Shared office anonymous behavior identification and short-time occupation prediction method and system, and medium

Abstract

The application provides a shared office anonymous behavior recognition and short-time occupation prediction method. Relying on zone-level visual, carbon dioxide, motion, power and other multi-type anonymous sensor collection data, following a three-level privacy protocol, the original image is destroyed at the edge, and only aggregated anonymous information is output. The system detects three types of occupation events: number of people, in-place rest, and extracts six-dimensional cause features to construct anonymous evidence packages. A two-level architecture of LLM parser and review unit is used to infer the posterior probability of causes such as toilet and lunch break, and the bottom is verified by mode, sensor and confidence. Based on the pre-calibrated truncated normal return distribution, the personnel return probability and expected empty time are calculated, and the air conditioning control flags such as pre-cooling, maintenance and rollback are generated differently. When the confidence is low, the conservative strategy is executed. This scheme can distinguish different off-site behaviors to dynamically control VRV air conditioners, solve the problem that traditional schemes only coarsely judge the presence, are prone to false shutdown, and the energy-saving comfort is unbalanced, and at the same time provide an offline Bayesian alternative reasoning scheme to adapt to offline scenes.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensing and control technology for building environments, and particularly to a method, system, and medium for identifying anonymous behavior and predicting short-term occupancy in shared office spaces. Specifically, it relates to a method, system, and medium for identifying the causes of anonymous behavior and predicting short-term occupancy in shared office spaces for multiple users. Background Technology

[0002] With the development of the concept of occupancy-centric control, occupancy information is gradually becoming the core input for the coordinated optimization of HVAC, ventilation and thermal comfort in commercial buildings. However, the output of existing occupancy sensing and prediction technologies is still mainly based on whether there is a person (presence), the number of people (count), or coarse-grained occupancy level (low / medium / high), which are cause-blind representations.

[0003] In shared office spaces, short-term occupancy changes are often caused by different underlying behaviors, which vary significantly in duration and return probability. Toilet breaks typically last about 3 to 8 minutes with a high return probability; smoking typically lasts about 8 to 15 minutes and recurs periodically according to individual habits; getting water or taking a tea break typically lasts about 2 to 5 minutes; temporary absences typically last about 15 to 40 minutes; and lunch breaks or leaving get off work have a very low probability of returning in the near future. Furthermore, short naps (where the person remains at their workstation but activity levels significantly decrease) are a special case. Systems based on passive infrared (PIR) motion detection may misjudge vacancy and incorrectly shut off the air conditioning due to prolonged inactivity.

[0004] Existing methods for directly identifying individual behavior (facial recognition, pedestrian re-identification, wearable devices or work card positioning, etc.) are neither compliant nor easy to implement in real offices where people are mixed and privacy is sensitive. Large Language Models (LLM) have shown advantages in context integration and few-sample inference, but there is currently no publicly available solution for using them to infer the cause of short-term displacement and generate a reliable air conditioning control interface without identifying individuals.

[0005] In summary, existing technologies suffer from several drawbacks: the occupancy perception technology is unrelated to the cause of occupancy, individual identification schemes are not feasible in privacy-sensitive scenarios, and they cannot cover the state of resting while in place. Therefore, how to identify the cause of short-term occupancy without identifying individual identity and thereby improve short-term prediction and control is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, this invention provides a method for anonymous behavior identification and short-term occupancy prediction in shared office spaces. This method enables the inference of the causes and return times of short-term occupancy changes without identifying individual identities, thereby supporting shared air conditioning systems in making risk-aware decisions to maintain, revert, or pre-cool.

[0007] To achieve the above objectives, the shared office anonymous behavior identification and short-term occupancy prediction method provided by the present invention includes the following steps:

[0008] S1. Collect multimodal raw observation data of the office space through the edge-side anonymous sensing architecture. The edge side does not output any images, pixels, or biometric data that can identify individuals or personal identities, but only outputs anonymous statistics of partition aggregation.

[0009] S2. Detect short-term occupancy change events in office space based on the aforementioned multimodal raw observation data;

[0010] S3. For each short-term occupancy change event, extract the event cause-related features without distinguishing or identifying any individual's identity, and encapsulate the features and the corresponding time window observation data into a structured anonymized evidence package.

[0011] S4. Input the anonymous evidence into the cause identification and inference unit, and infer the posterior probability distribution corresponding to the cause of each predefined person's departure behavior.

[0012] S5. Retrieve the return time distribution function pre-calibrated for each cause, and combine it with the posterior probability distribution of the cause to calculate the probability of personnel returning and the expected idle time of the partition under the specified duration.

[0013] S6. Generate a dedicated control flag for the shared air conditioning system by combining three types of indicators: return probability, expected idle time, and causal reasoning confidence. When the reasoning confidence is lower than the preset threshold or there is evidence conflict in the multimodal observation data, the abandonment judgment logic is triggered, and the control flag reverts to the conservative baseline of maintaining the operating state of the previous device.

[0014] In some embodiments, the anonymous sensing architecture in step S1 includes multiple types of dedicated anonymous sensing units: a district-level anonymous visual counting unit, an inlet direction event unit, a carbon dioxide and temperature and humidity unit, a PIR motion detection unit, a device power unit, and an air supply status unit.

[0015] The data flows unidirectionally according to a three-level privacy protocol: the original P0 image is discarded directly at the sensing edge and is not stored or uploaded to the cloud; the P1 layer contains aggregated data on the number of people in each partition, entry and exit counts, and activity candidate tags circulating within the local area network; only the P2 layer's anonymous evidence package in JSON format is input to the cause identification and inference unit.

[0016] In some embodiments, the entrance direction event unit uses a double infrared beam with staggered installation for soft curtain-type entrances that cannot stably trigger the door magnet. It determines the entrance and exit direction based on the order in which the two infrared beams are blocked and outputs a structured event containing a timestamp and the entrance and exit direction.

