An Intelligent Optical Emotion Intervention Method Based on a Triadic Closed-Loop Learning Mechanism
By employing a ternary closed-loop learning mechanism to perform event-level merging and online updates of the emotional light intervention system, the instability of parameter mapping relationships in existing technologies is resolved, thereby improving the stability and individual adaptability of emotional light intervention.
Patent Information
- Application Number
- CN202611026031.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-10
- Publication Date
- 2026-11-17
AI Technical Summary
When existing emotional light intervention systems continuously collect feedback data within a sliding time window, they tend to repeatedly update parameter mapping relationships by treating multiple consecutive window labels from the same light intervention as independent learning samples. This can lead to parameter combinations being overweighted, overweighted, or prematurely triggering conservative strategies, affecting the stability and individual adaptability of subsequent emotional light interventions.
A ternary closed-loop learning mechanism is adopted, which defines the continuous feedback process formed after the distribution of multidimensional light intervention parameters as an intervention event. By performing event-level merging of multiple candidate window labels within the same intervention event, event-level intervention result labels are generated and updated online based on the triplet, thereby reducing the interference of temporal correlation between adjacent window labels.
It reduces the risk of repeated reinforcement or weakening of parameter mapping relationships, improves the accuracy of the correspondence between closed-loop learning samples and actual light intervention processes, reduces the impact of short-term body movement, signal anomalies or environmental changes on the results, and avoids premature triggering of conservative parameter sets.
Smart Images

Figure CN123101807A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent lighting control, emotion computing, multimodal signal processing, and human-computer interaction, and particularly to an intelligent light emotion intervention method based on a ternary closed-loop learning mechanism. The ternary closed-loop learning mechanism is specifically used to associate the event-initial emotional state vector, the event light intervention parameters, and the event-level intervention result label, and update the parameter mapping relationship based on the triple formed by the three. Background Technology
[0002] In recent years, mood-adjusting lighting, health lighting, and sleep-aid lighting have gradually evolved from static scene control to dynamic control based on physiological feedback. Existing solutions typically determine the user's state by collecting data such as heart rate, heart rate variability, skin conductance, body movement, or ambient light, and adjust color temperature, illuminance, color, or dynamic change modes accordingly to achieve effects such as relaxation, soothing, focus, or sleep assistance. Some solutions also continue to collect user feedback data after light intervention and correct subsequent lighting parameters based on the feedback results.
[0003] However, in emotional light intervention systems that use sliding time windows to continuously collect feedback data, the output of a single light parameter usually has a continuous effect in multiple consecutive windows. The heart rate variability, skin conductance, body movement changes, and emotional state changes collected in subsequent windows are often continuous responses to the same light intervention and are not independent of each other. Existing closed-loop update methods usually treat the intervention result labels generated by each feedback window as independent learning samples, or only trigger parameter mapping relationship updates based on the same labels in multiple consecutive windows, without distinguishing the temporal correlation and homologous intervention relationship between consecutive window labels.
[0004] For example, after the system outputs a set of soothing light parameters with low illumination, warm color, and low dynamic change frequency, the user showed increased heart rate variability or decreased skin conductance in multiple subsequent windows. If the system treats these consecutive windows as multiple valid intervention samples, it will mistakenly regard the continuous feedback of a single light intervention as multiple independent verifications, thus repeatedly increasing the weight of the light parameter combination. Conversely, if multiple windows after an intervention are marked as invalid due to brief physical movements, environmental changes, or abnormal signals, the same parameter combination may be repeatedly downweighted, and the conservative parameter set may be triggered prematurely. Therefore, existing technologies still lack a label merging and update constraint mechanism for intervention events within the three-element closed-loop learning framework. This can easily lead to over-strengthening, over-weakening, or bias towards the outcome of a single intervention, affecting the stability and individual suitability of subsequent emotional light interventions. Summary of the Invention
[0005] This invention provides an intelligent light-based emotion intervention method based on a ternary closed-loop learning mechanism. This method addresses the problem that existing light-based emotion intervention systems, when continuously collecting feedback data using a sliding time window and performing online learning, tend to repeatedly update parameter mapping relationships by treating multiple consecutive window labels after the same light intervention as independent learning samples. This leads to excessive weighting, excessive deweighting, or premature triggering of conservative strategies in parameter combinations.
[0006] The ternary closed-loop learning mechanism in this invention does not directly use the candidate window labels generated by a single sliding time window as the basis for online updates. Instead, it defines the continuous feedback process formed after the issuance of multidimensional light intervention parameters as an intervention event, and performs event-level merging on multiple candidate window labels within the same intervention event to obtain an event-level intervention result label. Then, based on the event-initial emotional state vector, the event light intervention parameters, and the event-level intervention result label, the parameter mapping relationship is updated online once, thereby reducing the interference of the temporal correlation of adjacent window labels on the closed-loop learning results.
