A method, device, and medium for predicting thermal insulation performance based on machine learning
By reconstructing temperature response information through thermal boundary stripping and three-path thermal response relationship, thermal bridge disturbance sites are identified and sparse corrections are performed. This solves the problems of boundary loading influence and local structural heat transfer anomalies in thermal insulation performance prediction, and improves the stability and reliability of prediction results.
Patent Information
- Application Number
- CN202611130284.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-28
- Publication Date
- 2026-11-17
AI Technical Summary
Existing methods for predicting thermal insulation performance have difficulty distinguishing between the effects of boundary loading and local structural heat transfer anomalies, resulting in insufficient stability of the prediction results.
By acquiring material structure information, test boundary information, and temperature response information of thermal insulation components, a thermal test sample set is generated. The temperature response information is reconstructed through thermal boundary stripping and three-path thermal response relationship, thermal bridge perturbation sites are determined, sparse correction is performed, and thermal insulation prediction features are generated. Finally, a machine learning model is used to predict thermal insulation performance.
This improved the stability and reliability of thermal insulation performance prediction results, clarified the structural origin of thermal bridge disturbance sites, and enhanced the interpretability of prediction results.
Smart Images

Figure CN123115606A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of thermal insulation prediction technology, and in particular to a method, device and medium for predicting thermal insulation performance based on machine learning. Background Technology
[0002] With the increasing demand for building energy efficiency assessment and industrial thermal insulation performance testing, the characterization of thermal insulation components is gradually shifting from steady-state heat transfer parameter calculation to dynamic temperature response analysis. Multilayer material structures, partially connected structures, and heterogeneous infill structures generate complex hot-end loading responses, thickness-side hysteresis responses, and cold-end transfer responses during heat transfer, resulting in significant spatiotemporal coupling characteristics in thermal insulation performance evaluation. In recent years, methods such as infrared thermography, heat flow testing, finite element heat conduction modeling, and machine learning prediction have been increasingly applied to thermal insulation performance assessment. By collecting material structural parameters and temperature response curves, a nonlinear mapping relationship between the component's structural state and thermal insulation performance indicators is established. Especially in complex thermal insulation components and multi-condition thermal testing scenarios, machine learning methods have become an important research direction for thermal performance prediction and sample pattern extraction.
[0003] However, existing methods for predicting thermal insulation performance typically use temperature response information and material structure information directly to train machine learning models. However, the response offset generated by the test boundary conditions is often mixed with thermal bridge disturbances inside the material structure in the same temperature response curve, making it difficult for the model input features to distinguish between the effects of boundary loading and local structural heat transfer anomalies. At the same time, the thermal bridge residuals formed at joint locations, fastener locations, and edge compression locations lack clear structural backtracking constraints, which can easily lead to insufficient stability of the prediction features, thereby affecting the consistency of the thermal insulation performance prediction results. Summary of the Invention
[0004] In view of the aforementioned existing problems, the present invention is proposed.
[0005] Therefore, this invention provides a machine learning-based method for predicting thermal insulation performance to solve the problem of difficulty in distinguishing the effects of boundary loading and local structural heat transfer anomalies.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for predicting thermal insulation performance based on machine learning, comprising, Obtain material structure information, test boundary information, and temperature response information of thermal insulation components, and group them according to the same component number and temperature measurement location to generate a thermal test sample set; Thermal boundary stripping is performed on the thermal test sample set to separate the response offset caused by the test boundary information from the temperature response information, and the temperature response information is reconstructed according to the three-path thermal response relationship to generate thermal insulation response characteristics. The thermal insulation response features are embedded into the material structure information. The thermal bridge perturbation sites are determined based on the back projection positions of the response residuals in the thermal insulation response features into the material structure information. The response segments occupied by the thermal bridge perturbation sites are then sparsely corrected to generate thermal insulation prediction features. A machine learning model is trained using a thermal test sample set, and the model is used to predict the thermal insulation performance based on the thermal insulation prediction features. The predicted thermal insulation performance value is then corrected using the training coverage boundary of the thermal bridge perturbation site to generate the predicted thermal insulation performance result.
[0007] As a preferred embodiment of the machine learning-based thermal insulation performance prediction method of the present invention, the component number is a unique identifier used to distinguish different thermal insulation components and to group the material structure information, test boundary information and temperature response information of the same thermal insulation component into the same thermal test sample set. The temperature measurement position is a fixed spatial position on the thermal insulation component used to collect temperature response information, so as to limit the merging range of test boundary information and temperature response information in the thermal test sample set at the same position; The three-path thermal response relationship is the response arrangement relationship formed by the hot end loading response, the thickness hysteresis response, and the cold end transfer response in the temperature response information according to the order of heat transfer.
[0008] As a preferred embodiment of the machine learning-based thermal insulation performance prediction method of the present invention, wherein: the generation of thermal insulation response features specifically includes, Within the thermal test sample set, a single sample range is divided according to the same component number and temperature measurement location. The test boundary information within a single sample range is mapped to the time occupancy location and temperature measurement occupancy location in the temperature response information to generate boundary occupancy segments. Embed the boundary occupancy segment into the temperature response information, separate the response offset caused by the test boundary information along the temperature response information, and generate the boundary stripping response segment. The boundary stripping response fragment is used to connect the three-channel thermal response relationship. The response bits of the boundary stripping response fragment are rearranged to generate the three-channel response bit sequence. The temperature response information is reconstructed based on the three-path response bit sequence. The reconstructed temperature response information is then normalized in terms of response scale and arranged according to the three-path thermal response relationship to generate thermal insulation response characteristics.
[0009] As a preferred embodiment of the machine learning-based thermal insulation performance prediction method of the present invention, the response scale normalization is to unify the amplitude range and align the time scale of the three-path response bit sequence in the reconstructed temperature response information to form normalized temperature response information.
[0010] As a preferred embodiment of the machine learning-based thermal insulation performance prediction method of the present invention, wherein: determining the thermal bridge disturbance location specifically involves, Set the component number and temperature measurement location in the material structure information as the structural index, merge the material hierarchy and local structural locations under the structural index to form a structural location item; The structural location items are connected in series according to the material hierarchy relationship in the structural location items, and adjacent structural channels are established in the local structural locations in the structural location items to generate a material structural channel chain. The thermal insulation response characteristics are written into the material structure channel chain, and the three-path response bit sequence in the thermal insulation response characteristics is attached to the adjacent structural channels under the same temperature measurement position to generate the structural load-bearing response chain. Based on the material hierarchy and adjacent structural channels in the structural load-bearing response chain, the response features that do not continuously shift with the three-path response sequence are extracted from the thermal insulation response features, and the response residual fragments are output. The three-path thermal response relationship carried by the response residual fragment is transformed into a back-projection gating relationship, and the response residual fragment is projected onto the material structure channel chain according to the back-projection gating relationship to generate a phase-gated back-projection fragment. The phase-gated backprojection segment is subjected to structural channel competition and repositioning. The backprojection position in the phase-gated backprojection segment that simultaneously bears the disconnection of material hierarchy relationship and the abrupt change of adjacent structural channel is determined as the thermal bridge disturbance position.
[0011] As a preferred embodiment of the machine learning-based thermal insulation performance prediction method of the present invention, wherein: the generation of thermal insulation prediction features specifically includes, The thermal bridge disturbance is returned to the structural load-bearing response chain, and the three-path response range occupied by the thermal bridge disturbance is delineated in the structural load-bearing response chain as the disturbance response band. The response residual fragments are embedded into the perturbation response band, and the correction range of the perturbation response band is narrowed according to the position occupied by the response residual fragments in the perturbation response band to generate sparse correction slots. Based on the material hierarchy relationship and adjacent structural channels of the sparse correction slot in the material structure channel chain, the response residual fragment in the sparse correction slot is subjected to graph-time expansion sparse compensation to obtain the graph-time compensation chain. The time-compensation chain is backfilled to the position of the response residual segment corresponding to the sparse correction slot, and the response segments in the disturbance response band located outside the sparse correction slot are retained in the original response position, and the domain-locked response is output. By updating the response fragment occupied by thermal bridge perturbation bits in the insulation response features through domain locking response, insulation prediction features are generated.