[0017] In some embodiments, the short-term occupancy change events in step S2 are divided into three categories: departure events with a decrease in the number of people in the partition, arrival events with an increase in the number of people in the partition, and resting events with no change in the number of people but a decrease in activity.

[0018] The rules for determining the "resting in place" event are as follows: if the number of people in a zone is always greater than zero within a set time window, the PIR unit trigger frequency decreases more than the standard relative to the 30-minute baseline, and there is no corresponding decrease in carbon dioxide concentration, it is determined that the person has not left the site, and the air conditioning shutdown action is prohibited.

[0019] In some embodiments, the causal features extracted in step S3 are six-dimensional anonymous causal feature vectors, which are, in order: event scale features, exit direction features, time period prior features, event recurrence features, equipment power coupling features, and environmental attenuation features.

[0020] In some embodiments, each dimension of the feature has a specific causal differentiation function:

[0021] The event scale feature is used to distinguish between a single person leaving their post for a short period of time and a group of people leaving a meeting / workday.

[0022] The exit direction features are distinguished by the type of door used for restrooms, smoking, going out, meetings, and lunch break departures.

[0023] The time-period prior features introduce the prior probability of behavior occurring at each time period throughout the day;

[0024] The event reproducibility features are used to identify periodic smoking disengagement behavior;

[0025] The power coupling characteristics of the equipment are adjusted based on the short-term return probability of the on / off state of the workstation equipment.

[0026] The environmental degradation characteristic is corrected for the actual number of people present in the zone by the rate of carbon dioxide decline.

[0027] In some embodiments, the cause identification and inference unit in step S4 is configured with a two-level verification and inference architecture, consisting of an LLM parser and an external review unit;

[0028] The LLM parser receives P2-level anonymous evidence packets and, under the dual constraints of a predefined set of causes and a fixed output JSON format, outputs structured data containing the posterior probabilities of all causes, the highest confidence cause label, and the reasoning basis. The external review unit independently verifies the output results without changing the underlying parameters of the large model and is compatible with various LLM parser backends.

[0029] In some embodiments, the additional review unit performs three independent checks in parallel:

[0030] Pattern validity check: Verify that the output JSON fields are complete and that the cause enumeration values ​​are within the preset set;

[0031] Sensor consistency check: LLM identifies conflicting evidence when a sensor is displaced but CO2 concentration is rising positively.

[0032] Confidence threshold check: If the maximum causal posterior probability is lower than the CONF_TH threshold, the decision is abandoned; if any check fails, the conservative baseline control flag is directly enabled.

[0033] In some embodiments, the cause identification and inference unit is set to an offline alternative embodiment: a Bayesian scoring framework; the posterior probability of the cause is proportional to the product of the prior weight of the time period and the likelihood factor of each feature, all causes are normalized to obtain the probability distribution, and the likelihood factor is calibrated by the site pre-experiment, without the need to call the large model network interface.

[0034] In some embodiments, the predefined potential cause categories include at least two or more categories such as toilet use, water / tea break, smoking, short rest, temporary absence, meeting gatherings and dispersals, and lunch break / off of get off work. For each cause category, the time for people to leave and return is fitted using a truncated normal distribution, and three sets of distribution parameters, namely mean, standard deviation, and upper and lower limits of time truncation, are pre-tested and calibrated.

[0035] In some embodiments, the return probability calculation formula in step S5 is as follows:

[0036] p_return(τ)=ΣP(C_k|features)×F_k(τ),

[0037] In the formula, P(C_k|features) is the posterior probability of cause Ck, and F_k is the truncated normal return time cumulative distribution of the k-th cause (parameters are calibrated by pre-experiment).

[0038] This invention provides a shared anonymity behavior identification and short-term occupancy prediction system for performing the prediction method described above, the prediction system comprising:

[0039] Anonymous sensing units are used to collect multimodal anonymous sensing data in the office space and perform edge privacy data destruction;

[0040] The event detection unit is used to identify three types of short-term occupancy change events: increase or decrease in the number of people in a zone, and short rest in place.

[0041] The causal feature extraction and evidence package construction unit is used to extract six-dimensional causal features and generate a P2-level anonymous JSON evidence package;

[0042] The causation identification and inference unit includes an LLM parser and a review unit or an offline Bayesian scoring module, which outputs the posterior probability distribution of each potential causation.

[0043] The occupancy duration prediction unit calculates the return probability and expected idle duration for a specified duration based on the calibrated return distribution.

[0044] The air conditioning control interface unit outputs a control flag to the shared VRV air conditioning system based on the prediction results, and triggers conservative fallback control when an anomaly is detected.

[0045] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the shared anonymity behavior identification and short-term occupancy prediction method as described above.

[0046] Compared with the prior art, the present invention has the following advantages:

[0047] The shared office anonymous behavior identification and short-term occupancy prediction method provided by this invention is particularly applicable to the occupant-center control (OCC) front end of a shared variable refrigerant flow (VRV / VRF) air conditioning system. The technical objective is to shift from detecting the specific behavior of a specific individual to constructing a structured anonymous evidence package from anonymized multimodal contextual evidence without identifying the individual's identity. This package is then processed by a cause identification and inference unit with a large language model at its core and an additional review unit to infer the posterior probability distribution and return time distribution of short-term occupancy change events on predefined potential cause categories. This probability distribution is then used to drive the reliable control decision of the shared air conditioning system. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall process of the shared office anonymous behavior identification and short-term occupancy prediction method shown in the embodiment of the present invention;

[0049] Figure 2 This is an architecture diagram of the shared office anonymous behavior identification and short-term occupancy prediction system shown in an embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram of the composition and three-level privacy data flow of the shared office anonymous behavior identification and short-term occupancy prediction system shown in an embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram illustrating the divergence relationship among the three sensors in in-situ rest identification according to an embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram of the overall process framework of the shared office anonymous behavior identification and short-term occupancy prediction method shown in the embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram of the LLM parser and review unit logic of the cause identification and inference unit shown in the embodiment of the present invention;

[0054] Figure 7 This is a schematic diagram of the decision logic for generating control-related flags based on the return probability and the expected idle time, as shown in an embodiment of the present invention.

