Building intelligent operation and maintenance system and method based on large model

Abstract

The application discloses a building intelligent operation and maintenance system and method based on a large model, and relates to the field of intelligent detection. Firstly, one-dimensional convolutional coding is performed on the first device energy consumption time sequence to extract local features, and time sequence propagation coding is used to capture dynamic changes on the time axis. Then, based on the device attribute semantic association matrix, the first device energy consumption features are nonlinearly mapped into the second device semantic space, so that the second device energy consumption prediction fuses the influence of the upstream device operating state. By comparing the difference between the prediction and the actual monitoring data, the second device hidden energy consumption fluctuation caused by the first device is identified. This method breaks through the limitation of traditional isolated modeling, realizes feature transmission across the device semantic space, accurately captures the energy transmission anomaly between devices, effectively reduces false positives and false negatives, improves the accuracy and efficiency of anomaly detection, and reduces the need for manual verification.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent detection, and more particularly, to a building intelligent operation and maintenance system and method based on a large model. BACKGROUND

[0002] With the continuous improvement of the complexity of modern building systems, building operation and maintenance management has evolved from single device monitoring to multi-system collaborative optimization. Currently, large public buildings generally deploy building automation systems (BAS), energy management systems (EMS), and other intelligent platforms, which continuously collect operation parameters and energy consumption time series data of heating, ventilation, air conditioning, elevators, lighting, and other devices.

[0003] Traditional operation and maintenance methods mainly rely on threshold alarm mechanisms and isolated device analysis based on statistical models, which are difficult to effectively analyze the complex coupling relationship between devices. In particular, in high-density sensor deployment scenarios, the time series correlation of device operating states presents non-stationary and multi-scale characteristics. For example, air conditioning system abnormalities may indirectly cause fluctuations in energy consumption of ventilation equipment through air flow paths, and traditional rule engine-based correlation analysis methods cannot capture such cross-system implicit fault propagation chains. In addition, existing time series prediction models often ignore the mapping relationship between device attribute semantic space and energy consumption dynamic characteristics when dealing with multi-device collaborative optimization, resulting in feature drift problems in cross-device anomaly reasoning. When faced with complex energy transfer and working condition collaborative relationships between devices, traditional methods often produce a large number of false positives or false negatives, forcing operation and maintenance personnel to repeatedly manually check to confirm device states, which not only significantly increases operation and maintenance costs, but also may lead to a response lag in critical device abnormal conditions.

[0004] Therefore, a building intelligent operation and maintenance solution based on a large model is desired. SUMMARY

[0005] To solve the above technical problems, the present application is proposed.

[0006] According to an aspect of the present application, a building intelligent operation and maintenance method based on a large model is provided, which includes:

[0007] Obtaining a time queue of energy consumption data of a first device in a building;

[0008] Obtaining a time queue of energy consumption data of a second device in the building;

[0009] Performing local time series-based recursive message passing coding on the time queue of energy consumption data of the first device to obtain a first device energy consumption time series pattern message passing coding vector;

[0010] coding the time queue of the energy consumption data of the second device based on the first device energy consumption time sequence pattern message passing coding vector to obtain a generated time queue of the energy consumption data of the second device;

[0011] calculating an energy consumption data time sequence distribution difference value between the time queue of the energy consumption data of the second device and the generated time queue of the energy consumption data of the second device;

[0012] confirming whether the energy consumption of the second device is abnormal based on a comparison between the energy consumption data time sequence distribution difference value and a preset threshold value.

[0013] According to another aspect of the present application, there is provided a large model-based building intelligent operation and maintenance system, which comprises:

[0014] an energy consumption data acquisition module of the first device, configured to acquire a time queue of energy consumption data of a first device in a building;

[0015] an energy consumption data acquisition module of the second device, configured to acquire a time queue of energy consumption data of a second device in the building;

[0016] a first device energy consumption time sequence coding module, configured to perform local time sequence-based recursive message passing coding on the time queue of the energy consumption data of the first device to obtain a first device energy consumption time sequence pattern message passing coding vector;

[0017] a second device energy consumption time sequence coding module, configured to perform energy consumption time sequence coding and decoding on the time queue of the energy consumption data of the second device based on the first device energy consumption time sequence pattern message passing coding vector to obtain a generated time queue of the energy consumption data of the second device;

[0018] an energy consumption data difference calculation module, configured to calculate an energy consumption data time sequence distribution difference value between the time queue of the energy consumption data of the second device and the generated time queue of the energy consumption data of the second device;

[0019] an abnormality judgment module, configured to confirm whether the energy consumption of the second device is abnormal based on a comparison between the energy consumption data time sequence distribution difference value and a preset threshold value.