[0012] As a preferred embodiment of the machine learning-based thermal insulation performance prediction method of the present invention, the graph-time expansion sparsity compensation refers to expanding the response residual fragments in the sparse correction groove along the three-path response bit sequence according to the material hierarchy relationship and adjacent structural channels in the material structure channel chain, and converting the expanded response residual fragments into a graph-time compensation chain.
[0013] As a preferred embodiment of the machine learning-based thermal insulation performance prediction method of the present invention, wherein: the generation of thermal insulation performance prediction results specifically includes, A machine learning model is trained using a thermal test sample set, and the sample prediction bias generated during the training process is fed back to the thermal bridge perturbation position to generate a perturbation position bias sequence. Arrange the sample prediction biases belonging to the same hot bridge perturbation position in the perturbation position bias sequence according to the deviation direction, define the deviation interval continuously occupied by the sample prediction bias, and close the boundary position of the deviation interval to generate the training coverage boundary. The thermal insulation performance is predicted by using a machine learning model to predict the thermal insulation performance of the thermal insulation prediction features. The thermal bridge perturbation sites corresponding to the thermal insulation prediction features are then assigned to the training coverage boundary to obtain the prediction boundary assignment result. Based on the predicted boundary alignment results, the predicted thermal insulation performance value is aligned with the deviation interval in the training coverage boundary to determine the boundary correction segment of the predicted thermal insulation performance value. The predicted thermal insulation performance value is then corrected using the boundary correction segment to generate the predicted thermal insulation performance result.
[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the machine learning-based thermal insulation performance prediction method described in the first aspect of the present invention.
[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the machine learning-based thermal insulation performance prediction method described in the first aspect of the present invention.
[0016] The beneficial effects of this invention are as follows: Through back-projection positioning and time-of-graph sparsity compensation, structured correction of thermal bridge disturbance sites is achieved. After embedding material structure information into the thermal insulation response features, the response residuals can fall back to the thermal bridge disturbance sites along the material hierarchy and adjacent structural channels, giving local anomalies in the temperature response a clear structural origin, rather than merely manifesting as curve amplitude deviations. The disturbance response band can limit the three-path response range occupied by the thermal bridge disturbance sites, the sparse correction slots can narrow the local correction objects, and the time-of-graph compensation chain can convert response residual fragments into backfillable compensation expressions, thereby preserving the continuous relationship between the hot-end loading response, the thickness hysteresis response, and the cold-end transfer response. The machine learning model performs predictions based on the structured-corrected thermal insulation prediction features, improving the stability, reliability, and interpretability of the thermal insulation performance prediction results. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a flowchart of a machine learning-based method for predicting thermal insulation performance.
[0019] Figure 2 A flowchart for generating thermal insulation response characteristics.
[0020] Figure 3 The flowchart for determining the thermal bridge disturbance position.
[0021] Figure 4 A flowchart for generating thermal insulation prediction features.
[0022] Figure 5 This is a comparison chart of integral prediction errors for cold junction temperature rise.
[0023] Figure 6 The graph shows a comparison of the stability of sparse compensation prediction. Detailed Implementation
[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0026] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0027] Reference Figures 1-6 This is one embodiment of the present invention, which provides a machine learning-based method for predicting thermal insulation performance, including the following steps: S1. Obtain the material structure information, test boundary information, and temperature response information of the thermal insulation components, and merge them according to the same component number and temperature measurement location to generate a thermal test sample set.
[0028] S1.1. The material layer relationships, material layer thickness, material density, porosity, joint location, fastener location, and edge compression location of the thermal insulation components are compiled into material structure information. The material layer relationships are derived from material drawings, the joint locations and fastener locations are derived from production batch records and manual positioning results, and the material layer thickness, material density, porosity, and edge compression location are derived from thickness measurement results, density measurement results, porosity detection results, and edge compression measurement results.
[0029] The hot end temperature, cold end temperature, ambient temperature, heating time, component thickness, and heat flow direction are organized into test boundary information. The hot end temperature and cold end temperature are derived from temperature control results, the ambient temperature is derived from ambient temperature acquisition results, the heating time is derived from timing results, the component thickness is derived from thickness measurement results, and the heat flow direction is derived from manual calibration results.
[0030] The hot-end temperature curve from the hot-end temperature measurement results, the cold-end temperature curve from the cold-end temperature measurement results, and the thickness temperature curve from the thickness temperature measurement point acquisition results are organized into temperature response information according to the acquisition time, measurement location, and temperature value.
[0031] Using the component number field and temperature measurement location field as merging indexes, material structure information, test boundary information, and temperature response information are grouped into sample items under the same component number and the same temperature measurement location. When the component number field and temperature measurement location field are consistent, and the acquisition time in the temperature response information falls within the test start and end time range indicated by the timing results, the sample item is marked as a valid sample item. When the component number field is inconsistent, the temperature measurement location field is inconsistent, the component number field is missing, the temperature measurement location field is missing, or the acquisition time in the temperature response information does not fall within the test start and end time range indicated by the timing results, a merging anomaly mark is retained in the sample item. The valid sample items and the sample items with retained merging anomaly marks are summarized to generate a thermal test sample set.
[0032] Among them, the component number is a unique identifier used to distinguish different thermal insulation components and to merge the material structure information, test boundary information and temperature response information of the same thermal insulation component into the same thermal test sample set; the temperature measurement position is a fixed spatial location on the thermal insulation component used to collect temperature response information, so as to limit the merging range of test boundary information and temperature response information in the thermal test sample set at the same position.
[0033] S2. Perform thermal boundary stripping on the thermal test sample set, separate the response offset caused by the test boundary information from the temperature response information, and reconstruct the temperature response information according to the three-path thermal response relationship to generate thermal insulation response features.
[0034] S2.1 Within the thermal test sample set, divide the sample range into individual sample ranges based on the same component number and temperature measurement location. Map the test boundary information within each sample range to the time occupancy location and temperature measurement occupancy location in the temperature response information to generate boundary occupancy segments. Specifically, Valid sample items are read from the thermal test sample set. According to the component number field and the temperature measurement location field, the material structure information, test boundary information, and temperature response information of the same component number and the same temperature measurement location are classified into a single sample range. The test start and end time range indicated by the heating duration in the test boundary information is pushed into the acquisition time arrangement position of the temperature response information. The hot end temperature is written into the temperature measurement location arrangement position corresponding to the hot end temperature curve as a reference for the hot end loading amplitude. The cold end temperature is written into the temperature measurement location arrangement position corresponding to the cold end temperature curve as a cold end constraint reference. The ambient temperature is written into the temperature measurement location arrangement position corresponding to the thickness temperature curve as a reference for the external temperature effect. The component thickness is written into the thickness temperature curve side under the same temperature measurement location as the thickness temperature measurement span. The heat flow direction is written into the thickness temperature curve side under the same temperature measurement location as the arrangement direction identifier of the hot end loading response, thickness hysteresis response, and cold end transmission response.
[0035] Determine whether the acquisition time in the temperature response information is within the test start and end time range indicated by the timing results. Check whether the temperature measurement location fields in the material structure information, test boundary information, and temperature response information are consistent. Verify whether the hot end temperature curve, cold end temperature curve, and thickness temperature curve are complete. If the acquisition time is within the test start and end time range, the temperature measurement location fields are consistent, the hot end temperature curve is complete, the cold end temperature curve is complete, and the thickness temperature curve is complete, then combine the acquisition time period and temperature measurement location segment occupied by the test boundary information in the temperature response information into a boundary occupancy segment. If any of the following occurs: the acquisition time is not within the test start and end time range, the temperature measurement location fields are inconsistent, the hot end temperature curve is missing, the cold end temperature curve is missing, or the thickness temperature curve is missing, then write a boundary mapping anomaly mark in the single sample range and stop generating boundary occupancy segments in the single sample range.