[0055] Figure 8 This is a statistical chart showing the distribution of control flags according to an embodiment of the present invention;

[0056] Figure 9 This is a heat map showing the distribution of control flags by date and hour in an embodiment of the present invention;

[0057] Figure 10 This is a statistical chart showing the distribution of the inferred causes of off-site events according to an embodiment of the present invention;

[0058] In the attached figures, the following labels are used:

[0059] 300-Shared Anonymous Behavior Identification and Short-Term Occupancy Prediction System;

[0060] 310 - Anonymous Sensing Unit;

[0061] 320 - Event Detection Unit;

[0062] 330 - Causal Feature Extraction and Evidence Package Construction Unit;

[0063] 340 - Causal Identification and Inference Unit;

[0064] 350 - Occupancy Duration Prediction Unit;

[0065] 360-Air Conditioning Control Interface Unit;

[0066] S1-S6: Steps. Detailed Implementation

[0067] See Figure 1-10 One embodiment of the present invention provides a method for identifying anonymous behavior and predicting short-term occupancy in shared office spaces. See [link to relevant documentation]. Figure 1 It includes the following steps:

[0068] S1. Collect multimodal raw observation data of the office space through the edge-side anonymous sensing architecture. The edge side does not output any images, pixels, or biometric data that can identify individuals or personal identities, but only outputs anonymous statistics of partition aggregation.

[0069] S2. Detect short-term occupancy change events in office space based on the aforementioned multimodal raw observation data;

[0070] S3. For each short-term occupancy change event, extract the event cause-related features without distinguishing or identifying any individual's identity, and encapsulate the features and the corresponding time window observation data into a structured anonymized evidence package.

[0071] S4. Input the anonymous evidence into the cause identification and inference unit, and infer the posterior probability distribution corresponding to the cause of each predefined person's departure behavior.

[0072] S5. Retrieve the return time distribution function pre-calibrated for each cause, and combine it with the posterior probability distribution of the cause to calculate the probability of personnel returning and the expected idle time of the partition under the specified duration.

[0073] S6. Generate a dedicated control flag for the shared air conditioning system by combining three types of indicators: return probability, expected idle time, and causal reasoning confidence. When the reasoning confidence is lower than the preset threshold or there is evidence conflict in the multimodal observation data, the abandonment judgment logic is triggered, and the control flag reverts to the conservative baseline of maintaining the operating state of the previous device.

[0074] In this embodiment, the anonymized sensing architecture in step S1 includes multiple types of dedicated anonymized sensing units: a district-level anonymized visual counting unit, an entrance direction event unit, a carbon dioxide and temperature / humidity unit, a PIR motion detection unit, a device power unit, and an air supply status unit. Data flows unidirectionally according to a three-level privacy protocol: the P0 raw image is discarded directly at the sensing edge, without storage or cloud uploading; the P1 layer contains aggregated data on the number of people in each partition, entry / exit counts, and activity candidate tags circulating within the local area network; only the P2 layer's JSON-formatted anonymized evidence package is input to the cause identification and inference unit. More specifically, the P0-level raw image is discarded directly after edge processing, never leaving the acquisition device, without being saved, and without cloud uploading; the P1 layer contains the anonymized number of people in each partition, entry / exit counts, and activity candidate tags, which remain within the edge computing unit of the local area network; the P2 layer contains the anonymized evidence package in plain text or JSON format, which is the only visible input to the large language model parser.

[0075] The entrance direction event unit employs a staggered dual infrared beam system for soft curtain-type entrances where door magnets cannot be stably triggered. It determines the entry / exit direction based on the order in which the two infrared beams are blocked, outputting a structured event containing a timestamp and the entry / exit direction. The district-level anonymous visual counting unit only outputs the number of people in each district and activity candidate labels at the edge, without outputting any images, pixels, or biometric features that can identify an individual.

[0076] In this embodiment, the short-term occupancy change events in step S2 are divided into three categories: departure events (reduction in the number of people in a zone), arrival events (increase in the number of people in a zone), and rest events (unchanged number of people but decreased activity). The rules for determining rest events are as follows: if the number of people in a zone remains constant above zero within a set time window, the PIR unit trigger frequency decreases significantly relative to the 30-minute baseline, and there is no corresponding decrease in carbon dioxide concentration, it is determined that the person has not left, and the air conditioning shutdown action is prohibited. Specifically, when the number of people in a zone remains constant and greater than zero within a set time window, the unit time trigger frequency of the motion detection unit decreases by more than a set proportion compared to the previous 30-minute baseline, and the carbon dioxide concentration does not show a corresponding decrease, it is determined that the zone is in a low-activity state rather than an unoccupied state, and the air conditioning shutdown action is prohibited accordingly. In this embodiment, short-term occupancy change events are detected by the multimodal data. These short-term occupancy change events include events with a decrease in the number of people in a zone, events with an increase in the number of people in a zone, and events with decreased activity under the condition that the number of people in a zone remains constant.

[0077] In this embodiment, the causal features extracted in step S3 are a six-dimensional anonymous causal feature vector, namely: event scale feature, exit direction feature, time period prior feature, event recurrence feature, equipment power coupling feature, and environmental degradation feature. Each dimension of the feature has a specific causal differentiation function: the event scale feature is used to distinguish between short-term single-person departure and batch meeting / get off work-off departure; the exit direction feature distinguishes between toilet, smoking, outing, meeting, and lunch break departure channels based on the type of door; the time period prior feature introduces the prior probability of behavior occurring at various times of the day; the event recurrence feature is used to identify periodic smoking departure behavior; the equipment power coupling feature corrects the short-term return probability based on the on / off status of the workstation equipment; and the environmental degradation feature corrects the actual number of people present in each zone based on the carbon dioxide decline rate.