[0007] To achieve the above objectives, the present invention adopts the following technical solution: An intelligent optical emotion intervention method based on a three-element closed-loop learning mechanism, executed by a processor, includes: acquiring user physiological signals collected by a physiological acquisition device and environmental information collected by an environmental sensor; The user's physiological signals and environmental information are time-aligned, denoised, and feature-extracted according to a preset sliding time window to obtain the current window feature Ft; The current emotional state vector Et is determined based on the current window feature Ft; based on the emotional state vector Et, the user's individual baseline parameter Bu, and the environmental parameter Ct, the multidimensional light intervention parameter Pt for the current moment is generated through the parameter mapping relationship Mt and sent to the light output terminal for execution; When the multidimensional light intervention parameter Pt is sent and executed, the corresponding intervention event Ik is established, and the intervention event Ik is bound to the event start emotional state vector Ek, the event light intervention parameter Pk, and the event start time Tk. During the period from the start of the intervention event Ik to the satisfaction of the event termination conditions, feedback data FBt after the intervention is obtained based on multiple consecutive sliding time windows, and multiple candidate window labels yt are generated. After the intervention event Ik meets the event termination condition, multiple candidate window labels yt belonging to the same intervention event Ik are merged at the event level to obtain an event-level intervention result label Yk; Based on triples<Ek,Pk,Yk> Perform an event-level online update on the parameter mapping relationship Mt, and use the updated parameter mapping relationship for subsequent multidimensional optical intervention parameter decisions.
[0008] Preferably, the event-level intervention result label Yk is one of the following: effective, partially effective, invalid, unavailable, or uncertain; the processor determines Yk by label consistency merging and time-weighted merging. When the two merging results are inconsistent, Yk is determined to be uncertain.
[0009] The embodiments of the present invention have at least the following beneficial effects: First, the feedback labels generated by the continuous sliding window are attributed to the intervention event, avoiding the misidentification of the continuous feedback of the same light intervention as multiple independent intervention samples, and reducing the risk of repeated reinforcement or weakening of the parameter mapping relationship; Second, by merging multiple candidate window labels within the same intervention event at the event level, the update unit of online learning is upgraded from window-level samples to event-level samples, thereby improving the accuracy of the correspondence between closed-loop learning samples and the actual light intervention process. Third, by judging window availability, judging label consistency, and merging time weights, the impact of short-term physical movement, signal abnormalities, or environmental changes on the labels of event-level intervention results can be reduced. Fourth, by limiting the parameter mapping relationship update to only once for the same intervention event, and requiring that consecutive valid or consecutive invalid counts come from different intervention events, we can avoid premature triggering of conservative parameter sets or excessive weighting of valid parameter combinations due to high correlation between adjacent windows. Attached Figure Description
[0010] Figure 1 This is the overall flow of the method of the present invention; Figure 2 This is a schematic diagram of the ternary closed-loop learning mechanism of the present invention; Figure 3 This is a schematic diagram illustrating the establishment and feedback window allocation for intervention events in this invention; Figure 4 This is a schematic diagram of the event-level merging of candidate window labels in this invention; Figure 5 This is a schematic diagram of the event-level online update of the parameter mapping relationship in this invention; Figure 6 This is a flowchart of the conservative parameter set switching process of the present invention. Detailed Implementation
[0011] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the described embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.
[0012] In this embodiment of the invention, the processor can be located in a mobile terminal, gateway, smart lighting controller, wearable terminal, or cloud server; the physiological acquisition device is used to acquire user physiological signals, the environmental sensor is used to acquire environmental information, and the light output terminal is used to execute multidimensional light intervention parameters; the processor can communicate with the physiological acquisition device, environmental sensor, and light output terminal via wired or wireless means.
[0013] User physiological signals include at least two of the following: heart rate variability-related signals, skin conductance signals, and body movement signals; heart rate variability-related signals can be obtained from pulse wave signals, electrocardiogram signals, or equivalent physiological signals; environmental information includes at least ambient illuminance, and may also include at least one of the following: ambient color temperature, time period, and spatial scene type.
[0014] In this embodiment of the invention, the sliding time window refers to a time segment used for continuously collecting and processing physiological signals, environmental information, and intervention feedback data; the window length can be from 5 seconds to 60 seconds, and the step size can be from 0.5 seconds to 5 seconds. The window feature Ft is a set of features constructed from data within a sliding time window, including at least two of the following: heart rate variability features, skin conductance features, body motion intensity features, and ambient illuminance features.
[0015] In this embodiment of the invention, the emotional state vector is denoted as Et, which is used to characterize the user's emotional state at the current moment. It preferably includes a valence parameter vt and an arousal parameter at. The valence parameter vt and the arousal parameter at are in the range of [-1, 1] after being limited. The lower vt is, the stronger the negative emotional tendency is, and the higher at is, the higher the arousal level is. The multidimensional light intervention parameter is denoted as Pt, which is used to drive the light output terminal to perform emotional intervention. It preferably includes at least three of the following: correlated color temperature CCTt, target illuminance Lt, dynamic change frequency ft, and gradual change time τt.