[0036] Among them, the single sample range is the merged range of material structure information, test boundary information and temperature response information defined by the same component number and the same temperature measurement location in the thermal test sample set; the boundary mapping anomaly mark is the anomaly identification mark retained in the single sample range when the test boundary information cannot accurately fall into the acquisition time period and temperature measurement location period of the temperature response information.
[0037] S2.2. Embed the boundary occupancy segment into the temperature response information, separate the response offset caused by the test boundary information along the temperature response information, and generate a boundary stripping response segment. Specifically, The boundary occupancy segment is embedded into the temperature response information within the same single sample range. The stripping window is defined by the acquisition time period and temperature measurement location segment in the boundary occupancy segment, so that the hot end temperature curve, cold end temperature curve, and thickness temperature curve form a boundary occupancy curve segment within the stripping window. Within the stripping window, a boundary effect reference is established by combining the loading amplitude reference of the hot end temperature, the constraint benchmark of the cold end temperature, the external temperature effect reference of the ambient temperature, the test start and end boundaries of the heating time, and the arrangement direction of the heat flow direction. The initial slope, peak position, stable segment position, and temperature value change segment in the boundary occupancy curve segment are read along the acquisition time, and the consistency is verified with the continuous trend of the adjacent curve segment at the front end and the adjacent curve segment at the back end of the stripping window, and the position of the response shape offset caused by the test boundary information is calibrated.
[0038] If the response shape offset position falls within the stripping window, the response coordinates of the boundary-occupied curve segment are recalibrated according to the boundary action reference. The amplitude starting point of the hot-end loading response, the time position of the thickness hysteresis response, and the stable position of the cold-end transmission response are adjusted to the same boundary reference, and the recalibrated hot-end loading response, thickness hysteresis response, and cold-end transmission response are retained. If the response shape offset position deviates from the stripping window, the boundary stripping anomaly mark is retained in a single sample range. The hot-end loading response, thickness hysteresis response, and cold-end transmission response after the response coordinate recalibration are combined into a boundary stripping response segment.
[0039] The stripping window is defined by the acquisition time period and temperature measurement location segment in the boundary occupancy segment, and is used to locate and separate the temperature response information range that causes the response offset due to the test boundary information. The boundary occupancy curve segment is formed by extracting the curve portion that falls within the defined range from the hot end temperature curve, cold end temperature curve and thickness temperature curve after the acquisition time period and temperature measurement location segment are defined by the stripping window.
[0040] S2.3. Using the boundary stripping response fragment to connect the three-path thermal response relationship, rearrange the response bits of the boundary stripping response fragment to generate the three-path response bit sequence, specifically as follows: The hot-end loading response in the boundary peeling response segment is used as the front response position of the three-path thermal response relationship, the thickness hysteresis response in the boundary peeling response segment is used as the middle response position of the three-path thermal response relationship, and the cold-end transfer response in the boundary peeling response segment is used as the end response position of the three-path thermal response relationship.
[0041] The start time, peak time, settling time, and temperature change segment of the hot-end loading response, thickness hysteresis response, and cold-end transfer response at the same temperature measurement location are arranged along the heat flow direction and acquisition time sequence. The hot-end loading response is arranged on the heat input side, the thickness hysteresis response is arranged on the thickness transfer side, and the cold-end transfer response is arranged on the heat output side. When the acquisition times of the hot-end loading response, thickness hysteresis response, and cold-end transfer response overlap, the response positions are retained according to the heat transfer order of hot-end loading response first, thickness hysteresis response in the middle, and cold-end transfer response last.
[0042] When there are broken segments in the hot-end loading response, the thickness hysteresis response, and the cold-end transfer response, the response position break mark is retained at the response position where the broken segment is located; the hot-end loading response, the thickness hysteresis response, and the cold-end transfer response with the completed response position order are concatenated to generate a three-path response position sequence.
[0043] Among them, the three-path thermal response relationship is the response arrangement relationship formed by the hot end loading response, the thickness hysteresis response and the cold end transfer response in the temperature response information according to the order of heat transfer.
[0044] The response position rearrangement is based on the heat flow direction, acquisition time sequence, and response event location. The heat flow direction defines the spatial order of the hot-end loading response, thickness-induced hysteresis response, and cold-end transfer response from the heat input side to the heat output side. The acquisition time sequence defines the order of start time, peak time, and settling time at the same temperature measurement location. The response event location defines the order of the hot-end loading response entering the front response position, the thickness-induced hysteresis response entering the middle response position, and the cold-end transfer response entering the final response position. When the heat flow direction, acquisition time sequence, and response event location are consistent, the response position rearrangement is completed according to the order of the hot-end loading response, thickness-induced hysteresis response, and cold-end transfer response. When acquisition times overlap, the heat flow direction determines the response position affiliation, and the response event location corrects the succession order between different response positions.
[0045] S2.4. Based on the three-path response bit sequence, reconstruct the temperature response information, normalize the response scale of the reconstructed temperature response information, and arrange the normalized temperature response information according to the three-path thermal response relationship to generate thermal insulation response characteristics. Specifically, Using the front-end response bit, middle response bit, and end response bit in the three-channel response bit sequence as the arrangement skeleton, the start time of the hot-end loading response is first determined. Then, the acquisition time period adjacent to the hot-end loading response in the thickness hysteresis response is continued after the front-end response bit, and the acquisition time period adjacent to the thickness hysteresis response in the cold-end transfer response is continued after the middle response bit. When there is a time interval between adjacent acquisition time periods, conformal segmented cubic Hermite interpolation is used to complete the transition curve segment. The temperature value and temperature change slope of the acquisition point at the front of the time interval and the temperature value and temperature change slope of the acquisition point at the back of the time interval are used as interpolation constraints to generate the transition curve segment connecting the front-end response bit and the back-end response bit. When the temperature change slope is missing at the front-end acquisition point and the back-end acquisition point, linear interpolation between the temperature values at both ends is used to generate the transition curve segment. When the direction of temperature change before and after adjacent acquisition time periods cannot be determined, the transition curve segment completion is not performed, and the reconstruction break mark is retained. After completing the time continuation of the three-channel response bit sequence and the transition curve segment completion, the reconstructed temperature response information is formed.
[0046] In the reconstructed temperature response information, the start time of the hot-end loading response is used as the time reference, and the acquisition time of the thickness hysteresis response and the cold-end transfer response is converted into relative time positions. The hot-end loading response, thickness hysteresis response and cold-end transfer response in the reconstructed temperature response information are unified into the same amplitude representation interval. The lowest temperature value in the same single sample range is used as the amplitude start point, and the highest temperature value in the same single sample range is used as the amplitude end point. The temperature values at different response positions are converted into the same amplitude scale, and the acquisition time at different response positions is aligned to the same time scale to form normalized temperature response information.
[0047] The normalized temperature response information is arranged according to the three-path thermal response relationship of hot-end loading response, thickness hysteresis response and cold-end transfer response, so that the normalized temperature response information retains the sequential relationship between the heat input side, the thickness transfer side and the heat output side, and generates thermal insulation response characteristics.
[0048] S3. Embed the thermal insulation response features into the material structure information, determine the thermal bridge perturbation sites based on the back projection position of the response residuals in the thermal insulation response features into the material structure information, and perform sparse correction on the response segments occupied by the thermal bridge perturbation sites to generate thermal insulation prediction features.