[0078] More specifically, the causal feature vector includes at least six dimensions: event scale features representing the synchronous change in the number of people within the same time window, exit direction features representing where people enter and exit, time period prior features representing the probability of different causes occurring within a day, event recurrence features representing repeated entry and exit at the same exit within a set time window, equipment power coupling features representing whether the equipment in the vacated workstation area is still in operation, and environmental decay features representing the rate of change in carbon dioxide concentration.

[0079] In this embodiment, the distinction between smoking and toilet-related causes is determined jointly based on the exit direction characteristics, event recurrence characteristics, and prior time characteristics: when leaving via an exit leading to the outside, balcony, or stairs and repeatedly entering and leaving the building within a set time window, the posterior probability of smoking is increased; when leaving via an exit leading to the internal hygiene area, occurring alone, and lasting for less than a set threshold, the posterior probability of toilet-related causes is increased. The event scale characteristic is used to distinguish between short-term individual departures and mass departures: when the decrease in the number of people in a zone does not exceed a set value, it is determined to be a short-term individual departure; when the simultaneous decrease in the number of people in a zone reaches or exceeds a set value, it is determined to be a mass departure event such as meeting dispersal or lunch break departure. The equipment power coupling characteristic is used to correct the return probability: when the critical equipment in the vacated work area remains operational after the event occurs, the near-term return probability of the event is increased; when the critical equipment is turned off, the near-term return probability is reduced and the posterior probability of lunch break / departure is increased.

[0080] In this embodiment, the cause identification and inference unit in step S4 is configured with a two-level verification and inference architecture, consisting of an LLM parser and an external review unit. The LLM parser receives P2-level anonymous evidence packets and, under the dual constraints of a predefined set of causes and a fixed output JSON format, outputs structured data containing the posterior probabilities of all causes, the highest confidence cause label, and the reasoning basis. The external review unit independently verifies the output results without modifying the underlying parameters of the large model and is compatible with various LLM parser backends.

[0081] The external review unit performs three independent checks in parallel: pattern validity check: verifying the completeness of the output JSON fields and that the cause enumeration values ​​are within a preset set; sensor consistency check: marking evidence conflict when the LLM determines that the location is off but the CO2 concentration is rising positively; confidence threshold check: marking abandonment of discrimination if the maximum causal posterior probability is lower than the CONF_TH threshold; if any check fails, the conservative baseline control flag is directly activated.

[0082] In this embodiment, the causal feature vector and the multimodal observation data of the corresponding window are organized into an anonymous evidence package, which is input into the causal identification and inference unit. The anonymous evidence package is then mapped to the posterior probability distribution of the short-term occupancy change event on a predefined set of potential causal categories.

[0083] Specifically, in this embodiment, the cause identification and inference unit adopts a two-level architecture of a large language model parser and an additional review unit as the preferred implementation: the large language model parser receives the anonymous evidence package, utilizes the contextual understanding and few-sample transfer capabilities of the large language model, and outputs structured JSON containing the posterior probability distribution of causes, the highest confidence cause label, and the inference basis under system prompts that constrain the set of cause categories and the output format; the additional review unit performs mode legality checks, sensor consistency verification, and confidence threshold reviews on the structured JSON without modifying the parameters of the large language model, and is universal for any large language model backend. The additional review unit performs the following checks: (a) Pattern validity check: verify whether the output JSON contains all required fields and whether the enumerated values ​​belong to the pre-declared cause set; (b) Sensor consistency check: when the large language model determines that the person is off-site but the slope of the carbon dioxide concentration exceeds the set positive threshold and it is a single person off-site, it is marked as a sensor conflict; (c) Confidence threshold check: when the maximum posterior probability is lower than the set confidence threshold, the discrimination flag is abandoned; if any check fails, the relevant flags are controlled to fall back to the conservative baseline that maintains the state of the previous moment.

[0084] In this embodiment, the cause identification and inference unit is set to an offline alternative implementation: a Bayesian scoring framework; the posterior probability of the cause is proportional to the product of the prior weight of the time period and the likelihood factor of each feature, and all causes are normalized to obtain the probability distribution. The likelihood factor is calibrated by the site pre-experiment, without the need to call the large model network interface.

[0085] Specifically, the cause identification and inference unit adopts a Bayesian scoring framework as an alternative embodiment: the posterior probability of a cause is proportional to the product of the prior weight of the cause in time period and the likelihood factor under each cause feature. After normalizing all causes, the posterior probability distribution is obtained. The likelihood factor is calibrated by on-site pre-experimentation. This alternative embodiment does not call the large language model interface and is suitable for local deployment scenarios without network connection. The remaining steps are completely consistent with the preferred embodiment above.

[0086] In this embodiment, the predefined potential cause categories include at least two or more: restroom use, water / tea break, smoking, short rest at work, temporary absence, meeting gatherings / gatherings, and lunch break / off of get off work. For each cause category, a truncated normal distribution is used to fit the return time of personnel leaving their positions. Preliminary experiments are conducted to calibrate three sets of distribution parameters: mean, standard deviation, and upper and lower limits of the time truncation. The return probability calculation formula in step S5 is as follows:

[0087] p_return(τ)=ΣP(C_k|features)×F_k(τ),

[0088] In the formula, P(C_k|features) is the posterior probability of cause Ck, and F_k is the truncated normal return time cumulative distribution of the k-th cause (parameters are calibrated by pre-experiment).

[0089] In this embodiment, the return probability within a given time range τ is equal to the sum of the products of the posterior probability of each cause category and the truncated normal cumulative distribution of the return of that cause within τ; the expected idle time is obtained by weighted summation of the posterior probability of each cause category with its expected return time, and an upper limit is taken to avoid abnormal estimates.

[0090] In this embodiment, the return probability and expected idle time within a given time range are calculated using the posterior probability distribution, based on the pre-calibrated return time distribution for each potential cause category. The return time distribution parameters, time period prior weights, and typical exit directions for each cause category are obtained through on-site pre-experiment calibration before formal operation, so that the method can be adapted to the behavioral patterns and architectural layout of specific office spaces.