[0020] Compared with the prior art, the building intelligent operation and maintenance system and method based on a large model provided by the application adopt a data processing and analysis algorithm based on large model artificial intelligence. The technical concept constructs a dynamic correlation energy consumption reasoning mechanism between devices through a large model. Firstly, one-dimensional convolution coding is performed on the energy consumption time sequence of a first device to extract local time domain correlation features of the running mode of the first device. Then, time sequence propagation coding is used to capture the dynamic evolution law of the features on the time axis. Subsequently, based on a device attribute semantic correlation matrix, the energy consumption features of the first device are nonlinearly mapped to the semantic space of a second device, so that the energy consumption prediction of the second device can fuse the influence of the running state of the upstream device. Finally, by comparing the time sequence distribution difference between the predicted sequence and the actual monitoring data, the indirect energy consumption fluctuation of the second device caused by the abnormality of the first device is identified. The scheme breaks through the limitation of traditional isolated modeling, accurately captures the energy transmission abnormality of the second device caused by the failure of the first device through dynamic feature transmission across the semantic space of the devices, integrates the originally scattered threshold alarm events into complete implicit fault propagation chain traceability analysis, significantly reduces the false positives and false negatives caused by the unmodeled working condition coupling between devices, improves the accuracy and timeliness of anomaly detection, and reduces the workload of manual checking. BRIEF DESCRIPTION OF DRAWINGS

[0021] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application taken in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of embodiments of the present application and are incorporated in and constitute a part of this specification, illustrate embodiments of the present application and serve to explain the present application, and do not constitute a limitation of the present application. In the drawings, like reference numerals refer to like parts or steps throughout.

[0022] Figure 1 A flowchart of the building intelligent operation and maintenance method based on a large model according to an embodiment of the present application.

[0023] Figure 2 A flowchart of step S130 in the building intelligent operation and maintenance method based on a large model according to an embodiment of the present application.

[0024] Figure 3 A flowchart of step S132 in the building intelligent operation and maintenance method based on a large model according to an embodiment of the present application.

[0025] Figure 4 A flowchart of step S140 in the building intelligent operation and maintenance method based on a large model according to an embodiment of the present application.

[0026] Figure 5 A block diagram of the building intelligent operation and maintenance system based on a large model according to an embodiment of the present application. DETAILED DESCRIPTION

[0027] Embodiments of the present disclosure will be described below in greater detail with reference to the accompanying drawings. While certain embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be embodied in various forms and should not be construed as being limited to the embodiments set forth herein, but rather, these embodiments are provided so that the present disclosure can be more thoroughly and completely understood. It is understood that the drawings of the present disclosure and the embodiments are for exemplary purposes only and are not intended to limit the scope of protection of the present disclosure.

[0028] It is understood that each of the steps described in the method embodiments of the present disclosure can be performed in different orders and / or in parallel. In addition, the method embodiments can include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this regard.

[0029] To solve the problems in the background art, the technical concept of the present application is to use a data processing and analysis algorithm based on large model artificial intelligence. The technical concept constructs an energy consumption reasoning mechanism for inter-device dynamic correlation through a large model. First, one-dimensional convolution coding is performed on the energy consumption time sequence of a first device to extract local time domain correlation features of its running mode. Then, time sequence propagation coding is used to capture the dynamic evolution law of the features on the time axis. Subsequently, based on a device attribute semantic correlation matrix, the energy consumption features of the first device are nonlinearly mapped to the semantic space of a second device, so that the energy consumption prediction of the second device can integrate the influence of the running state of the upstream device. Finally, by comparing the time sequence distribution difference between the predicted sequence and the actual monitoring data, the indirect energy consumption fluctuation of the second device caused by the abnormality of the first device is identified. This scheme breaks through the limitations of traditional isolated modeling, accurately captures the energy transmission abnormalities caused by the first device failure to the second device through dynamic feature transmission across the semantic space of devices, integrates the originally scattered threshold alarm events into complete implicit fault propagation chain traceability analysis, significantly reduces false positives and false negatives caused by the uncoupling of device working conditions, improves the accuracy and timeliness of anomaly detection, and reduces the workload of manual verification.

[0030] Figure 1 A flowchart of the large model-based building intelligent operation and maintenance method according to an embodiment of the present application. As shown in FIG. 1, the method includes the following steps. Figure 1As shown, the building intelligent operation and maintenance method based on a large model according to the embodiment of the present application comprises: S110, acquiring a time queue of energy consumption data of a first device in a building; S120, acquiring a time queue of energy consumption data of a second device in the building; S130, performing local time sequence-based recursive message passing coding on the time queue of energy consumption data of the first device to obtain a first device energy consumption time sequence pattern message passing coding vector; S140, based on the first device energy consumption time sequence pattern message passing coding vector, performing energy consumption time sequence coding and decoding on the time queue of energy consumption data of the second device to obtain a generated time queue of energy consumption data of the second device; S150, calculating an energy consumption data time sequence distribution difference value between the time queue of energy consumption data of the second device and the generated time queue of energy consumption data of the second device; S160, based on a comparison between the energy consumption data time sequence distribution difference value and a preset threshold, confirming whether the energy consumption of the second device is abnormal.