[0049] S3.1. Set the component number and temperature measurement location in the material structure information as the structural index, merge the material hierarchy and local structural locations under the structural index to form a structural location item, specifically, Using the component number and temperature measurement location in the material structure information as the structural index entry, the material hierarchy of the insulation components is defined under the same component number, and the joint location, fastener location, and edge compression location are collected under the same temperature measurement location. The material hierarchy is used to indicate the interlayer arrangement order of the insulation components from the heat input side to the heat output side, and the joint location, fastener location, and edge compression location are used to indicate the local structural locations adjacent to the same temperature measurement location. The material hierarchy and local structural locations that have been collected by the structural index are combined into a structural location item.
[0050] S3.2. Connect the structural position items in series according to the material hierarchy relationship in the structural position items, and establish adjacent structural channels at local structural positions in the structural position items to generate a material structural channel chain. Specifically, Arrange the structural position items according to the material hierarchy in the structural position items, so that the structural position items closer to the heat input side are at the front end and the structural position items closer to the heat output side are at the end end; locate the joint positions between adjacent material layers, locate the fastener positions and edge compression positions at the same temperature measurement position, and use the joint positions as interlayer transition positions, the fastener positions as thickness penetration positions, and the edge compression positions as edge disturbance positions; connect adjacent structural position items along the interlayer transition positions, thickness penetration positions, and edge disturbance positions to generate adjacent structural channels; the arranged structural position items and adjacent structural channels form a material structural channel chain.
[0051] S3.3. Write the thermal insulation response characteristics into the material structure channel chain, and connect the three-path response bit sequence in the thermal insulation response characteristics to adjacent structural channels under the same temperature measurement position to generate a structural load-bearing response chain. Specifically, The material structure channel chain is used as the structural load-bearing path for the thermal insulation response characteristics. Structural position items with the same component number and temperature measurement location as the thermal insulation response characteristics are selected within the material structure channel chain. After the three-path response sequence enters the material structure channel chain, the front response position occupies the structural position item on the heat input side, the middle response position covers the interlayer transition position, the thickness penetration position, and the edge disturbance position along the adjacent structural channel, and the end response position occupies the structural position item on the heat output side. The hot-end loading response falls into the structural position item on the heat input side with the front response position, the thickness hysteresis response falls into the adjacent structural channel with the middle response position, and the cold-end transmission response falls into the structural position item on the heat output side with the end response position. The material hierarchy relationship retains the structural arrangement from the heat input side to the heat output side, and the three-path thermal response relationship retains the transmission arrangement of the hot-end loading response, the thickness hysteresis response, and the cold-end transmission response. These two are combined at the same temperature measurement location to form a structural load-bearing response chain.
[0052] S3.4. Based on the material hierarchy and adjacent structural channels in the structural load-bearing response chain, extract the response features in the thermal insulation response characteristics that do not continuously shift with the three-path response sequence, and output the response residual fragments, specifically, Along the material hierarchy of the structural load-bearing response chain, the positional movement of the three-path response position sequence is traced from the structural position item on the heat input side to the structural position item on the heat output side. The segment connecting the front response position to the middle response position and then to the end response position is used as the transition reference. When adjacent structural channels pass through interlayer transition positions, thickness penetration positions, and edge disturbance positions, the temperature value change segment in the thermal insulation response feature under the same acquisition time period is read and compared with the response position in the transition reference. The part of the temperature value change segment that can follow the front response position, middle response position, and end response position in sequence is retained as the continuous transition response. The part of the temperature value change segment that stays at the material layer junction position, the transition position of adjacent structural channels, or the area adjacent to the local structural position and does not continue to move with the response position is regarded as the discontinuous transition response feature. The interception boundary is defined by the start time and end time of the discontinuous transition response feature, and the temperature measurement position and response position of the discontinuous transition response feature are retained, and the response residual segment is output.
[0053] Among them, the response residual fragment is used to characterize the dynamic response anomalies in the thermal insulation response characteristics that fail to be continuously transmitted along the material hierarchy and adjacent structural channels. The response residual fragment is not directly equivalent to the thermal bridge disturbance position. The joint position, fastener position, edge compression position and pore state change position form potential disturbance candidate positions in the material structure channel chain. The response residual fragment is used to indicate the dynamic occupancy range of the potential disturbance candidate positions that has already generated response deviation in the current thermal test sample.
[0054] S3.5. Convert the three-path thermal response relationship carried by the response residual fragment into a back-projection gating relationship, and project the response residual fragment onto the material structure channel chain according to the back-projection gating relationship to generate a phase-gated back-projection fragment. Specifically, The front-end response bit, intermediate response bit, and final response bit carried by the response residual fragment are associated with the structural position item on the heat input side, the adjacent structural channel, and the structural position item on the heat output side, respectively. The back-projection time gate is defined by the start and end acquisition times of the response residual fragment, the back-projection position gate is defined by the temperature measurement position of the response residual fragment, and the back-projection path gate is defined by the response position of the response residual fragment. The back-projection time gate, the back-projection position gate, and the back-projection path gate are combined to form a back-projection gating relationship. When the response residual fragment returns to the material structural channel chain along the back-projection gating relationship, the back-projection time gate first defines the projectable acquisition time period, the back-projection position gate defines the structural position item under the same temperature measurement position, and the back-projection path gate guides the front-end response bit to the structural position item on the heat input side, the intermediate response bit to the adjacent structural channel, and the final response bit to the structural position item on the heat output side. After the projection limitation is completed, the response residual fragment retains the acquisition time period, temperature measurement position, response position, and temperature value change segment to generate a phase-gated back-projection fragment.
[0055] S3.6. Perform structural channel competition repositioning on the phase-gated backprojection segments. The backprojection positions within the phase-gated backprojection segments that simultaneously bear the characteristics of material hierarchy breakage and abrupt changes in adjacent structural channels are identified as thermal bridge disturbance sites. Specifically, In the material structure channel chain, candidate relocation positions capable of receiving the projection of response residual segments are delineated. When a candidate relocation position enters the material hierarchy relationship verification, the material hierarchy arrangement, material layer thickness variation, and joint layer crossing of the candidate relocation position are read. When at least one of the following occurs—disruption of material hierarchy arrangement, abrupt change in material layer thickness, and joint layer crossing—the candidate relocation position is marked as a material hierarchy relationship break position. When a candidate relocation position enters the adjacent structural channel verification, the fastener position, edge compression position, and pore state change of the candidate relocation position are read. When at least one of the following occurs—fastener position crossing the adjacent structural channel, edge compression position intruding into the adjacent structural channel, and abrupt change in pore state—the candidate relocation position is marked as an adjacent structural channel abrupt change position. When a candidate relocation position simultaneously possesses the relocation attributes of both a material hierarchy relationship break position and an adjacent structural channel abrupt change position, the candidate relocation position is used as a back projection position, and the back projection position is determined as a thermal bridge disturbance position.
[0056] Among them, the material hierarchy relationship is broken when there is an interruption in the hierarchical arrangement between adjacent material layers in the material structure channel chain, a sudden change in material layer thickness, or a cross-layer joint position, which causes the three-path response position sequence to not shift continuously according to the material hierarchy relationship; the adjacent structural channel abrupt change refers to the occurrence of fastener penetration, edge compression intrusion, or pore state abrupt change in adjacent structural channels, which causes the response residual fragment to be discontinuously repositioned in adjacent structural channels.
[0057] It should be noted that thermal bridge disturbance sites are not determined by inferring unknown structural causes solely from response residual fragments, but rather by locating and screening potential disturbance candidate locations already provided by the material structure information. Response residual fragments provide dynamic response anomalies in the current thermal test sample, while the material structure channel chain provides a range of structural candidates formed by joint locations, fastener locations, edge compression locations, and locations of pore state changes. When a candidate location simultaneously accommodates the return placement of a response residual fragment, a break in the material hierarchy relationship, and a sudden change in adjacent structural channels, it is identified as a thermal bridge disturbance site in the current thermal test sample. When only a structural candidate location exists but does not accommodate a response residual fragment, it is uncertain whether it is a thermal bridge disturbance site in the current thermal test sample.