[0091] In step S6 of this embodiment, a control-related flag for the shared air conditioning system is generated based on the return probability, expected idle time and confidence level. When the confidence level is lower than a set threshold or when multimodal evidence conflicts with each other, an abandonment judgment mechanism is triggered to control the relevant flag to fall back to a conservative baseline that maintains the state of the previous moment.

[0092] See Figure 2 Another embodiment of the present invention provides a shared anonymous behavior identification and short-term occupancy prediction system 300, used to execute the prediction method described above. The prediction system includes: an anonymous sensing unit 310, used to collect multimodal anonymous sensing data of office space and perform edge privacy data destruction; an event detection unit 320, used to identify three types of short-term occupancy change events: increase or decrease in the number of people in a partition, and short rest in place; a causal feature extraction and evidence package construction unit 330, used to extract six-dimensional causal features and generate a P2-level anonymous JSON evidence package; a causal identification and inference unit 340, including an LLM parser and a review unit or an offline Bayesian scoring module, outputting the posterior probability distribution of each potential cause; an occupancy duration prediction unit 350, which calculates the return probability of a specified duration and the expected vacancy duration based on the calibrated return distribution; and an air conditioning control interface unit 360, which outputs a control flag to the shared VRV air conditioning system according to the prediction results, and triggers conservative fallback control when the confidence is abnormal.

[0093] Specifically, the anonymized sensing unit is used to collect multimodal data, and the district-level anonymized visual counting unit only outputs the number of anonymous partitions and activity candidate labels at the edge; the event detection unit is used to perform step S2 as described in claim 1; the causal feature extraction and evidence package construction unit and the causal identification and inference unit are used to perform steps S3 and S4 as described in claim 1, with a large language model parser and an external review unit as the preferred implementation; the occupancy time prediction unit is used to perform step S5 as described in claim 1; and the air conditioning control interface unit is used to perform step S6 as described in claim 1.

[0094] The following describes in detail the shared anonymity behavior identification and short-term occupancy prediction system and method provided by the embodiments of the present invention, with reference to the accompanying drawings.

[0095] 1. System Composition and Anonymized Data Flow

[0096] In step S1 of this embodiment, the system is deployed in a shared office divided into workstation areas and meeting / discussion areas, and the controlled object is the shared VRV air conditioning system. Figure 3 As shown in the system composition and three-level data privacy flow diagram, the system includes: a zone-level anonymized visual counting unit (ceiling-mounted top view, edge side only outputs the number of people in each zone), an entrance direction event unit (hard door magnetic / soft curtain dual infrared beam), a CO2 and temperature / humidity unit, a PIR motion detection unit, a device power unit, and an air supply status unit. Data flows unidirectionally according to the three-level privacy protocol: P0 original image is discarded at the edge; P1 anonymized aggregated data remains on the local area network; P2 level JSON evidence packet is the sole input to the inference unit.

[0097] 2. Detection of short-term occupancy change events

[0098] In step S2 of this embodiment, refer to Figure 4 The processing unit uses a 5-minute analysis window to detect three types of events: ① departure events with a difference in the number of people Δn≤-1; ② arrival events with a difference of Δn≥1; ③ in-place rest events—the number of people remains unchanged and is greater than zero, the PIR trigger frequency drops sharply by more than a set proportion compared to the 30-minute baseline, and CO2 does not show a corresponding decay, which are judged as in-place low activity states, and shutdown is prohibited based on this.

[0099] 3. Characteristics and Evidence of Anonymity

[0100] In step S3 of this embodiment, six-dimensional anonymized features are extracted for each departure event: event size (single / batch, etc.), exit direction proxy (conference room door event, M11 building departure count), equipment power coupling (whether the socket power exceeds the threshold), CO2 environmental decay slope, reproducibility of departure events (number of times within the first 90 minutes), and time period (hours). These features, along with window quantities such as CO2 mean and PIR duty cycle, construct a P2-level JSON evidence package. The predefined potential cause categories and their typical characteristics in this embodiment are shown in Table 1. It should be emphasized that these categories are event-level potential causes, and the system does not attribute them to any specific individual.

[0101] Table 1. Potential Causes and Typical Characteristics of Short-Term Occupancy Change Events

[0102]

[0103] Table 2 shows the anonymous causal feature vectors extracted for each event and their effects.

[0104] Table 2. Composition and Function of Anonymity-Causing Feature Vectors

[0105]

[0106] 4. LLM parser + inspection unit

[0107] See Figure 5 , Figure 6 As shown, the preferred embodiment of the cause identification and inference unit in step S4 of this example adopts a two-level architecture of a large language model parser and an external review unit. The LLM parser (with DeepSeek Chat as the backend) receives the P2 level evidence package and, with system prompts on constrained cause sets and JSON format, uses the LLM context understanding capability to output structured JSON containing the posterior probability distribution of causes, the highest confidence cause label, and the inference basis. The external review unit (reviewer) performs three independent checks on the output: (a) pattern validity check—fields are complete and enumeration values ​​are valid; (b) sensor consistency check—if the LLM determines that the sensor is out of place but CO2 increases significantly, it is marked as a sensor conflict; (c) confidence threshold—if the maximum posterior is lower than CONF_TH, the absolute_flag is set. If any check fails, the control flag is rolled back to a conservative baseline that maintains the state of the previous time step.

[0108] 5. Return time prediction and control flags (corresponding to S5 / S6)

[0109] In steps S5 and S6 of this embodiment, the return probability p_return(τ) = ΣP(C_k|features) × F_k(τ), where F_k is the truncated normal cumulative return time distribution of the k-th cause (parameters calibrated by pre-experiment), and τ is 5 minutes and 15 minutes. Based on the return probability and the expected idle time, according to... Figure 7 Logical generation control decisions: In-position short break → KEEP_OCC_REST; Low confidence → LOW_CONF_FALLBACK_HOLD; Batch arrival of meetings → PRE_COOL; 15-minute return probability ≥ 0.6 → KEEP_COOL; Lunch break / off of get off work → SETBACK.