[0031] In step S110, a time queue of energy consumption data of a first device in a building is acquired. It should be understood that the time queue of energy consumption data of the first device in the building refers to a sequence formed by sequentially recording the energy consumption data of the first device at different time points in time order. For example, the electrical energy consumption values of the first device (such as a heating, ventilation and air conditioning device) are collected at certain time intervals (such as every minute, every hour, etc.), and these values are arranged in the order of collection time to form the energy consumption data time queue of the device. These data reflect the changes in the energy consumption of the device over a period of time. In particular, the energy consumption of the device in the building is an important basis for evaluating the running state, performance of the device and the overall energy utilization efficiency of the building. The first device is one of the many devices in the building, and its energy consumption data is crucial for understanding the working state of the device itself and the relationship with other devices. By acquiring the energy consumption data time queue, the energy consumption pattern, energy consumption trend, etc. of the device can be analyzed to further discover possible abnormal conditions or potential problems of the device. For example, if it is found that the energy consumption of a device suddenly increases significantly in a certain period of time, it may mean that the device has a fault or abnormal operation, which needs to be further checked and maintained.

[0032] In step S120, a time queue of energy consumption data for the second equipment within the building is acquired. Accordingly, the time queue of energy consumption data for the second equipment within the building is a collection of energy consumption data from different times, arranged chronologically. For example, for elevator equipment in a building, its power consumption is collected at fixed time intervals (e.g., every 5 minutes). Arranging these power consumption data that change over time sequentially forms the time queue of energy consumption data for this second equipment, visually presenting the dynamic changes in energy consumption of the equipment within a certain time range. As part of the building's equipment system, the energy consumption of the second equipment is a crucial factor in assessing the overall energy consumption of the building and the operating status of the equipment. Different equipment has different energy consumption characteristics and variation patterns. Acquiring the time queue of energy consumption data for the second equipment helps to analyze the energy consumption pattern of that equipment individually, determine whether there are abnormal energy consumption situations, such as excessively high energy consumption or abnormal fluctuations, and thus promptly detect potential equipment malfunctions or inefficient operating states. In particular, in intelligent building operation and maintenance, equipment does not operate in isolation but rather has complex interrelationships. By acquiring the energy consumption data time queue of the second device, it is possible to compare and correlate it with the energy consumption data of the first device and other related devices, explore the relationships between devices such as energy transfer and operating condition coordination, and provide data support for revealing the hidden fault propagation chain between devices.

[0033] In step S130, the time queue of the energy consumption data of the first device is subjected to recursive message passing encoding based on local timing to obtain the energy consumption timing pattern message passing encoding vector of the first device.

[0034] Figure 2 This is a flowchart of step S130 in the intelligent building operation and maintenance method based on a large model according to an embodiment of this application. Specifically, in the embodiments of this application, as... Figure 2 As shown, step S130, which involves performing recursive message passing encoding based on local temporal sequence on the time queue of the energy consumption data of the first device to obtain the energy consumption time sequence pattern message passing encoding vector of the first device, includes: S131, performing energy consumption time sequence pattern feature extraction based on one-dimensional convolutional encoding on the time queue of the energy consumption data of the first device to obtain the sequence distribution of the local temporal domain correlation feature vector of the first device's energy consumption; S132, performing temporal propagation recursive structured encoding on the sequence distribution of the local temporal domain correlation feature vector of the first device's energy consumption to obtain the energy consumption time sequence pattern message passing encoding vector of the first device.

[0035] Specifically, the step S131 performs energy consumption time sequence pattern feature extraction based on one-dimensional convolution coding on the time queue of the energy consumption data of the first device to obtain a sequence distribution of first device energy consumption local time domain correlation feature vectors. It should be understood that, considering that the device energy consumption data has complex non-steady-state, multi-scale time sequence characteristics. Traditional threshold alarm mechanisms and statistical analysis methods often use fixed time window mean statistics or extreme value detection, which are difficult to effectively capture micro time sequence patterns with spatial locality such as transient current peaks at air conditioner compressor start-stop moments, ladder-shaped energy consumption curves characteristic of elevator acceleration stages, etc. These features often carry potential signs of early device failure, such as abnormal fluctuations in motor starting current caused by bearing wear, but traditional methods lack the ability to extract fine local time domain correlation features, which can easily be submerged in the overall energy consumption trend, resulting in insufficient sensitivity of abnormal detection. Therefore, the application performs energy consumption time sequence pattern feature extraction based on one-dimensional convolution coding on the time queue of the energy consumption data of the first device to obtain a sequence distribution of first device energy consumption local time domain correlation feature vectors. That is, the one-dimensional convolution coding technology is used to perform deep feature mining on the time queue of the energy consumption data of the first device, and the essence is to use a learnable convolution kernel to perform local sliding calculation along the time axis. This processing method can adaptively capture energy consumption fluctuation patterns of different time scales: larger convolution kernels identify long-period energy consumption transitions during device start-stop stages, and smaller convolution kernels focus on high-frequency oscillation details during operation. Through multi-layer convolution stacking, a sequence of local time domain correlation feature vectors with hierarchical representation ability is finally formed, such as in a heating, ventilation and air conditioning device, the periodic performance consumption peak corresponding to the refrigerant circulation period and the non-periodic performance consumption burr caused by abnormal leakage can be accurately separated.