[0058] S3.7. The thermal bridge disturbance is returned to the structural load-bearing response chain. Within the structural load-bearing response chain, the three-path response range occupied by the thermal bridge disturbance is delineated as the disturbance response band. Specifically, After the thermal bridge disturbance position is determined, the structural position item is returned to the response bearing position under the same component number and the same temperature measurement position in the structural bearing response chain. Around the response bearing position where the thermal bridge disturbance position is located, first delineate the response segment in the hot end loading response that connects the material layer in front of the thermal bridge disturbance position, then delineate the response segment in the thickness hysteresis response that passes through the adjacent structural channel, and finally delineate the response segment in the cold end transmission response that connects the material layer in back of the thermal bridge disturbance position. Connect the hot end loading response segment, the thickness hysteresis response segment, and the cold end transmission response segment in the order of front-end response position, middle response position, and end response position to form the three-path response range occupied by the thermal bridge disturbance position. The three-path response range retains the structural position relationship in the material structure channel chain, including the acquisition time period, temperature measurement position, response position, and material structure channel chain, as a disturbance response band, and is used to support the embedding of subsequent response residual segments and the narrowing of sparse correction slots.
[0059] S3.8. Embed the response residual fragments into the perturbation response band, and narrow the correction range of the perturbation response band according to the occupied position of the response residual fragments in the perturbation response band to generate sparse correction slots. Specifically, With the disturbance response band already generated, the response residual fragments are mapped and aligned with the three-channel response range of the disturbance response band in terms of acquisition time period, temperature measurement location, and response position. The temperature value change segments that fall into the area occupied by the response residual fragments in the hot-end loading response, thickness hysteresis response, and cold-end transfer response are isolated to form a local correction zone. The response fragments in the disturbance response band that do not fall into the local correction zone remain in place and do not participate in the narrowing operation. The sparse correction slot is defined by the time period, temperature measurement location, and response position of the local correction zone. The sparse correction slot is used to limit the scope of sparse compensation when the subsequent graph is generated, ensuring that only the area occupied by the response residual fragment is compensated, and carrying over to the next operation of the compensation chain when the graph is generated.
[0060] The correction range of the disturbance response band is the local response region in the disturbance response band that coincides with the response residual segment in terms of acquisition time period, temperature measurement location and response position, and needs to be compensated by entering the sparse correction slot.
[0061] S3.9. Based on the material hierarchy relationship and adjacent structural channels of the sparse correction slot in the material structure channel chain, perform graph-time expansion sparse compensation on the response residual fragments within the sparse correction slot to obtain the graph-time compensation chain, specifically as follows: Within the limited acquisition time, temperature measurement location, and response position of the sparse correction groove, the material hierarchy occupied by the sparse correction groove is read from the material structure channel chain. The interlayer transition position, thickness penetration position, and edge perturbation position traversed by the response residual fragment are determined along adjacent structural channels. Using the three-path response position sequence as the expansion order, the hot-end loading response portion of the response residual fragment is extended to the front response position, the thickness hysteresis response portion is extended to the middle response position, and the cold-end transfer response portion is extended to the end response position, thus extending the response residual fragment from the local residual segment within the sparse correction groove. The time-to-image compensation segment is expanded to cover the material hierarchy and adjacent structural channels. Then, according to the connection order of adjacent structural channels, the time-to-image compensation segment is divided into compensation nodes. Each compensation node retains the acquisition time period, response position, interlayer transition position, thickness penetration position, edge disturbance position, and temperature value change segment. The compensation nodes under the same response position are arranged in the order of acquisition time period, and the compensation nodes in adjacent material layers are connected according to the material hierarchy to form a time-to-image compensation chain that has both time arrangement relationship and interlayer connection relationship. The time-to-image compensation chain is used to subsequently backfill to the response residual segment position corresponding to the sparse correction groove.
[0062] S3.10. Backfill the time-compensation chain in the graph to the position of the response residual segment corresponding to the sparse correction slot, and retain the response segments in the disturbance response band that are outside the sparse correction slot in the original response position, and output the domain-locked response, specifically as follows: Using the acquisition time period, temperature measurement location, and response position corresponding to the sparse correction groove as the basis for backfill positioning, the acquisition time period in the time-to-time compensation chain is aligned with the acquisition time period of the sparse correction groove, the interlayer position in the time-to-time compensation chain is aligned with the material hierarchy relationship of the sparse correction groove in the material structure channel chain, and the temperature value change segment in the time-to-time compensation chain covers the local residual segment corresponding to the response residual segment. Within the disturbance response band, the hot-end loading response segment, the thickness hysteresis response segment, and the cold-end transfer response segment located outside the sparse correction groove do not participate in the compensation coverage and maintain their original response positions. The backfilled time-to-time compensation chain and the response segments retained outside the groove are connected according to the acquisition time sequence and response position to form a locked domain response that is only compensated within the sparse correction groove and remains locked outside the sparse correction groove. The locked domain response is used to subsequently update the response segments occupied by the thermal bridge disturbance position in the thermal insulation response characteristics.
[0063] S3.11. Update the response fragment occupied by thermal bridge perturbation sites in the insulation response features by locking the domain response to generate insulation prediction features, specifically, After the locked domain response enters the thermal insulation response feature, the acquisition time period, temperature measurement location, and response position in the locked domain response are aligned with the response segments occupied by thermal bridge disturbances to determine the response range to be updated in the thermal insulation response feature. Within the sparse correction slots within the response range to be updated, the position of the response residual segment is taken over by the time-to-time compensation chain. For response segments outside the sparse correction slots within the response range to be updated, the original response position is retained, so that the response segments occupied by thermal bridge disturbances are locally updated. For thermal insulation response features that do not fall into the response range to be updated, the original arrangement of the hot-end loading response, the thickness hysteresis response, and the cold-end transfer response is retained. The thermal insulation response features after local updates are used as thermal insulation prediction features for subsequent machine learning models to predict thermal insulation performance.
[0064] Among them, the response range to be updated is the range of response segments in the thermal insulation response characteristics that overlap with the domain-locked response in terms of acquisition time period, temperature measurement location, and response position, and is occupied by thermal bridge disturbance positions.
[0065] It should be noted that, as Figure 6 As shown, with the increase of thermal bridge disturbance intensity, the mean absolute error of the cold end temperature rise integral prediction for the non-sparse compensation feature, the full-disturbance response band compensation feature, and the non-locked domain retention compensation feature all show an upward trend. In the figure of this invention, the sparse compensation feature is generally in a low error range in the main comparison interval. This result shows that the sparse correction slot can limit the compensation range to the local response segment occupied by the thermal bridge disturbance position, and the locked domain response can retain the original response position outside the sparse correction slot, avoiding the drift of the non-disturbance response caused by full-band compensation, thereby improving the prediction stability of the thermal insulation prediction feature under different thermal bridge disturbance intensities.
[0066] S3.12. During the process of the thermal insulation prediction feature being transferred from S3.11 to S4.1 and then to S4.4, the thermal insulation prediction feature, thermal bridge perturbation position, locked domain response, response range to be updated, perturbation position deviation sequence and training coverage boundary are linked and interacted in a chain.