[0110] In this embodiment, the preliminary experimental calibration method for the cumulative distribution F_k(τ) of return time for each cause is as follows: Without changing the above-mentioned three-level anonymized data stream, historical short-term occupancy change events for a set duration (two consecutive weeks in this embodiment) are collected during the initial stage of system deployment. For each type of cause C_k, the measured departure time sample from departure to return is statistically analyzed. The maximum likelihood estimation of this sample is performed using a truncated normal distribution N(μ_k, σ_k²; [a_k, b_k]), where the lower bound a_k is the shortest observable departure time for that cause, and the upper bound b_k is the maximum idle time allowed by the analysis window. The resulting (μ_k, σ_k) is used as the calibration parameter of F_k(τ) and is periodically updated with subsequent running data. Consistent with the typical duration of each cause described in the background technology, the truncated normal parameters obtained in this embodiment are shown in Table 3 below:

[0111] Table 3. Cut-off normal parameters obtained from calibration

[0112]

[0113] Substituting the calibrated F_k(τ) into the return probability p_return(τ)=ΣP(C_k|features)×F_k(τ), we obtain the average 5-minute return probability of 0.2085 and the average 15-minute return probability of 0.5007 in this embodiment (see the summary conclusion below). The difference between the two is about 2.4 times, which quantitatively reflects the highly dispersed departure return time caused by different factors, thus providing a parameter basis for the generation of differentiated control flags.

[0114] 6. Verification of Operation Results and Control Effect (corresponding to S4-S6)

[0115] In S4-S6 of this embodiment, the processing unit operates with a 5-minute analysis window, covering the period from 18:00:00 on May 30, 2026 to 13:40:00 on June 6, 2026, forming a total of 1965 time windows and detecting 895 short-term occupancy change events. Among them, there were 428 departure events, 421 arrival events, and 46 in-place rest events. The number of departures and arrivals is basically balanced, indicating that the system can reflect the closed-loop dynamics of short-term entry and exit in a shared office space for multiple users; the in-place rest events are identified separately, indicating that the present invention can cover the risk of accidental shutdown that may occur when the PIR is not triggered for a long time.

[0116] Depend on Figure 8 As can be seen, the control flag MAINTAIN was the most frequent, occurring 420 times, accounting for 46.9% of all events; LOW_CONF_FALLBACK_HOLD occurred 255 times, accounting for 28.5%; SETBACK occurred 78 times, SETBACK_SOFT occurred 16 times, totaling 94 times, accounting for 10.5%; KEEP_COOL occurred 76 times, accounting for 8.5%; KEEP_COOL_WATCH occurred 3 times; KEEP_OCC_REST occurred 46 times, accounting for 5.1%; and PRE_COOL occurred only once. This distribution indicates that the control strategy is not simply to shut down the air conditioner based on the "no one" result, but rather to adopt a conservative approach when there is low confidence or conflicting evidence, maintaining cooling when a short-term return is expected, and only reverting to the setting during longer periods of inactivity, reflecting risk perception and safety control characteristics. Furthermore, the 255 instances of abandonment of discrimination (LOW_CONF_FALLBACK_HOLD) precisely correspond to the global abandonment of discrimination rate of 0.2849 (255 / 895≈28.5%). Among these, approximately 59.6% (255 / 428) of the off-site events triggered conservative backoff due to confidence levels falling below the threshold or conflicting evidence. This indicates that the additional review unit proactively provides a safety net when evidence is insufficient, avoiding the direct conversion of uncertain inferences into potentially erroneous air conditioning actions.

[0117] Depend on Figure 9 As can be seen, the control flags exhibit significant time-of-day differences across different dates and hours. During weekdays, the flags are mainly in conservative or comfort-priority states such as KEEP_COOL, MAINTAIN, and LOW_CONF_FALLBACK_HOLD. During lunch breaks, after get off work, and in the evening, the number of SETBACK-type flags increases. Some blank windows in the heatmap indicate no valid events or no need to trigger control updates, demonstrating that this invention uses an event-driven approach to output the control interface, rather than forcibly changing the air conditioning state for each time window.

[0118] Depend on Figure 10As can be seen, the primary cause of departure events was using the restroom, totaling 161 times, accounting for 37.6% of all departure events; lunch break / leaving get off work accounted for 126 times, or 29.4%; smoking accounted for 117 times, or 27.3%; meeting endings accounted for 20 times, or 4.7%; and temporary absences accounted for 4 times, or 0.9%. The coexistence of short-term high return probability causes and longer periods of vacancy demonstrates that relying solely on changes in the number of people as input is insufficient to support shared air conditioning control. Therefore, the posterior probability of the causes and the return time distribution are necessary in this invention. It should be noted that water / tea breaks and short rests were not separately counted in the departure cause distribution: the former did not reach the trigger threshold during this embodiment's time period (counted as 0 times), while the latter was counted as an independent event type with 46 occurrences. Neither of these affects the calculation of the posterior probability of departure causes. Furthermore, the two high-probability recent return causes of toilet use (37.6%) and smoking (27.3%) accounted for 64.9% of the departure events. If the traditional method based solely on the number of people is adopted, these short-term departures are easily misjudged as requiring return, resulting in premature return and a second cooling-off period after the return of personnel. On the other hand, lunch break / off of get off work (29.4%) is a true long-term vacancy with a low probability of return and should be returned. This invention distinguishes between the two through posterior causal analysis, which is the core gain of its relative number-based method.

[0119] The abandonment discrimination rate in this embodiment is 0.2849, consistent with the average confidence level of 0.4727 for off-site events. This indicates that the review unit will proactively back off when there is low to medium confidence or insufficient evidence, avoiding the direct conversion of uncertain inferences into aggressive controls. The average 5-minute return probability for off-site events is 0.2085, the average 15-minute return probability is 0.5007, and the average expected idle time is 61.67 minutes. This indicates that short-term returns and longer idle times coexist in real office scenarios, requiring the combination of return probability, expected idle time, and confidence level to generate control flags. These results support the technical effectiveness of the present invention's "anonymous cause identification—return time prediction—risk perception control." In particular, the average 5-minute return probability (0.2085) differs from the average 15-minute return probability (0.5007) by about 2.4 times, indicating that the return time of off-site events is highly dispersed. Any traditional strategy that adopts a fixed backoff delay cannot simultaneously take into account both "fast return" and "long idle time" events. However, this invention dynamically calculates the return probability and expected idle time based on the posterior a priori cause, and can output differentiated control flags for events with different causes.

