithm is used to extract semantic relationships between entities, and a set of multi-class relationship triples is generated based on entity co-occurrence frequency and context semantic association strength. The triple set is subjected to knowledge fusion processing to form a multidimensional knowledge graph topology structure under a unified namespace; Based on graph databases, persistent storage operations of knowledge graphs are performed, and the indexes of entity nodes and relation edges in the graphs are optimized and their attributes are expanded.

7. The intelligent laundry room environmental monitoring and control method according to claim 6, characterized in that, Step S3 further includes using semantic role labeling and point mutual information (PMI) algorithm to rank entity relationship strength, selecting the top 10% of entity pairs with semantic association to generate device-environment, environment-operation, and operation-abnormal triplet, and using point embedding model and complex embedding model to disambiguate synonymous entities and redundant relationships.

8. The intelligent laundry room environmental monitoring and control method according to claim 1, characterized in that, Step S4 specifically includes: A topological analysis of the entity relationship network in the knowledge graph is performed, and a graph traversal algorithm is used to identify the set of potential reasoning paths associated with the current environment state, thereby constructing a candidate reasoning path library. Based on the execution feedback data of historical reasoning paths, Bayesian statistical methods are used to calculate the path execution probability of each path under similar environmental conditions. The current environment state feature vector is matched with the historical path context features, and the cosine similarity algorithm is used to calculate the path fit index. Based on the path execution probability and the path suitability index, a weighted summation algorithm is used to generate dynamic weight values ​​for each candidate path. The dynamic weight values ​​are normalized, and a path sampling operation is performed based on the probability roulette selection strategy to generate a priority sequence for multi-path inference.

9. The intelligent laundry room environmental monitoring and control method according to claim 1, characterized in that, Step S5 specifically includes: Based on the current state vector of the laundry room environment, perform multi-path matching operation in the knowledge graph to obtain a set of multiple reasoning paths semantically associated with the current state; For each path in the set of reasoning paths, a knowledge graph embedding method based on graph attention network is used to map the entities and relations in the path to a unified low-dimensional semantic space to obtain a vectorized representation of the path. Based on the vectorized representation of the path and historical path execution feedback data, a long short-term memory network is used to model the temporal evolution characteristics of the path, predict the evolution trend of the path in future time steps, and generate a path evolution state vector. The path evolution state vector is subjected to multilayer perceptron classification processing to identify the abnormal type label corresponding to the path evolution, and output the prediction results of possible abnormal states in future time under each path. Based on the path weight allocation mechanism and the confidence score function of the path evolution state vector, the confidence index of the abnormal state prediction result corresponding to each inference path is calculated.

10. The intelligent laundry room environmental monitoring and control method according to claim 2, characterized in that, In step S8, a reinforcement learning training sample set is constructed based on the execution feedback data of historical inference paths. Each sample includes the environment state, inference path, feedback result and corresponding reward value. The path selection strategy is based on a deep Q network and is continuously optimized by combining experience playback and periodic retraining.

Images