[0067] In the pre-judgment stage, the thermal insulation prediction feature is used as the admission object, and the thermal bridge perturbation position, the locked domain response, and the response range to be updated are used as the judgment objects. The overlap relationship between the collection time period, temperature measurement position, and response position is used to verify whether the thermal bridge perturbation position falls within the response range to be updated, and whether the locked domain response completely takes over the response segment occupied by the thermal bridge perturbation position. When the thermal bridge perturbation position falls within the response range to be updated and the locked domain response completely takes over the response segment occupied by the thermal bridge perturbation position, the thermal insulation prediction feature is marked as the training admission state, and the position value of the thermal bridge perturbation position and the compensation value of the locked domain response are passed to the input layer of the machine learning model. When the thermal bridge perturbation position does not fall within the response range to be updated, the thermal bridge perturbation position is returned to S3.8 to re-limit the correction range of the perturbation response band. When the locked domain response does not completely take over the response segment occupied by the thermal bridge perturbation position, the locked domain response is returned to S3.10 to re-align the compensation chain and sparse correction slot when the graph is aligned.
[0068] In the case-by-case processing phase, the thermal insulation prediction features in the training admission state are used as training inputs in S4.1. The machine learning model processes the thermal insulation prediction features in the order of the input layer, the first fully connected layer, the second fully connected layer, and the output layer, and outputs sample prediction values. The sample prediction values are compared with the measured values of thermal insulation performance to form sample prediction bias. The sample prediction bias carries the component number, temperature measurement location, thermal bridge perturbation position, and training batch sequence into the perturbation position bias sequence. When the sample prediction biases belong to the same thermal bridge perturbation position and have the same deviation direction, the sample prediction biases are arranged in the order of the training batch. When the sample prediction biases belong to the same thermal bridge perturbation position but have different deviation directions, the sample prediction biases are arranged as positive sample prediction bias, negative sample prediction bias, and zero bias. When the sample prediction biases do not belong to the same thermal bridge perturbation position, the sample prediction biases are re-aggregated to the corresponding thermal bridge perturbation positions according to the component number and temperature measurement location.
[0069] In the subsequent call phase, the perturbation position deviation sequence formed in the case-by-case processing phase is passed to S4.2. S4.2 then delineates continuous deviation segments according to the deviation direction and training batch order under the same thermal bridge perturbation position, and the continuous deviation segments form a deviation interval. After the training coverage boundary is generated, the lower boundary, upper boundary, minimum deviation value, and maximum deviation value of the interval are passed to S4.3. This allows S4.3 to output the thermal insulation performance prediction value while incorporating the thermal bridge perturbation position carried by the thermal insulation prediction feature into the matched training coverage boundary, forming the prediction boundary repositioning result.
[0070] During the result verification stage, the thermal bridge perturbation position in the predicted boundary repositioning result is repositioned to match the training coverage boundary. The predicted insulation performance value is then compared with the matched deviation interval to verify the deviation status. When the thermal bridge perturbation position falls within the training coverage boundary and the predicted insulation performance value corresponds to the positive sample prediction deviation interval, the predicted boundary repositioning result enters the positive boundary correction path. When the thermal bridge perturbation position falls within the training coverage boundary and the predicted insulation performance value corresponds to the negative sample prediction deviation interval, the predicted boundary repositioning result enters the negative boundary correction path. When the thermal bridge perturbation position falls within the training coverage boundary and the predicted insulation performance value corresponds to the zero deviation position, the predicted boundary repositioning result enters the retention path. When the thermal bridge perturbation position does not fall within the training coverage boundary, the perturbation position deviation sequence under the same thermal bridge perturbation position is returned to S4.2 to redefine the deviation interval and regenerate the training coverage boundary.
[0071] During the feedback correction phase, the predicted thermal insulation performance value is corrected based on the path determined in the result verification phase. When the prediction boundary realignment result enters the positive boundary correction path, the positive boundary correction segment is defined by the minimum and maximum deviation values in the positive sample prediction deviation interval, and the predicted thermal insulation performance value is lowered according to the deviation direction corresponding to the positive boundary correction segment. When the prediction boundary realignment result enters the negative boundary correction path, the negative boundary correction segment is defined by the minimum and maximum deviation values in the negative sample prediction deviation interval, and the predicted thermal insulation performance value is raised according to the deviation direction corresponding to the negative boundary correction segment. When the prediction boundary realignment result enters the retention path, the predicted thermal insulation performance value is retained. The predicted thermal insulation performance value after being lowered, raised, or retained is used as the predicted thermal insulation performance value, and the predicted thermal insulation performance value retains the corresponding thermal bridge perturbation position, training coverage boundary, and boundary correction segment.
[0072] It should be further explained that the chain-linked interaction involves the continuous flow of thermal insulation prediction features, thermal bridge perturbation positions, lock-in responses, perturbation position deviation sequences, and training coverage boundaries between S3.12 and S4.4, including field-by-field transmission, deviation write-back, boundary calls, and result feedback. When the thermal insulation prediction features enter S4.1, the three-path response position sequence is used as the response input field, the position value of the thermal bridge perturbation position is used as the position input field, and the compensation value of the lock-in response is used as the compensation input field. After S4.1 generates the sample prediction deviation, it is processed according to the component number, temperature measurement location, thermal bridge perturbation position, and training batch order. S4.2 Write the perturbation bit deviation sequence; S4.2 After reading the perturbation bit deviation sequence and generating the training coverage boundary, provide the lower bound of the interval, the upper bound of the interval, the minimum deviation value, and the maximum deviation value to S4.3; S4.3 Retrieve the matching training coverage boundary according to the thermal bridge perturbation bit and generate the prediction boundary repositioning result; S4.4 Determine the boundary correction segment according to the prediction boundary repositioning result, and write the corrected thermal insulation performance prediction value, thermal bridge perturbation bit, training coverage boundary, and boundary correction segment into the thermal insulation performance prediction result, thereby completing the chain linkage from feature input to deviation write-back and from boundary call to prediction correction.
[0073] S4. Train a machine learning model using a thermal test sample set, and use the machine learning model to predict the thermal insulation performance of the thermal insulation prediction features to obtain the predicted thermal insulation performance value. Correct the predicted thermal insulation performance value with the training coverage boundary of the thermal bridge perturbation site to generate the predicted thermal insulation performance result.
[0074] S4.1. Train a machine learning model using the thermal test sample set, and feed back the sample prediction bias generated during the training process to the thermal bridge perturbation position to generate a perturbation position bias sequence. Specifically, The thermal insulation prediction features in the thermal test sample set are used as training input, and the measured thermal insulation performance values under the same component number and the same temperature measurement location are used as training labels. The measured thermal insulation performance values are the integral values of cold end temperature rise within the test start and end time range. The integral values of cold end temperature rise are calculated from the initial cold end temperature in the cold end temperature curve, the cold end temperature at each sampling time, and the time interval between adjacent samplings, as shown in the following formula: ; In the formula, This represents the integral value of the cold junction temperature rise. This indicates the initial cold junction temperature of the cold junction temperature curve at the start of the test. This represents the cold end temperature curve, which is a function of how the cold end temperature changes over time. Indicates the start time of the test. The cold end temperature curve indicates the temperature at the first... The cold junction temperature at each sampling time. Indicates the first The time interval between each sampling moment and the previous sampling moment This indicates the number of sampling points after the start of the test; the cold end temperature rise integral value is used to characterize the cumulative degree of heat transfer from the insulation component to the cold end under the same test boundary; the hot end temperature curve and the thickness temperature curve are not directly used as training labels, but are used for insulation prediction feature construction and three-path response consistency verification.
[0075] It should be noted that the machine learning model adopts a feedforward neural network structure, which includes an input layer, a first fully connected layer, a second fully connected layer, and an output layer. Before the thermal insulation prediction features enter the input layer, the hot-end loading response, thickness hysteresis response, and cold-end transfer response in the three-path response bit sequence are resampled into 128 sampling points to form a 384-dimensional response input vector. The thermal bridge perturbation bits are converted into a 4-dimensional position input vector composed of material level number, local structure type number, temperature measurement position number, and response position number. The compensation values in the locked domain response are resampled into a 128-dimensional compensation input vector. The 384-dimensional response input vector, the 4-dimensional position input vector, and the 128-dimensional compensation input vector are concatenated to form a 516-dimensional model input vector. The input layer receives the 516-dimensional model input vector. The first fully connected layer has 256 neurons and uses the ReLU activation function, the second fully connected layer has 64 neurons and uses the ReLU activation function, and the output layer has 1 neuron and uses a linear output function. The output layer outputs the sample prediction value, which is the cold-end temperature rise integral prediction value.