[0036] Specifically, the step S132 is to perform time sequence propagation recursive structured coding on the sequence distribution of the first device energy consumption local time domain correlation feature vector to obtain the first device energy consumption time sequence mode message passing coding vector. Further, considering that the time sequence features of device energy consumption often have complex dynamic evolution rules. The traditional method uses fixed time window statistical modeling or simple recurrent neural network processing, which is difficult to effectively capture implicit correlations across time steps such as gradual shift of ventilation device energy consumption within hours after air conditioning system anomaly. Although the prior art can extract local time domain features, it lacks the ability to model the dynamic propagation mechanism of features on the time axis, resulting in the inability to distinguish between normal operating condition fluctuations and early abnormality-induced time sequence mode distortion. For example, when the HVAC compressor bearing is worn, its energy consumption curve will show the feature of gradually increasing harmonic components in a specific frequency band. If only isolated time window analysis is used, these weak time sequence mode changes are easily misjudged as random noise and their potential systematic evolution trend is ignored. Based on this, the first device energy consumption time sequence mode message passing coding vector is obtained by performing time sequence propagation recursive structured coding on the sequence distribution of the first device energy consumption local time domain correlation feature vector.

[0037] Specifically, first, the recurrent neural network is used to capture the preliminary time sequence dependence of the first device energy consumption local features to generate an initial sequence transfer coding vector sequence. Then, the dynamic importance of different time step feature vectors is quantified by a time domain credibility adjustment coefficient (such as identifying the higher diagnostic value of transient features in the compressor start-stop phase than in the steady-state running phase), and an air domain credibility adjustment coefficient is introduced to reflect the structural influence of the device in the network topology (such as the key node attribute of the air conditioner host compared to the end fan). The message passing space-time collaborative adjustment coefficient formed by the fusion of the two can dynamically adjust the information flow intensity of the features on the time propagation path, for example, enhance the persistence propagation weight of abnormal harmonic components on the time axis, and suppress the transmission attenuation of normal operating condition fluctuations. In this way, the first device energy consumption time sequence mode message passing coding vector obtained is a high-level feature representation of the original energy consumption data after multiple processing. It not only integrates local and global energy consumption feature information, but also considers the dynamic propagation and interaction of features on the time sequence, which can more accurately depict the energy consumption behavior mode of the first device, provide more representative and discriminant feature vectors for subsequent device state evaluation, fault diagnosis and energy consumption prediction tasks, and help improve the accuracy and reliability of these tasks.

[0038] Figure 3 The flowchart of step S132 in the building intelligent operation and maintenance method based on a large model according to the embodiment of the present application. Specifically, in the embodiment of the present application, as shown in FIG. 1, the first device energy consumption time sequence mode message passing coding vector is obtained by performing time sequence propagation recursive structured coding on the sequence distribution of the first device energy consumption local time domain correlation feature vector. Figure 3As shown, step S132, which involves performing time-series propagation recursive structured encoding on the sequence distribution of the first device energy consumption local temporal domain correlation feature vector to obtain the first device energy consumption temporal pattern message passing encoding vector, includes: S132-1, performing sequence encoding based on a recurrent neural network on the sequence distribution of the first device energy consumption local temporal domain correlation feature vector to obtain the sequence distribution of the sequence-transmitted first device energy consumption local temporal domain correlation initial feature vector; S132-2, performing message passing based on spatiotemporal reliability adjustment on the sequence distribution of the sequence-transmitted first device energy consumption local temporal domain correlation initial feature vector to obtain the first device energy consumption temporal pattern message passing encoding vector.

[0039] Specifically, step S132-1 involves performing sequence encoding based on a recurrent neural network on the sequence distribution of the local temporal correlation feature vector of the first device's energy consumption to obtain the sequence distribution of the initial local temporal correlation feature vector of the first device's energy consumption for sequence transmission. This process can be represented by the following formula:

[0040] I = {v1, v2, ..., v} i ,...,v t}

[0041] RNN(I) = {h1,h2,...,h} i ,...,h t}

[0042] Where I is the sequence distribution of the local time-domain correlation feature vector of the energy consumption of the first device, v1, v2, v i and v t Let h1, h2, and ht be the 1st, 2nd, i-th, and t-th local time-domain correlation feature vectors of the first device's energy consumption, respectively, in the sequence distribution of the local time-domain correlation feature vectors of the first device's energy consumption. RNN(I) is a sequence encoding of I based on an RNN structure. i and h t These are the 1st, 2nd, i-th, and t-th initial feature vectors of the local temporal correlation of the energy consumption of the first device in the sequence transmission, respectively.