[0120] Of the 428 off-site events, this invention determined 79 events (76 KEEP_COOL events + 3 KEEP_COOL_WATCH events, accounting for 18.5%) to be of high return probability and maintained cooling, avoiding premature rollback and secondary pre-cooling after personnel return, which is a problem in traditional large-scale solutions for such short-term off-site events. It determined 94 events (78 SETBACK events + 16 SETBACK_SOFT events, accounting for 22.0%) to be of low return probability and rolled back to achieve energy savings. Furthermore, it maintained comfort control for 46 on-site rest events, avoiding false shutdowns of the PIR system caused by prolonged inactivity. These differentiated decisions collectively constitute the quantifiable gains in energy consumption and thermal comfort of this invention compared to the cause-independent solution.

[0121] In this embodiment, the external review unit performed three independent reviews on all 895 events: pattern legality check, sensor consistency check, and confidence threshold check. The average highest posterior confidence of the off-site events was 0.4727, and the global abandonment discrimination rate was 0.2849. Among them, approximately 59.6% of the off-site events were marked as abandonment discrimination because the maximum posterior probability was lower than the confidence threshold CONF_TH or because they triggered sensor consistency conflicts that contradicted the CO2 slope and off-site judgment. The corresponding control flags were then reverted to a conservative baseline that maintained the state of the previous moment, rather than directly executing air conditioning actions based on uncertain inferences. This mechanism enables the system to output differentiated control when there is sufficient evidence and to provide a safety net when there is insufficient evidence, explicitly transforming the uncertainty of causal inferences into risk-controllable control behavior.

[0122] To further verify the accuracy of the cause identification, 128 off-site events were randomly selected during the aforementioned runtime period and supplemented with manual on-site records as ground truth labels. The consistency rate between the highest confidence cause labels output by the large language model parser and the manually labeled ground truths (i.e., the cause identification accuracy rate) was 83.6%. Among them, the precision rates for the three main causes—toilet use, smoking, and lunch break / getting off work—were 88.0%, 81.8%, and 84.6%, respectively, and the recall rates were 89.8%, 77.1%, and 86.8%, respectively. The above results indicate that, without identifying individual identities and relying solely on anonymous evidence packages, this invention can still identify short-term off-site causes with acceptable accuracy, and its misjudgments are mainly concentrated on low-confidence samples that have been conservatively preserved by the review unit, thus avoiding erroneous air conditioning actions.

[0123] This embodiment uses a cause-independent scheme based solely on the number of people in a zone or PIR motion detection and employing a fixed backoff delay as a baseline. Based on the cause distribution and control flag distribution statistically analyzed in this embodiment, the expected gain of the present invention relative to this baseline is estimated as follows:

[0124] Table 4. Comparison Gains of Baseline Schemes with Relative Number of People

[0125]

[0126] Another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the shared anonymity behavior identification and short-term occupancy prediction method as described above.

[0127] In summary, the anonymous behavior causation identification and short-term occupancy prediction method for shared office spaces provided by this invention is based on anonymized sensor data such as district-level anonymized visual counting, entrance direction events, carbon dioxide and temperature / humidity, motion detection, and equipment power to detect three types of short-term occupancy change events. Without identifying individual identities, a six-dimensional causal feature vector is extracted and an anonymous evidence package is constructed. This package is then processed by a causal identification and inference unit based on a large language model and an additional reviewer unit to obtain the posterior probability distribution of potential causal factors such as using the restroom, smoking, fetching water, taking a short break, temporary absence, meeting gatherings / dispersals, and lunch breaks / leaving get off work. Furthermore, the return probability is calculated by combining the return time distribution of each causal factor, generating control-related flags such as maintaining cooling, reverting to previous settings, pre-cooling, or maintaining ventilation. This invention eliminates the need for individual identification and tracking, upgrading the representation of the number of people unrelated to causal factors to causal perception-based short-term prediction. It also ensures the credibility and secure control of the output through a three-level privacy protocol, an additional review, and a rejection mechanism.

[0128] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only.

[0129] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying anonymous behavior and predicting short-term occupancy in shared office spaces, characterized in that: Includes the following steps: S1. Collect multimodal raw observation data of the office space through the edge-side anonymous sensing architecture. The edge side does not output any images, pixels, or biometric data that can identify individuals or personal identities, but only outputs anonymous statistics of partition aggregation. S2. Detect short-term occupancy change events in office space based on the aforementioned multimodal raw observation data; S3. For each short-term occupancy change event, extract the event cause-related features without distinguishing or identifying any individual's identity, and encapsulate the features and the corresponding time window observation data into a structured anonymized evidence package. S4. Input the anonymous evidence into the cause identification and inference unit, and infer the posterior probability distribution corresponding to the cause of each predefined person's departure behavior. S5. Retrieve the return time distribution function pre-calibrated for each cause, and combine it with the posterior probability distribution of the cause to calculate the probability of personnel returning and the expected idle time of the partition under the specified duration. S6. Generate a dedicated control flag for the shared air conditioning system by combining three types of indicators: return probability, expected idle time, and causal reasoning confidence. When the reasoning confidence is lower than the preset threshold or there is evidence conflict in the multimodal observation data, the abandonment judgment logic is triggered, and the control flag reverts to the conservative baseline of maintaining the operating state of the previous device.