[0016] In one embodiment, before determining the emotional state vector Et, the processor first performs individualized baseline correction on the current window feature Ft based on the user's individual baseline parameters Bu. The user's individual baseline parameters Bu include at least one of baseline heart rate variability HRVbase,u, baseline electroskin activity EDAbase,u, baseline body motion intensity Abase,u, resting heart rate HRbase,u, and comfort threshold. The processor calculates relative deviation features based on the heart rate variability, electroskin activity, and body motion intensity of the current window, for example, h^ut=Norm((HRVt-HRVbase,u) / (HRVbase,u+ε)), e^ut=Norm((EDAt-EDAbase,u) / (EDAbase,u+ε)), mot^ut=Norm((At-Abase,u) / (Abase,u+ε)), where ε is a preset constant to prevent the denominator from being zero. The processor calculates relative deviation features based on h^ut, e^ut, and mot^ut. The valence parameter vt and arousal parameter at are calculated to make the emotional state vector reflect the user's state changes relative to their baseline.
[0017] The target emotional range is denoted as Ωg, which is a preset range determined by the processor within the valence-arousal two-dimensional space based on the current intervention target. For example, the soothing intervention target corresponds to Ωg={(v, a)|0.20≤v≤0.80, -0.50≤a≤0.20}, the sleep-aid intervention target corresponds to Ωg={(v, a)|0.00≤v≤0.60, -0.80≤a≤0.00}, and the focus maintenance intervention target corresponds to Ωg={(v, a)|0.10≤v≤0.70, 0.00≤a≤0.50}. The processor uses the Ωg corresponding to the current intervention target when determining ΔDt and judging whether the emotional state has entered the target emotional range.
[0018] The parameter mapping relationship is denoted as Mt, which is used to represent the mapping relationship between the emotional state vector, the user's individual baseline parameters, environmental parameters and multidimensional light intervention parameters; the intensity of body movement is uniformly denoted as At, and the amount of change in body movement is uniformly denoted as ΔAt, in order to avoid confusion with the parameter mapping relationship Mt.
[0019] An intervention event is denoted as Ik, which refers to the feedback acquisition process formed during the continuous action of a multidimensional light intervention parameter after it is issued and executed. The candidate window label is denoted as yt, which refers to the temporary evaluation result generated based on feedback data within a single sliding time window; The event-level intervention outcome label is denoted as Yk, which refers to the evaluation result obtained after merging multiple candidate window labels yt belonging to the same intervention event Ik. Yk can be effective, partially effective, ineffective, unavailable, or uncertain.
[0020] like Figure 1 and Figure 2 As shown, the ternary closed-loop learning mechanism refers to the processor determining and executing the event-level intervention parameters Pk based on the event's initial emotional state vector Ek, then generating the event-level intervention result label Yk based on the candidate window labels within the intervention event, and finally determining the result based on the triplet.<Ek,Pk,Yk> The closed-loop process of updating the parameter mapping relationship Mt; This mechanism does not use the candidate window labels of each sliding time window as independent update samples, but rather uses the intervention event as the update unit.
[0021] In one embodiment, the physiological acquisition device continuously acquires at least two of the user's heart rate variability-related signals, skin conductance signals, and body movement signals, while the environmental sensor continuously acquires ambient illuminance; the processor performs time alignment, noise reduction, and feature extraction on the above data according to a preset sliding time window to obtain the current window feature Ft; Heart rate variability characteristics may include at least one of SDNN, RMSSD, LF, HF, or LF / HF; skin conductance characteristics may include at least one of skin conductance level, skin conductance response frequency, or skin conductance variation amplitude. The body dynamic intensity characteristic At can be represented by the acceleration modulus, angular velocity modulus, or body dynamic energy value; the ambient illuminance characteristic can be represented by the current ambient illuminance value or the change in ambient illuminance.
[0022] The user's individual baseline parameter Bu is established during the initialization phase and is invoked in both the emotional state vector calculation and the generation of light intervention parameters. Bu includes at least three of the following: baseline heart rate variability (HRVbu), baseline skin conductance (EDAbu), baseline body motion intensity (Abu), baseline ambient illuminance (ELbu), preferred color temperature (CCTbu), and illuminance comfort baseline (Lbu). The processor continuously collects data for a preset number of sliding time windows while the user is in a non-interventional or resting state, and performs statistical analysis on heart rate variability, skin conductance, body motion intensity, ambient illuminance, and user manual dimming records to obtain the aforementioned baseline parameters. When historical data is insufficient, the processor uses preset default baseline parameters and updates Bu with low weight after the subsequent effective intervention event ends.
[0023] The processor determines the current emotional state vector Et based on the current window feature Ft; the emotional state vector Et can be output by the emotional state determination model or obtained by rule calculation; the emotional state determination model can be a regression model, a time series model, a rule model or a combination thereof.