[0043] It can be understood that the timing characteristics of the device energy consumption often have long-range dependence and dynamic evolution characteristics. Although the feature extraction of one-dimensional convolution can separate the microscopic fluctuations in the operation of the device, the characteristics of each time window are in an isolated state, and it is difficult to represent the systematic evolution law of the energy consumption pattern in the aging process of the device. For example, when the heating, ventilation and air conditioning evaporator is frosting, its energy consumption curve will show a long-term trend of gradually decaying periodic peak amplitude. If the feature correlation across dozens of sampling points is only dependent on local convolution coding, it is easy to be divided into discrete abnormal fragments and lose the key information of the evolution path. Therefore, the sequence distribution of the first device energy consumption local time domain correlation feature vector is encoded based on a recurrent neural network to perform deep time sequence modeling on the first device energy consumption local time domain correlation feature sequence, and the sequence distribution of the sequence transfer first device energy consumption local time domain correlation initial feature vector is obtained. The recurrent connection mechanism of RNN gives it an intrinsic timing memory capability, which can dynamically couple the convolution features at the current time with the hidden states of the previous time steps. For example, when the air conditioner is in the variable frequency speed regulation stage, its energy consumption curve will show periodic oscillation characteristics. RNN can couple and analyze the compressor frequency change characteristics captured in the previous time window with the power fluctuation characteristics at the current time through continuous updating of the hidden state. This timing memory characteristic enables the system to identify the feature propagation pattern generated by the air conditioning system in a specific operating stage (such as refrigeration mode switching), and these patterns often have a strong correlation with the time-delayed abnormal energy consumption of the subsequent ventilation device.

[0044] Specifically, the step S132-2, the first device energy consumption time sequence pattern message passing encoding vector is obtained by the message passing based on the space-time credibility adjustment of the sequence distribution of the sequence passing first device energy consumption local time domain correlation initial feature vector, including: calculating the first device energy consumption local time domain correlation feature time domain credibility adjustment coefficient of each sequence passing first device energy consumption local time domain correlation initial feature vector in the sequence distribution of the sequence passing first device energy consumption local time domain correlation initial feature vector; the first device energy consumption local time domain correlation feature space domain credibility adjustment coefficient of each sequence passing first device energy consumption local time domain correlation initial feature vector in the sequence distribution of the sequence passing first device energy consumption local time domain correlation initial feature vector; based on the first device energy consumption local time domain correlation feature time domain credibility adjustment coefficient and the first device energy consumption local time domain correlation feature space domain credibility adjustment coefficient of each sequence passing first device energy consumption local time domain correlation initial feature vector, the first device energy consumption local time domain correlation feature message passing space-time collaborative adjustment coefficient of each sequence passing first device energy consumption local time domain correlation initial feature vector is constructed; based on the first device energy consumption local time domain correlation feature message passing space-time collaborative adjustment coefficient of each sequence passing first device energy consumption local time domain correlation initial feature vector, the message passing aggregation based on the space-time modulation of each sequence passing first device energy consumption local time domain correlation initial feature vector is carried out to obtain the first device energy consumption time sequence pattern message passing encoding vector.

[0045] Specifically, in the embodiment of the application, the first device energy consumption local time domain correlation feature time domain credibility adjustment coefficient of each sequence passing first device energy consumption local time domain correlation initial feature vector in the sequence distribution of the sequence passing first device energy consumption local time domain correlation initial feature vector is calculated. This process can be represented by the following formula:

[0046]

[0047] Where, v i is the i th first device energy consumption local time domain correlation feature vector in the sequence distribution of the first device energy consumption local time domain correlation feature vector, h i is the i th sequence passing first device energy consumption local time domain correlation initial feature vector in the sequence distribution of the sequence passing first device energy consumption local time domain correlation initial feature vector, W 1i is the corresponding learnable weight matrix of h i , W 2i is the corresponding learnable weight matrix of v i , and α and β are trainable weighting hyperparameters respectively, tanh is the hyperbolic tangent activation function, is the matrix multiplication, v t is the score weight vector, is hi The corresponding first device energy consumption local time domain correlation feature energy fraction, softmax is a normalization function, is h i The corresponding first device energy consumption local time domain correlation feature time domain credibility adjustment coefficient.

[0048] Correspondingly, the device energy consumption time series data often has significant temporal heterogeneity characteristics. For example, the running state of the air conditioning system during the morning and evening peak periods will produce sharp fluctuations due to sudden changes in environmental temperature, while the energy consumption curve during the regular period presents a smooth characteristic. If a uniform weighted time series processing method is used, it is difficult to distinguish the data quality difference of different time periods. When the sensor accidental noise or device short-time overload causes the local time period feature vector credibility to drop sharply, if all time step features are processed without discrimination, noise amplification effect is easily introduced in the subsequent cross-device reasoning link, causing abnormal detection misjudgment. In this application, the first device energy consumption local time domain correlation feature time domain credibility adjustment coefficient of each sequence transfer first device energy consumption local time domain correlation initial feature vector in the sequence distribution of the sequence transfer first device energy consumption local time domain correlation initial feature vector is calculated. That is, the time domain credibility adjustment coefficient is introduced to dynamically adjust the weight of the sequence transfer first device energy consumption local time domain correlation initial feature vector. This mechanism constructs an adaptive evaluation model based on the attention mechanism, which can capture the data quality fluctuations of different running stages of the air conditioning system: in the stable running stage of the device, the feature vector obtains a high confidence score due to the data smoothness; and in the compressor start-stop instant or the sensor accidental packet loss period, the system automatically reduces the weight coefficient of the corresponding time step. For example, when the air conditioner filter accumulates dust, causing the fan power to gradually rise, the time domain credibility adjustment coefficient will enhance the weight of the middle running data, suppress the interference information of the initial stable stage and the late abnormal saturation stage, and accurately focus on the key time window of the fault feature propagation.