[0076] The sample prediction value is compared with the integral value of the cold-end temperature rise. When the sample prediction value is greater than the integral value of the cold-end temperature rise, the difference between the sample prediction value and the integral value of the cold-end temperature rise is taken as a positive sample prediction bias; when the sample prediction value is less than the integral value of the cold-end temperature rise, the difference between the sample prediction value and the integral value of the cold-end temperature rise is taken as a negative sample prediction bias; when the sample prediction value is equal to the integral value of the cold-end temperature rise, the sample prediction bias is recorded as zero bias. During the training process, the mean squared error loss function is used to constrain the sample prediction bias, and the first fully connected layer and the second fully connected layer are updated through backpropagation and gradient descent. The weight parameters in the output layer enable the machine learning model to reduce the sample prediction bias round by round. After each training round, the sample prediction bias is written into the same thermal bridge perturbation position according to the component number, temperature measurement position, thermal bridge perturbation position, and training batch order to form a perturbation position bias term. The perturbation position bias term includes the position value of the thermal bridge perturbation position, the deviation value of the sample prediction bias, the deviation direction of the sample prediction bias, and the training batch order. The perturbation position bias terms belonging to the same thermal bridge perturbation position are arranged according to the training batch order to generate a perturbation position bias sequence. The perturbation position bias sequence is used to define the training coverage boundary in the future.
[0077] S4.2 Arrange the sample prediction biases belonging to the same hot bridge perturbation position in the perturbation position bias sequence according to the direction of deviation, define the deviation intervals continuously occupied by the sample prediction biases, and close the boundary positions of the deviation intervals to generate the training coverage boundary. Specifically, Sample prediction deviations belonging to the same thermal bridge disturbance site are grouped together, and the component number, temperature measurement location, and training batch order corresponding to the sample prediction deviation are retained. The sample prediction deviations are divided into positive sample prediction deviations, negative sample prediction deviations, and zero deviations according to the direction of deviation. Positive sample prediction deviations are arranged in the direction higher than the measured value of thermal insulation performance, negative sample prediction deviations are arranged in the direction lower than the measured value of thermal insulation performance, and zero deviations are placed between positive and negative sample prediction deviations.
[0078] Under the same deviation direction, adjacent sample prediction deviations are connected according to the training batch order. Sample prediction deviations with the same deviation direction and continuous batch position are connected into a continuous deviation segment. The starting training batch, ending training batch, minimum deviation value and maximum deviation value occupied by the continuous deviation segment are organized into a deviation interval. The starting boundary of the deviation interval is used as the lower boundary of the interval, and the ending boundary of the deviation interval is used as the upper boundary of the interval. The lower boundary, upper boundary, minimum deviation value and maximum deviation value together define the training coverage boundary under the same thermal bridge perturbation position. The training coverage boundary is used to receive the thermal bridge perturbation position corresponding to the thermal insulation prediction feature and complete the prediction boundary positioning.
[0079] S4.3. The thermal insulation performance of the thermal insulation prediction features is predicted using a machine learning model to obtain predicted thermal insulation performance values. The thermal bridge perturbation sites corresponding to the thermal insulation prediction features are then incorporated into the training coverage boundary to obtain the prediction boundary relocation results. Specifically... The thermal insulation prediction features of the thermal insulation component to be predicted are fed into the input layer of the machine learning model. The input layer receives a 384-dimensional response input vector formed by the three-path response bit sequence, a 4-dimensional position input vector formed by the thermal bridge perturbation bit, and a 128-dimensional compensation input vector formed by the domain lock response. The first fully connected layer extracts the temperature value change features in the three-path response bit sequence. The second fully connected layer connects the temperature value change features, the position values of the thermal bridge perturbation bit, and the compensation values of the domain lock response into a prediction joint feature. The output layer converts the prediction joint feature into the cold end temperature rise integral prediction value and uses the cold end temperature rise integral prediction value as the thermal insulation performance prediction value.
[0080] Among them, the index of the thermal insulation performance prediction value is the cold end temperature rise integral prediction value, which is used to characterize the predicted cumulative degree of heat transfer from the thermal insulation component to the cold end under the same test boundary; the smaller the cold end temperature rise integral prediction value, the lower the cumulative temperature rise at the cold end under the same test boundary, and the stronger the thermal insulation performance of the thermal insulation component to be predicted.
[0081] The thermal bridge disturbance position carried by the thermal insulation prediction feature is matched with the same component number, the same temperature measurement position, and the same thermal bridge disturbance position in the training coverage boundary. When the match is successful, the lower boundary, upper boundary, minimum deviation value, maximum deviation value, and deviation interval type in the training coverage boundary are read, and the predicted thermal insulation performance value, thermal bridge disturbance position, training coverage boundary, and deviation interval type are written into the same relocation record to form the prediction boundary relocation result. When the match fails, the prediction boundary relocation anomaly mark is retained, and the predicted thermal insulation performance value is listed as the prediction value to be reviewed.
[0082] It should be noted that, as Figure 5 As shown, within the same test sample number range, the absolute error of the cold-end temperature rise integral prediction of the original temperature curve-feedforward neural network prediction technology fluctuates significantly overall, with obvious error peaks appearing in the local magnification interval. The error curves of the thermal boundary stripping three-path reconstruction prediction technology, the back-projection positioning prediction technology without sparse compensation technology, and the time-map sparse compensation prediction technology without boundary correction technology decrease sequentially. The prediction technology of this invention maintains the lowest error distribution in the main comparison interval, indicating that after the thermal insulation prediction feature carries the thermal bridge disturbance position into the machine learning model, it can reduce the prediction deviation caused by the combined effect of test boundary disturbance and local thermal bridge disturbance, and improve the stability of cold-end temperature rise integral prediction.
[0083] S4.4. Based on the predicted boundary alignment results, the predicted thermal insulation performance value is aligned with the deviation interval in the training coverage boundary to determine the boundary correction segment of the predicted thermal insulation performance value. The predicted thermal insulation performance value is then corrected using this boundary correction segment to generate the predicted thermal insulation performance result. Specifically... The predicted thermal insulation performance value in the predicted boundary alignment result is matched with the training coverage boundary. During boundary matching, the lower boundary, upper boundary, minimum deviation value, and maximum deviation value of the interval in the training coverage boundary are read, and then the thermal bridge disturbance position corresponding to the predicted thermal insulation performance value is checked to see if it falls into the positive sample prediction deviation interval, the negative sample prediction deviation interval, or the zero deviation position.
[0084] When the predicted thermal insulation performance value corresponds to the positive sample prediction deviation range, the minimum and maximum deviation values in the positive sample prediction deviation range are used to define the positive boundary correction segment, and the predicted thermal insulation performance value is adjusted downward according to the deviation direction corresponding to the positive boundary correction segment. When the predicted thermal insulation performance value corresponds to the negative sample prediction deviation range, the minimum and maximum deviation values in the negative sample prediction deviation range are used to define the negative boundary correction segment, and the predicted thermal insulation performance value is adjusted upward according to the deviation direction corresponding to the negative boundary correction segment. When the predicted thermal insulation performance value corresponds to the zero deviation position, the predicted thermal insulation performance value is retained. The predicted thermal insulation performance value after the boundary correction segment is completed is taken as the predicted thermal insulation performance value.