[0024] In a rule-based implementation, the processor first performs baseline correction on the current window feature Ft based on the user's individual baseline parameter Bu, and then performs normalization processing. For example, the processor calculates the relative heart rate variability feature ht=Norm((HRVt-HRVbu) / max(HRVbu, ε)), the relative skin conductance feature et=Norm((EDAt-EDAbu) / max(EDAbu, ε)), the relative body motion intensity feature mott=Norm((At-Abu) / max(Abu, ε)) and the relative ambient illuminance feature lt=Norm((ELt-ELbu) / max(ELbu, ε)) based on HRVbu, EDAbu, Abu, and ELbu, respectively, where ELt is the average ambient illuminance of the current window and ε is a preset positive number to prevent the denominator from being zero.
[0025] The processor calculates the valence parameter vt and the arousal parameter at according to vt=c0+cB1·Bv+c1·ht-c2·et-c3·mott-c4·lt, at=d0+dB1·Ba+d1·et+d2·mott-d3·ht+d4·lt; where Bv is the valence baseline offset determined by the user's individual baseline parameter Bu, Ba is the arousal baseline offset determined by the user's individual baseline parameter Bu, and c0, cB1, c1, c2, c3, c4, d0, dB1, d1, d2, d3 and d4 are all preset coefficients. After amplitude limiting, the calculation results are located in the interval [-1, 1].
[0026] Through the baseline correction described above, the same HRVt, EDAt, or At absolute value will not be mechanically interpreted by the processor as the same emotional state; the processor uses the degree of deviation of the current user from its own Bu as the input of the emotional state vector Et, thereby avoiding distortion of vt and at due to differences in individual heart rate variability, skin conductance, or exercise habits.
[0027] During the parameter generation phase, the processor continues to call the user's individual baseline parameters Bu, including the preferred color temperature CCTbu, the illuminance comfort benchmark Lbu, and historical manual dimming records, to impose comfort constraints on the candidate light intervention parameter combinations.
[0028] For example, when the candidate target illuminance Lt exceeds the upper limit corresponding to the illuminance comfort baseline Lbu, the processor reduces the priority of the candidate parameter combination or limits Lt to the user's illuminance comfort threshold range.
[0029] The processor generates the multidimensional light intervention parameter Pt for the current moment through the parameter mapping relationship Mt based on the emotion state vector Et, the user's individual baseline parameter Bu, and the environmental parameter Ct. That is, Pt=Mt(Et, Bu, Ct). The environmental parameter Ct may include at least one of the following: current ambient illuminance, ambient color temperature, time period, or scene state.
[0030] In one embodiment, the parameter mapping relationship Mt is a candidate parameter combination priority table, which includes multiple emotional state intervals, multiple candidate light intervention parameter combinations, and a priority weight W corresponding to each candidate light intervention parameter combination. Based on the emotional state range to which Et belongs, and in conjunction with Bu and Ct, the processor selects the parameter combination with the highest priority weight from the candidate parameter combinations that meet the comfort threshold as Pt.
[0031] In one example, the processor sets the low valence threshold to Vlow=-0.20 and the high arousal threshold to Ahigh=0.40. When the valence parameter vt≤Vlow and the arousal parameter at≥Ahigh, the processor determines that the user is in a low valence, high arousal state and generates multi-dimensional light intervention parameters with lower correlated color temperature, lower target illuminance, lower dynamic change frequency and longer gradation time to create a soothing light environment. The above thresholds can be adjusted with low weight based on the user's individual baseline parameter Bu and historical effective intervention events.
[0032] In another example, when the user is focused on their needs and does not show signs of fatigue, the processor generates multi-dimensional light intervention parameters, including a neutral-to-cool correlated color temperature, a relatively stable target illuminance, and a low or zero dynamic change frequency.
[0033] The processor sends the multidimensional optical interference parameter Pt to the optical output terminal for execution; The light output terminal can be at least one of the following: table lamp, ceiling lamp, wall lamp, ambient lamp, eye mask-type light-emitting device, and wearable light-emitting device; When the optical output terminal supports color temperature control, the processor converts CCTt into a color temperature control instruction; when the optical output terminal only supports brightness percentage control, the processor converts the target illuminance Lt into a brightness percentage according to the preset illuminance-brightness calibration relationship. The dynamic frequency ft is used to generate a periodic dimming curve, and the gradient time τt is used to generate a gradient curve that transitions from the current control value to the target control value.
[0034] like Figure 3 As shown, when the multidimensional light intervention parameter Pt is sent to the light output terminal for execution, the processor establishes the corresponding intervention event Ik and binds the intervention event Ik with the event start emotional state vector Ek, the event light intervention parameter Pk, and the event start time Tk; Where Ek is the emotional state vector when the intervention event is established, and Pk is the multidimensional light intervention parameter sent and executed when the intervention event is established; Tk is the time when the parameter is issued or the time when the optical output terminal confirms the start of execution; the gradient time in the event optical intervention parameter Pk is denoted as τk.
[0035] In one embodiment, the processor obtains the multidimensional light intervention parameter Pt at the current moment and the multidimensional light intervention parameter P{t-1} executed at the previous moment, calculates at least two of the following: the correlated color temperature difference, the target illuminance difference, the dynamic change frequency difference, and the gradual change time difference, and normalizes the above differences. The processor obtains the parameter change DP by weighting the normalized differences; for example, DP = w1·Norm(|CCTt-CCT{t-1}|) + w2·Norm(|Lt-L{t-1}|) + w3·Norm(|ft-f{t-1}|) + w4·Norm(|τt-τ{t-1}|); where w1, w2, w3 and w4 are non-negative preset weights, and w1+w2+w3+w4=1; Norm(·) is the normalization function.