[0049] Specifically, in the embodiments of the present application, the first device energy consumption local time domain correlation feature spatial domain credibility adjustment coefficient of each sequence transfer first device energy consumption local time domain correlation initial feature vector in the sequence distribution of the sequence transfer first device energy consumption local time domain correlation initial feature vector is calculated. This process can be represented by the following formula:

[0050]

[0051] Where h i is the i-th sequence transfer first device energy consumption local time domain correlation initial feature vector in the sequence distribution of the sequence transfer first device energy consumption local time domain correlation initial feature vector, h i T is the transpose vector of h i , is the square of the F norm, s(hi ) is h i the corresponding first device energy consumption local time domain correlation feature space similarity score value, exp is the exponential function value with the natural constant e as the base, is h i the corresponding first device energy consumption local time domain correlation feature space confidence adjustment coefficient.

[0052] It can be understood that when analyzing the first device energy consumption feature, not only the time sequence information is important, but also the spatial structure relationship of the device in the whole system will have an impact on its energy consumption feature. Calculating the space confidence adjustment coefficient can take into account the spatial relationship and structural importance of the node (device) in the structured network, and comprehensively analyze the energy consumption feature. Therefore, by calculating the first device energy consumption local time domain correlation feature space confidence adjustment coefficient of each sequence transmission first device energy consumption local time domain correlation initial feature vector in the sequence distribution of the sequence transmission first device energy consumption local time domain correlation initial feature vector. That is, a space confidence adjustment coefficient is assigned to each vector in the sequence distribution of the sequence transmission first device energy consumption local time domain correlation initial feature vector, similar to the spatial confidence constraint factor assigning a spatial confidence quantitative value to the node, so as to represent the importance and information propagation potential of the device (node) in the energy consumption feature space structure.

[0053] Specifically, in the embodiment of the present application, based on the first device energy consumption local time domain correlation feature time domain confidence adjustment coefficient and the first device energy consumption local time domain correlation feature space confidence adjustment coefficient of each sequence transmission first device energy consumption local time domain correlation initial feature vector, the first device energy consumption local time domain correlation feature message transmission space-time collaborative adjustment coefficient of each sequence transmission first device energy consumption local time domain correlation initial feature vector is constructed. This process can be represented by the following formula:

[0054]

[0055] wherein, is h i the corresponding first device energy consumption local time domain correlation feature time domain confidence adjustment coefficient, is h i the corresponding first device energy consumption local time domain correlation feature space confidence adjustment coefficient, ω1 and ω2 are respectively and the contribution adjustment weighting coefficient of and sigmoid is the sigmoid function, is h i the corresponding first device energy consumption local time domain correlation feature message transmission space-time collaborative adjustment coefficient.

[0056] Accordingly, the propagation of device anomaly features is simultaneously constrained by both temporal dynamics and spatial structure. The bearing wear fault of the air conditioning system presents a gradual energy consumption escalation feature in the time dimension, and its spatial path through the duct network topology affects the end devices, with primary and secondary differences: the ventilation equipment closer to the fault source will preferentially exhibit anomalies, while the remote devices may exhibit delayed responses due to pipe network damping effects. Traditional single-dimensional confidence evaluation methods cannot capture this spatio-temporal coupling characteristic, and are prone to equate early anomaly features of key nodes with random fluctuations of edge nodes, resulting in dilution of core propagation path information during cross-device reasoning. Therefore, based on the first device energy consumption local time domain correlation feature time domain confidence adjustment coefficient and the first device energy consumption local time domain correlation feature space domain confidence adjustment coefficient, a first device energy consumption local time domain correlation feature message passing space-time collaborative adjustment coefficient of the sequence transmission first device energy consumption local time domain correlation initial feature vector is constructed. This step uses a gating fusion mechanism to dynamically balance the contribution weights of the time domain confidence adjustment coefficient (such as the diagnostic value of transient features during the compressor start-stop phase) and the space domain confidence adjustment coefficient (such as the structural importance of the host node in the air duct network). When processing the heating, ventilation and air conditioning system, the model automatically increases the collaborative weights of a specific time window (leakage feature significant period) and a key node (main air valve controller) at the initial stage of refrigerant leakage, so that the periodic pressure fluctuation feature caused by the host bearing wear can not only track its amplitude growth trend in the time dimension, but also strengthen its propagation intensity in the space dimension at the core node, forming an anomaly propagation graph with spatio-temporal joint sensitivity.