2. The method for identifying anonymous behavior and predicting short-term occupancy in shared office spaces according to claim 1, characterized in that: The anonymous sensing architecture in step S1 includes multiple types of dedicated anonymous sensing units: district-level anonymous visual counting unit, inlet direction event unit, carbon dioxide and temperature and humidity unit, PIR motion detection unit, device power unit and air supply status unit. The data flows unidirectionally according to a three-level privacy protocol: the original P0 image is discarded directly at the sensing edge and is not stored or uploaded to the cloud; the P1 layer contains aggregated data on the number of people in each partition, entry and exit counts, and activity candidate tags circulating within the local area network; only the P2 layer's anonymous evidence package in JSON format is input to the cause identification and inference unit.

3. The method for anonymous behavior identification and short-term occupancy prediction in shared office spaces according to claim 2, characterized in that: The entrance direction event unit uses a double infrared beam with staggered installation for soft curtain-type entrances that cannot stably trigger the door magnet. It determines the entrance and exit direction based on the order in which the two infrared beams are blocked and outputs a structured event containing a timestamp and the entrance and exit direction.

4. The method for identifying anonymous behavior and predicting short-term occupancy in shared office spaces according to claim 1, characterized in that: In step S2, short-term occupancy change events are divided into three categories: departure events where the number of people in a zone decreases, arrival events where the number of people in a zone increases, and resting events where the number of people remains the same but the activity level decreases. The rules for determining the "resting in place" event are as follows: if the number of people in a zone is always greater than zero within a set time window, the PIR unit trigger frequency decreases more than the standard relative to the 30-minute baseline, and there is no corresponding decrease in carbon dioxide concentration, it is determined that the person has not left the site, and the air conditioning shutdown action is prohibited.

5. The method for anonymous behavior identification and short-term occupancy prediction in shared office spaces according to claim 1, characterized in that: The causal features extracted in step S3 are a six-dimensional anonymous causal feature vector, which are as follows: event scale feature, exit direction feature, time period prior feature, event recurrence feature, equipment power coupling feature, and environmental attenuation feature.

6. The method for identifying anonymous behavior and predicting short-term occupancy in shared office spaces according to claim 5, characterized in that: Each dimension of characteristics has its own specific role in distinguishing causes: The event scale feature is used to distinguish between a single person leaving their post for a short period of time and a group of people leaving a meeting / workday. The exit direction features are distinguished by the type of door used for restrooms, smoking, going out, meetings, and lunch break departures. The time-period prior features introduce the prior probability of behavior occurring at each time period throughout the day; The event reproducibility features are used to identify periodic smoking disengagement behavior; The power coupling characteristics of the equipment are adjusted based on the short-term return probability of the on / off state of the workstation equipment. The environmental degradation characteristic is corrected for the actual number of people present in the zone by the rate of carbon dioxide decline.

7. The method for anonymous behavior identification and short-term occupancy prediction in shared office spaces according to claim 1, characterized in that: The cause identification and inference unit in step S4 is set up with a two-level verification and inference architecture, consisting of an LLM parser and an external review unit. The LLM parser receives P2 level anonymous evidence packets and, under the dual constraints of a predefined set of causes and a fixed output JSON format, outputs structured data containing the posterior probabilities of all causes, the highest confidence cause label, and the reasoning basis. The external review unit independently verifies the output results without altering the underlying parameters of the large model, and is compatible with various LLM parser backends.

8. The method for identifying shared anonymity behavior and predicting short-term occupancy according to claim 7, characterized in that: The external review unit performs three independent checks in parallel: Pattern validity check: Verify that the output JSON fields are complete and that the cause enumeration values ​​are within the preset set; Sensor consistency check: LLM identifies conflicting evidence when a sensor is displaced but CO2 concentration is rising positively. Confidence threshold check: If the maximum causal posterior probability is lower than the CONF_TH threshold, the decision is abandoned; if any check fails, the conservative baseline control flag is directly enabled.

9. The method for identifying shared anonymity behavior and predicting short-term occupancy according to claim 7, characterized in that: The cause identification and inference unit is configured with an offline alternative implementation: a Bayesian scoring framework; the posterior probability of a cause is proportional to the product of the prior weight of the time period and the likelihood factor of each feature, and all causes are normalized to obtain a probability distribution. The likelihood factor is calibrated by site pre-experimentation, without the need to call the large model network interface.

10. The method for identifying shared anonymity behavior and predicting short-term occupancy according to claim 1, characterized in that: The predefined potential cause categories include at least two or more categories such as using the toilet, getting water / tea break, smoking, taking a short rest at work, temporary absence, meeting gatherings and dispersals, and lunch break / off of get off work. For each cause category, the time for people to leave and return is fitted using a truncated normal distribution, and the mean, standard deviation, and upper and lower limits of the time truncation are calibrated in the pre-experiment.

11. The method for identifying shared anonymity behavior and predicting short-term occupancy according to claim 10, characterized in that: The formula for calculating the return probability in step S5 is as follows: p_return(τ)=ΣP(C_k|features)×F_k(τ), In the formula, P(C_k|features) is the posterior probability of cause Ck, and F_k is the truncated normal return time cumulative distribution of the k-th cause (parameters are calibrated by pre-experiment).

12. A shared anonymity behavior identification and short-term occupancy prediction system, characterized in that: The prediction system is used to perform the prediction method as described in any one of claims 1-11, and includes: Anonymous sensing units are used to collect multimodal anonymous sensing data in the office space and perform edge privacy data destruction; The event detection unit is used to identify three types of short-term occupancy change events: increase or decrease in the number of people in a zone, and short rest in place. The causal feature extraction and evidence package construction unit is used to extract six-dimensional causal features and generate a P2-level anonymous JSON evidence package; The causation identification and inference unit includes an LLM parser and a review unit or an offline Bayesian scoring module, which outputs the posterior probability distribution of each potential causation. The occupancy duration prediction unit calculates the return probability and expected idle duration for a specified duration based on the calibrated return distribution. The air conditioning control interface unit outputs a control flag to the shared VRV air conditioning system based on the prediction results, and triggers conservative fallback control when an anomaly is detected.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the shared anonymity behavior identification and short-term occupancy prediction method as described in any one of claims 1-11.

Images