[0036] When the parameter change DP is greater than the preset parameter change threshold, the processor creates a new intervention event; when the parameter change DP is less than or equal to the preset parameter change threshold, the processor does not create a new intervention event, but instead assigns the feedback data corresponding to the current sliding time window to the previous intervention event.
[0037] The termination condition for intervention event Ik includes at least one of the following: the change in the current multidimensional light intervention parameter relative to the event light intervention parameter Pk is greater than a preset parameter change threshold; The duration of the intervention event has reached the preset maximum event duration; the number of available candidate window labels within the event has reached the preset number of evaluation windows; The user's emotional state enters the target emotional range and remains continuously for a preset number of hours; or the user actively terminates the light intervention.
[0038] Before executing the judgment that "the user's emotional state enters the target emotional range and is maintained for a preset number of windows", the processor first reads the target emotional range Ωg according to the current intervention target. Ωg is limited by four boundary values, namely Ωg={(v, a)|vg1≤v≤vg2,ag1≤a≤ag2}, where vg1 is the lower limit of valence, vg2 is the upper limit of valence, ag1 is the lower limit of arousal, and ag2 is the upper limit of arousal. The above four boundary values can be determined by factory configuration, user target selection, or statistical results of historical effective intervention events.
[0039] The target emotional interval Ωtarget refers to the region in the valence-arousal plane where the emotional state vector Et falls. If Et = (vt, at), and the range of vt is [-1, 1] and the range of at is [0, 1], then the target emotional interval Ωtarget can be expressed as: [Vmin, Vmax]×[Amin, Amax]; where Vmin and Vmax are the lower limit and upper limit of the target valence, respectively, and Amin and Amax are the lower limit and upper limit of the target arousal, respectively. In a relaxing and soothing scenario, Ωtarget can be set to [0.10, 0.60] × [0.20, 0.50]; in a bedtime companionship scenario, Ωtarget can be set to [0.00, 0.50] × [0.10, 0.35]; in a focused work scenario, Ωtarget can be set to [0.10, 0.70] × [0.35, 0.65]; the processor can select the corresponding target emotional range based on the scenario type, the user's individual baseline parameter Bu, or the user's preference record.
[0040] The current emotional distance Dt is the minimum Euclidean distance from Et to Ωtarget; when Et is inside Ωtarget, Dt=0; when Et is outside Ωtarget, Dt=sqrt(dv^2+da^2), where dv=max(Vmin-vt, 0, vt-Vmax) and da=max(Amin-at, 0, at-Amax).
[0041] During the period from the start of the intervention event Ik to the satisfaction of the event termination conditions, the processor acquires the post-intervention feedback data FBt based on multiple consecutive sliding time windows; the feedback data FBt includes at least two of the following: heart rate variability change ΔHRVt, skin electrical activity change ΔEDAt, body movement change ΔAt, and the improvement in the distance of the emotional state to the target emotional zone ΔDt.
[0042] In one embodiment, the processor uses at least one sliding time window before the establishment of the intervention event Ik as the baseline window to obtain the baseline heart rate variability HRVbase, baseline skin conductance activity EDAbase, baseline body movement intensity Abase, and baseline emotional distance Dbase; the processor uses the current sliding time window after the establishment of the intervention event Ik as the feedback window to obtain the current heart rate variability HRVt, current skin conductance activity EDAt, current body movement intensity At, and current emotional distance Dt.
[0043] Wherein, the current emotional distance Dt is the distance between the current emotional state vector Et and the current target emotional interval Ωg; let Ωg={(v, a)|vg1≤v≤vg2,ag1≤a≤ag2}, the processor first calculates vclip=min(max(vt,vg1),vg2) and alip=min(max(at,ag1),ag2), and then calculates the current emotional distance according to Dt=sqrt((vt-vclip)^2+λa·(at-aclip)^2), where λa is the arousal distance weight; if Et is located within Ωg, then vclip=vt, alip=at and Dt=0.
[0044] ΔHRVt=HRVt-HRVbase, ΔEDAt=EDAt-EDAbase, ΔAt=At-Abase, ΔDt=Dbase-Dt; If ΔHRVt is positive, ΔEDAt is negative, ΔAt is negative and ΔDt is positive, it usually indicates that the intervention feedback of the current window is changing in the direction of improvement; When any indicator in the feedback data FBt cannot be obtained from the acquired signal, the processor does not use that indicator in the calculation of the window intervention score St, and St is calculated from at least two available feedback indicators.