[0057] More specifically, in the embodiments of the present application, based on the first device energy consumption local time domain correlation feature message passing space-time collaborative adjustment coefficient of each sequence transmission first device energy consumption local time domain correlation initial feature vector, a time-space modulation based message passing aggregation is performed on each sequence transmission first device energy consumption local time domain correlation initial feature vector to obtain the first device energy consumption time sequence pattern message passing encoding vector, including:

[0058] The curvature-driven space-time multi-mode fusion geometric regularization is performed on the first device energy consumption local time domain correlation feature message passing space-time collaborative adjustment coefficient of each sequence transmission first device energy consumption local time domain correlation initial feature vector to obtain the first device energy consumption local time domain correlation feature message passing space-time collaborative optimization adjustment coefficient of each sequence transmission first device energy consumption local time domain correlation initial feature vector. This process can be represented by the following formula:

[0059]

[0060]

[0061] wherein, is h ia corresponding first device energy consumption local time domain correlation feature time domain credibility adjustment coefficient, is h i a corresponding first device energy consumption local time domain correlation feature spatial domain credibility adjustment coefficient, is h i a corresponding first device energy consumption local time domain correlation feature message passing space-time coordination adjustment coefficient, G i is h i a corresponding first device energy consumption local time domain correlation feature constant curvature space mapping value, sin is a sine function, π is a circular constant, R i is h i a corresponding first device energy consumption local time domain correlation feature hyperspherical progressive representation value, is h i a corresponding first device energy consumption local time domain correlation feature space-time coupling metric weight factor, is an optimized first device energy consumption local time domain correlation feature message passing space-time coordination optimization adjustment coefficient;

[0062] a first device energy consumption local time domain correlation feature message passing space-time coordination optimization adjustment coefficient based on the respective sequence transmission first device energy consumption local time domain correlation initial feature vectors, the respective sequence transmission first device energy consumption local time domain correlation initial feature vectors are calculated by position weighting and sum to obtain the first device energy consumption time sequence mode message passing encoding vector, which can be represented by the following formula:

[0063]

[0064] wherein h i is the i-th sequence transmission first device energy consumption local time domain correlation initial feature vector in the sequence distribution of the sequence transmission first device energy consumption local time domain correlation initial feature vectors, is an optimized first device energy consumption local time domain correlation feature message passing space-time coordination optimization adjustment coefficient, t is the number of vectors in the sequence distribution of the sequence transmission first device energy consumption local time domain correlation initial feature vectors, and z is the first device energy consumption time sequence mode message passing encoding vector.

[0065] In particular, the first device energy consumption local time domain correlation feature time domain credibility adjustment coefficient and the first device energy consumption local time domain correlation feature spatial domain credibility adjustment coefficient respectively represent the respective sequence transmission first device energy consumption local time domain correlation initial feature vectors h iUnder the condition of time and space dimension information weight distribution, if the time and space dimensions are regarded as the global time and space single degree of freedom modeling paradigm, the negative correlation effect in the single mode attention mechanism needs to be avoided to prevent the non-flat representation of the time and space composite manifold caused by the negative correlation effect, so as to prevent the representation performance of the multi-dimensional fusion space from being degraded.

[0066] Based on this, first, the first device energy consumption local time domain correlation initial feature vector h i The corresponding first device energy consumption local time domain correlation feature time domain credibility adjustment coefficient And the first device energy consumption local time domain correlation feature space domain credibility adjustment coefficient The first device energy consumption local time domain correlation feature constant curvature space mapping value G i And the first device energy consumption local time domain correlation feature hyperspherical progressive representation value R i :

[0067]

[0068] Then, the metric tensor reference of the collaborative space is taken as the space-time joint constraint reference of message passing, and the first device energy consumption local time domain correlation initial feature vector h i The first device energy consumption local time domain correlation feature message passing space-time collaborative adjustment coefficient Iterative optimization is performed:

[0069]

[0070] When the constraint condition Holds, the single mode representation under the space-time dimension And Will tend to the quasi-flat space-time collaborative attractor, thereby embodying the quasi-Euclidean characteristics of the message passing adjustment coefficient in the composite manifold. The method realizes the geometric preservation optimization of the fusion space through the regularization compensation of the single mode negative curvature, and significantly improves the multi-mode fusion representation ability of the space-time collaborative adjustment coefficient.

[0071] Finally, based on the first device energy consumption local time domain correlation feature message passing space-time collaborative optimization adjustment coefficient of the first device energy consumption local time domain correlation initial feature vector of each sequence, the first device energy consumption local time domain correlation initial feature vector of each sequence is calculated by position weighting to obtain the first device energy consumption time sequence mode message passing encoding vector. In this way, through the weighted sum calculation, the model can adaptively enhance the abnormal features transmitted by the key path (main air pipe pressure monitoring point) of the air conditioning system in a specific operation stage (such as the end of the refrigerant charging period), so that the subsequent cross-device mapping stage generates a ventilation device prediction sequence that can accurately reflect the influence of such space-time coupled anomalies.