[0085] This embodiment also provides a computer device applicable to the case of a machine learning-based thermal insulation performance prediction method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the machine learning-based thermal insulation performance prediction method proposed in the above embodiment.
[0086] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0087] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the machine learning-based thermal insulation performance prediction method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0088] In summary, this invention achieves structured correction of thermal bridge perturbation sites through back-projection localization and graph-time sparsity compensation. After embedding material structure information into the thermal insulation response features, the response residuals can fall back to the thermal bridge perturbation sites along material hierarchical relationships and adjacent structural channels, giving local anomalies in the temperature response a clear structural origin, rather than merely manifesting as curve amplitude deviations. The perturbation response band can limit the three-path response range occupied by the thermal bridge perturbation sites, the sparse correction slots can narrow the local correction objects, and the graph-time compensation chain can convert response residual fragments into backfillable compensation expressions, thereby preserving the continuous relationship between the hot-end loading response, the thickness-wise hysteresis response, and the cold-end transfer response. The machine learning model performs predictions based on the structured-corrected thermal insulation prediction features, improving the stability, reliability, and interpretability of the thermal insulation performance prediction results.
[0089] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A machine learning based thermal insulation performance prediction method, characterized by: include, Obtain material structure information, test boundary information, and temperature response information of thermal insulation components, and group them according to the same component number and temperature measurement location to generate a thermal test sample set; Thermal boundary stripping is performed on the thermal test sample set to separate the response offset caused by the test boundary information from the temperature response information, and the temperature response information is reconstructed according to the three-path thermal response relationship to generate thermal insulation response characteristics. The thermal insulation response features are embedded into the material structure information. The thermal bridge perturbation sites are determined based on the back projection positions of the response residuals in the thermal insulation response features into the material structure information. The response segments occupied by the thermal bridge perturbation sites are then sparsely corrected to generate thermal insulation prediction features. A machine learning model is trained using a thermal test sample set, and the model is used to predict the thermal insulation performance based on the thermal insulation prediction features. The predicted thermal insulation performance value is then corrected using the training coverage boundary of the thermal bridge perturbation site to generate the predicted thermal insulation performance result. 2.The machine learning based thermal insulation performance prediction method of claim 1, wherein: The component number is a unique identifier used to distinguish different thermal insulation components and to group the material structure information, test boundary information and temperature response information of the same thermal insulation component into the same thermal test sample set. The temperature measurement position is a fixed spatial position on the thermal insulation component used to collect temperature response information, so as to limit the merging range of test boundary information and temperature response information in the thermal test sample set at the same position; The three-path thermal response relationship is the response arrangement relationship formed by the hot end loading response, the thickness hysteresis response, and the cold end transfer response in the temperature response information according to the order of heat transfer. 3.The machine learning based thermal insulation performance prediction method of claim 2, wherein: The generated thermal insulation response characteristics are specifically as follows: Within the thermal test sample set, a single sample range is divided according to the same component number and temperature measurement location. The test boundary information within a single sample range is mapped to the time occupancy location and temperature measurement occupancy location in the temperature response information to generate boundary occupancy segments. Embed the boundary occupancy segment into the temperature response information, separate the response offset caused by the test boundary information along the temperature response information, and generate the boundary stripping response segment. The boundary stripping response fragment is used to connect the three-channel thermal response relationship. The response bits of the boundary stripping response fragment are rearranged to generate the three-channel response bit sequence. The temperature response information is reconstructed based on the three-path response bit sequence. The reconstructed temperature response information is then normalized in terms of response scale and arranged according to the three-path thermal response relationship to generate thermal insulation response characteristics. 4.The machine learning based thermal insulation performance prediction method of claim 3, wherein: The response scale normalization is to unify the amplitude range and align the time scale of the three-channel response bit sequence in the reconstructed temperature response information to form normalized temperature response information. 5.The machine learning based thermal insulation performance prediction method of claim 4, wherein: The determination of the thermal bridge disturbance location specifically involves... Set the component number and temperature measurement location in the material structure information as the structural index, merge the material hierarchy and local structural locations under the structural index to form a structural location item; The structural location items are connected in series according to the material hierarchy relationship in the structural location items, and adjacent structural channels are established in the local structural locations in the structural location items to generate a material structural channel chain. The thermal insulation response characteristics are written into the material structure channel chain, and the three-path response bit sequence in the thermal insulation response characteristics is attached to the adjacent structural channels under the same temperature measurement position to generate the structural load-bearing response chain. Based on the material hierarchy and adjacent structural channels in the structural load-bearing response chain, the response features that do not continuously shift with the three-path response sequence are extracted from the thermal insulation response features, and the response residual fragments are output. The three-path thermal response relationship carried by the response residual fragment is transformed into a back-projection gating relationship, and the response residual fragment is projected onto the material structure channel chain according to the back-projection gating relationship to generate a phase-gated back-projection fragment. The phase-gated backprojection segment is subjected to structural channel competition and repositioning. The backprojection position in the phase-gated backprojection segment that simultaneously bears the disconnection of material hierarchy relationship and the abrupt change of adjacent structural channel is determined as the thermal bridge disturbance position. 6.The machine learning based thermal insulation performance prediction method of claim 5, wherein: The generated thermal insulation prediction features are specifically as follows: The thermal bridge disturbance is returned to the structural load-bearing response chain, and the three-path response range occupied by the thermal bridge disturbance is delineated in the structural load-bearing response chain as the disturbance response band. The response residual fragments are embedded into the perturbation response band, and the correction range of the perturbation response band is narrowed according to the position of the response residual fragments in the perturbation response band to generate sparse correction slots. Based on the material hierarchy relationship and adjacent structural channels of the sparse correction slot in the material structure channel chain, the response residual fragment in the sparse correction slot is subjected to graph-time expansion sparse compensation to obtain the graph-time compensation chain. The time-compensation chain is backfilled to the position of the response residual segment corresponding to the sparse correction slot, and the response segments in the disturbance response band located outside the sparse correction slot are retained in the original response position, and the locked domain response is output. By updating the response fragment occupied by thermal bridge perturbation bits in the insulation response features through domain locking response, insulation prediction features are generated.
7. The machine learning based thermal insulation performance prediction method of claim 6, wherein: The graph-time unfolding sparsity compensation refers to unfolding the response residual fragments in the sparse correction slots along the three-path response bit sequence based on the material hierarchy relationship and adjacent structural channels in the material structure channel chain, and converting the unfolded response residual fragments into a graph-time compensation chain.
8. The machine learning based thermal insulation performance prediction method of claim 7, wherein: The generated thermal insulation performance prediction results are specifically as follows: A machine learning model is trained using a thermal test sample set, and the sample prediction bias generated during the training process is fed back to the thermal bridge perturbation position to generate a perturbation position bias sequence. Arrange the sample prediction biases belonging to the same hot bridge perturbation position in the perturbation position bias sequence according to the deviation direction, define the deviation interval continuously occupied by the sample prediction bias, and close the boundary position of the deviation interval to generate the training coverage boundary. The thermal insulation performance is predicted by using a machine learning model to predict the thermal insulation performance of the thermal insulation prediction features. The thermal bridge perturbation sites corresponding to the thermal insulation prediction features are then assigned to the training coverage boundary to obtain the prediction boundary assignment result. Based on the predicted boundary alignment results, the predicted thermal insulation performance value is aligned with the deviation interval in the training coverage boundary to determine the boundary correction segment of the predicted thermal insulation performance value. The predicted thermal insulation performance value is then corrected using the boundary correction segment to generate the predicted thermal insulation performance result. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: When the processor executes the computer program, it implements the steps of the machine learning-based thermal insulation performance prediction method according to any one of claims 1 to 8.
10. A computer readable storage medium having stored thereon a computer program, characterized in that: When the computer program is executed by the processor, it implements the steps of the machine learning-based thermal insulation performance prediction method according to any one of claims 1 to 8.