[0045] In one implementation, the processor first calculates the window availability identifier Qt for each sliding time window. The processor obtains the proportion of missing physiological signals rm, the proportion of noisy sampling points rn, and the proportion of sampling points exceeding the effective amplitude range ro for the current sliding time window, and calculates the physiological signal quality score SQt according to SQt=clip(1-rm-rn-ro, 0, 1), where clip indicates that the calculation result is restricted to between 0 and 1; The processor also obtains the motion intensity At and the change in ambient illuminance for the current sliding time window. The change in ambient illuminance is the absolute value of the difference between the average ambient illuminance of the current window and the average ambient illuminance of the previous window.
[0046] When the physiological signal quality score SQt is lower than the preset signal quality threshold, or the body movement intensity At is higher than the preset body movement threshold, or the change in ambient illumination is higher than the preset environmental change threshold, the processor will determine the window availability flag Qt as unavailable and the corresponding candidate window label yt as unavailable.
[0047] When the physiological signal quality score SQt, body movement intensity At, and change in ambient illuminance all meet the corresponding threshold conditions, the processor determines the window availability flag Qt as available and allows the generation of valid, partially valid, or invalid candidate window labels based on the window intervention score St.
[0048] In one embodiment, the processor normalizes at least two of ΔHRVt, -ΔEDAt, -ΔAt and ΔDt to obtain the corresponding normalization improvement value; Normalization is performed using Norm(x)=clip((x-xmin) / (xmax-xmin),0,1), where xmin and xmax are preset normalization boundaries or obtained from historical samples, and clip means that the calculation result is restricted to between 0 and 1.
[0049] The processor performs a weighted summation of the normalized improvement values according to preset weights to obtain the window intervention score St; For example, St=b1·Norm(ΔHRVt)+b2·Norm(-ΔEDAt)+b3·Norm(-ΔAt)+b4·Norm(ΔDt); Among them, b1, b2, b3 and b4 are non-negative preset weights, and the sum of the weights involved in the calculation is 1; When St is greater than or equal to the first label threshold, the processor determines the candidate window label yt as valid; when St is less than the first label threshold and greater than or equal to the second label threshold, the candidate window label yt is determined as partially valid. When St is less than the second label threshold, the candidate window label yt is determined to be invalid; the first label threshold is greater than the second label threshold.
[0050] like Figure 4 As shown, when the intervention event Ik meets the event termination condition, the processor performs event-level merging on multiple candidate window labels yt belonging to the same intervention event Ik to obtain an event-level intervention result label Yk; Multiple candidate window labels yt within the same intervention event Ik are only used as data sources for generating event-level intervention result labels Yk, and are not used as independent learning samples to update the parameter mapping relationship Mt separately.
[0051] In one implementation, the processor removes sliding time windows labeled yt as unusable, thus obtaining a set of available candidate window labels; When the number of labels in the available candidate window label set is less than the preset minimum number of windows, the processor determines the event-level intervention result label Yk as unavailable and prohibits the intervention event Ik from participating in the online update of the parameter mapping relationship Mt.
[0052] When the number of labels in the available candidate window label set is greater than or equal to the preset minimum number of windows, the processor counts the number of valid labels ne, the number of partially valid labels np, and the number of invalid labels ni, and calculates the label consistency Rk according to Rk=max(ne, np, ni) / (ne+np+ni); When Rk is greater than or equal to the preset consistency threshold, the label type with the largest quantity is determined as the candidate event-level label; When Rk is less than the preset consistency threshold, the event-level intervention result label Yk is determined as uncertain, and this intervention event is prohibited from participating in the online update of the parameter mapping relationship Mt.
[0053] After the candidate event-level label is determined, the processor further performs time-weighted merging; the processor determines the time weight βt corresponding to the candidate window label yt according to the time difference between the end time Tt of the sliding time window corresponding to the candidate window label yt and the event start time Tk.
[0054] In an example, when Tt-Tk≤τk, βt=β0; when τk<Tt-Tk≤τk+Ts, βt=1; when Tt-Tk>τk+Ts, βt=β1; Wherein, τk is the gradient time in the event light intervention parameter Pk, Ts is the stability evaluation duration, 0<β0<1, 0<β1<1, and the sum of the time weights of the candidate window labels participating in the calculation is greater than 0.
[0055] The processor maps the valid label, partially valid label and invalid label to preset label scores respectively, wherein score(valid)=1, score(partially valid)=0.5, score(invalid)=0, and the unavailable label does not participate in the calculation; and calculates the event-level score Gk based on Gk=Σβt·score(yt) / Σβt; The processor determines the weighted event-level label according to the comparison result between Gk and the event-level label threshold; when the weighted event-level label is consistent with the candidate event-level label, the event-level intervention result label Yk is determined as the candidate event-level label; when the two are inconsistent, Yk is determined as uncertain.
[0056] As Figure 5 shown, the processor performs one event-level online update on the parameter mapping relationship Mt based on the triplet <Ek, Pk, Yk>; In this triplet, Ek is the emotional state input when the intervention event is established, Pk is the correspondingly executed light intervention parameter, and Yk is the event-level feedback result obtained by merging multiple candidate window labels in this intervention event.
[0057] In one embodiment, the parameter mapping relationship Mt includes a candidate parameter combination priority table, and the candidate parameter combination priority table includes a plurality of emotional state intervals, a plurality of candidate light intervention parameter combinations, and a priority weight W corresponding to each candidate light intervention parameter combination.
[0058] When the event-level intervention result label Yk is valid, the processor updates the priority weight corresponding to the event light intervention parameter Pk to Wnew=clip(Wold+η1, Wmin, Wmax); When Yk is partially valid, the priority weight is updated to Wnew=clip(Wold+η2, Wmin, Wmax), where η2 is less than η1; When Yk is invalid, the priority weight is updated to Wnew=clip(Wold-η3, Wmin, Wmax); when Yk is unavailable or uncertain, the priority weight W is not updated. Where η1, η2, and η3 are positive numbers, and Wmin and Wmax are the lower and upper limits of the priority weight, respectively.
[0059] The updated priority weights are used by the processor to select subsequent multidimensional light intervention parameters when the emotional state is similar to the event's initial emotional state vector Ek. Among them, similar emotional states are determined based on whether the distance between emotional state vectors is less than a preset emotional similarity threshold. For example, if Et=(vt, at) and Ek=(vk, ak), then we can determine whether the two are similar based on whether sqrt((vt-vk)^2+(at-ak)^2) is less than a preset emotional similarity threshold.
[0060] like Figure 6 As shown, in one embodiment, the processor records multiple event-level intervention result labels Yk according to the order in which the intervention events are established; when determining whether consecutive valid labels or consecutive invalid labels meet the parameter mapping relationship adjustment conditions, only event-level intervention result labels from different intervention events that are valid, partially valid or invalid are counted in the consecutive count, and multiple candidate window labels yt belonging to the same intervention event are not counted in the consecutive count.
[0061] When multiple consecutive event-level intervention result labels from different intervention events are all valid, the processor increases the output priority of the corresponding event light intervention parameters, or similar event light intervention parameters under similar emotional states. When multiple consecutive event-level intervention result labels from different intervention events are invalid, the processor reduces the output priority of the corresponding event light intervention parameters, or the output priority of similar event light intervention parameters in similar emotional states.
[0062] When multiple consecutive event-level intervention result labels from different intervention events are invalid, and the number of consecutive invalid results reaches a preset invalid number threshold, the processor will switch the subsequent multidimensional light intervention parameter decision to a conservative parameter set. The conservative parameter set includes at least two of the following: dynamic change frequency not higher than a preset conservative frequency threshold, single change amplitude of target illuminance not higher than a preset illuminance change threshold, single change amplitude of correlated color temperature not higher than a preset color temperature change threshold, and gradation time not lower than a preset gradation time threshold.
[0063] When the event-level intervention result label of at least one subsequent intervention event is valid or partially valid, and the proportion of available candidate window labels within that intervention event is greater than the preset recovery proportion threshold, the processor exits the conservative parameter set and resumes generating multidimensional light intervention parameters based on the updated parameter mapping relationship Mt.
[0064] In a home relaxation setting, physiological acquisition devices collect user heart rate variability-related signals, skin conductance signals, and body movement signals, while environmental sensors collect ambient illuminance. The processor obtains Et=(vt, at)=(-0.35, 0.62) in a certain window, where vt≤-0.20 and at≥0.40, thus recognizing that the user is in a state of low valence and high arousal. The processor determines the current intervention target as the soothing intervention target and reads Ωg={(v, a)|0.20≤v≤0.80, -0.50≤a≤0.20}, then generates Pk={2600K, 85lx, 0.05Hz, 40 seconds} and sends it to the main lamp and the desk lamp; the processor also establishes the intervention event Ik. The processor obtains the emotional state vector Ek=(-0.42, 0.78) in a certain window, where vk=-0.42 is lower than the low valence threshold Vl=-0.30, and ak=0.78 is higher than the high arousal threshold Ah=0.70. Therefore, the processor determines that the user is in a low valence, high arousal state. The processor generates event light intervention parameters Pk={2600K, 85lx, 0.05Hz, 40 seconds} and sends them to the main lamp and desk lamp for execution.
[0065] In subsequent sliding time windows, the processor calculates ΔHRVt, ΔEDAt, ΔAt and ΔDt respectively, and generates multiple candidate window labels; If the first two windows are within a transition period, the corresponding time weight is lower; if most available candidate window labels are valid within the subsequent stable evaluation window, and label consistency meets the threshold condition, and the weighted event-level label is consistent with the candidate event-level label, then the processor determines the event-level intervention result label Yk as valid, and based on...<Ek,Pk,Yk> Perform an online update on the parameter mapping relationship Mt; Even if there are multiple valid candidate window labels in the intervention event, only one event-level online update is performed; In another scenario, if a user gets up, the ambient light suddenly turns on, or the sensor malfunctions during an intervention event, the processor will determine the corresponding candidate window label as unavailable based on the window availability flag. If the number of available candidate window labels within an intervention event is insufficient, the processor will determine the event-level intervention result label as unavailable and prohibit the event from participating in online updates to avoid abnormal feedback polluting the parameter mapping relationship.
[0066] Those skilled in the art can make equivalent substitutions or modifications to the acquisition modality, the dimension of the emotional state vector, the form of the parameter mapping relationship, the event-level merging method, the type of optical output terminal, and the communication method without departing from the spirit and substance of the present invention, and all such substitutions or modifications should fall within the protection scope of the present invention.
Claims
1. A method for intelligent optical emotion intervention based on a ternary closed-loop learning mechanism, characterized in that, Executed by the processor: acquire user physiological signals and environmental information, construct window features Ft according to the sliding time window, and determine the emotional state vector Et; Based on Et, the user's individual baseline parameter Bu, and the environmental parameter Ct, a multidimensional light intervention parameter Pt is generated and sent to the light output terminal for execution; when Pt is executed, an intervention event Ik is established, and the event's initial emotional state vector Ek and the event's light intervention parameter Pk are bound together; Within Ik, acquire data from multiple feedback windows and generate candidate window labels yt; after Ik ends, merge multiple yt values within the same event into a single event-level intervention result label Yk; Based on triples<Ek,Pk,Yk> Perform an event-level online update on the parameter mapping relationship Mt. Multiple yt within the same Ik are not treated as independent samples and are updated separately for Mt.
2. The method according to claim 1, characterized in that, The user physiological signals include at least two of the following: heart rate variability-related signals, skin conductance signals, and body movement signals; the environmental information includes at least ambient illuminance; and the window feature Ft includes at least two of the following: heart rate variability features, skin conductance features, body movement intensity features, and ambient illuminance features.
3. The method according to claim 1, characterized in that, The emotional state vector Et includes a valence parameter vt and an arousal parameter at; The processor performs baseline correction and normalization on heart rate variability, skin conductance, body motion intensity, and ambient illuminance characteristics based on the user's individual baseline parameter Bu, to obtain relative heart rate variability, relative skin conductance, relative body motion intensity, and relative ambient illuminance characteristics, and outputs the vt and at through a preset rule model, regression model, or time series model.
4. The method according to claim 1, characterized in that, The multidimensional light intervention parameter Pt includes at least three of the following: correlated color temperature CCTt, target illuminance Lt, dynamic change frequency ft, and gradual change time τt. The light output terminal executes Pt according to at least one of the color temperature control command, brightness control command, periodic dimming curve, and gradient control curve.
5. The method according to claim 1, characterized in that, The establishment of the intervention event Ik includes: calculating the parameter change DP of the current multidimensional light intervention parameter Pt relative to the previous group of multidimensional light intervention parameters P{t-1}; when DP is greater than the preset parameter change threshold D0, a new intervention event is established; otherwise, the current feedback window is assigned to the previous intervention event.
6. The method according to claim 1, characterized in that, The termination condition of the intervention event Ik includes at least one of the following: The light intervention is triggered when the parameter change DP exceeds the preset parameter change threshold D0, the event duration reaches the preset maximum event duration, the number of available candidate window labels within the event reaches the preset number of evaluation windows, the emotional state enters the target emotional range and is maintained for a preset number of windows, or the user actively terminates the light intervention.
7. The method according to claim 1, characterized in that, The feedback window data includes at least two of the following: heart rate variability change ΔHRVt, skin conductance change ΔEDAt, body movement change ΔAt, and improvement in the distance from the emotional state to the target emotional zone ΔDt; the target emotional zone is predetermined in the valence-arousal two-dimensional space according to the intervention target, and is limited by the lower limit of valence, the upper limit of valence, the lower limit of arousal, and the upper limit of arousal. The processor calculates the window intervention score St based on at least two of the above, and generates valid, partially valid, or invalid candidate window labels yt based on the comparison result between St and the preset label threshold.
8. The method according to claim 7, characterized in that, Before generating candidate window labels yt, the processor determines the window availability identifier Qt based on the physiological signal quality score SQt, body movement intensity At, and change in ambient illuminance. When Qt is unavailable, the corresponding candidate window label yt is determined to be unavailable, and the window is prohibited from participating in parameter mapping relationship updates as a valid learning sample.
9. The method according to claim 1, characterized in that, The event-level merging includes: removing unusable candidate window labels, counting the number of valid labels ne, the number of partially valid labels np and the number of invalid labels ni, and calculating label consistency Rk=max(ne,np,ni) / (ne+np+ni); When Rk is greater than or equal to the preset consistency threshold, candidate event-level labels are determined; otherwise, Yk is determined to be uncertain. The candidate window labels are then time-weighted and merged according to the time weight βt to determine the final Yk.
10. The method according to claim 1, characterized in that, The parameter mapping relationship Mt includes a priority table for candidate parameter combinations; When Yk is valid, partially valid, or invalid, the processor increases, slightly increases, or decreases the priority weight W of Pk in an emotional state similar to Ek, respectively. When Yk is unavailable or uncertain, W is not updated. When multiple consecutive Yk values from different intervention events are invalid and the preset invalidity threshold is reached, switch to the conservative parameter set.