[0072] In step S140, the time queue of the energy consumption data of the second device is energy consumption time series coded based on the first device energy consumption time series pattern message passing encoding vector to obtain a generated time queue of the energy consumption data of the second device. Figure 4 The flowchart of step S140 in the method for building intelligent operation and maintenance based on a large model according to the embodiment of the present application is shown. Specifically, in the embodiment of the present application, as shown in the figure, the time queue of the energy consumption data of the second device is energy consumption time series coded based on the first device energy consumption time series pattern message passing encoding vector to obtain a generated time queue of the energy consumption data of the second device, which includes: S141, performing energy consumption time series mapping inference coding on the time queue of the energy consumption data of the second device based on the first device energy consumption time series pattern message passing encoding vector to obtain a second device energy consumption time series pattern inference encoding vector; and S142, performing energy consumption time series feature decoding prediction on the second device energy consumption time series pattern inference encoding vector to obtain a generated time queue of the energy consumption data of the second device. Figure 4

[0073] Specifically, in the embodiment of the present application, S141, performing energy consumption time series mapping inference coding on the time queue of the energy consumption data of the second device based on the first device energy consumption time series pattern message passing encoding vector to obtain a second device energy consumption time series pattern inference encoding vector includes: constructing an attribute semantic association matrix between the first device and the second device; and performing energy consumption time series feature cross-semantic space mapping on the first device energy consumption time series pattern message passing encoding vector using the attribute semantic association matrix to obtain the second device energy consumption time series pattern inference encoding vector.

[0074] It should be understood that, considering that the implicit association between devices often contains complex physical coupling and energy transfer mechanisms. For example, the supply air temperature of an air conditioning system and the fan speed of a ventilation device belong to different device parameters, but they form a strong association through air duct static pressure balance; while the illumination adjustment of a lighting system and the air conditioning load may only have a weak energy coupling relationship. If the association analysis method based on rules or statistics is limited by the static nature and one-sidedness of the manually defined association rules, it is difficult to adaptively capture such dynamic semantic associations across systems, resulting in frequent misjudgment phenomena such as false association of air conditioner compressor failure to lighting system energy consumption fluctuations. Based on this, in the technical solution of the present application, an attribute semantic association matrix between the first device and the second device is constructed. That is, the attribute semantic association matrix is a quantitative representation of the relationship between devices, which can provide a basic data structure for subsequent various analysis tasks. Whether it is device fault diagnosis, energy consumption prediction, or optimization of device operation strategy, it is necessary to first understand the association relationship between devices, and the attribute semantic association matrix can clearly present these relationships and provide a basis for further calculation and inference. ​

[0075] Specifically, the attribute semantic association matrix between the first device and the second device is constructed, and the specific implementation process is as follows:

[0076] First of all, data collection and arrangement. With the help of intelligent platforms such as building automation system (BAS) and energy management system (EMS) deployed in the building, the running parameters and energy consumption time series data of the first device and the second device are continuously collected. These data cover the basic attributes of the device (such as device type, specification and model, etc.), running state parameters (such as temperature, pressure, speed, etc.) and energy consumption data (energy consumption value at different time points). At the same time, the spatial position information of the device is collected, and its layout in the building is clarified, because the spatial position may affect the energy transfer and interaction between devices. For example, air conditioning equipment and ventilation equipment on the same floor and close distance may have closer association.

[0077] Then, feature engineering processing is carried out. The collected device running parameters and energy consumption data are cleaned to remove outliers and noise data, ensuring data quality. Then, key features are extracted, such as the change trend of device energy consumption, periodicity characteristics, and correlation characteristics with other devices. For different types of devices, appropriate features are selected according to their working principles and characteristics. For example, for air conditioning systems, refrigerating capacity, heating power, compressor running frequency, etc. are important features; for lighting systems, illumination, on-time, etc. are key features. These features are quantified and standardized to make the features of different devices comparable.

[0078] Then, the similarity between devices is calculated. Based on the extracted and processed features, the similarity between the first device and the second device is calculated using appropriate similarity measurement methods. Common methods include Euclidean distance, cosine similarity, etc. Taking Euclidean distance as an example, the distance between two devices in the feature space is calculated. The closer the distance, the higher the similarity between the devices, and the greater the potential association possibility. For example, if the Euclidean distance between the first air conditioning device and the second ventilation device in energy consumption change trend, running time and other features is small, it indicates that they have similarity in these aspects, and there may be some association.

[0079] Finally, the attribute semantic association matrix is constructed. According to the calculated similarity between devices, a two-dimensional matrix is constructed. The rows and columns of the matrix correspond to the first device and the second device respectively, and the value of the matrix element represents the degree of association between the two devices. The device pair with high similarity has a larger value of the corresponding element in the matrix; the device pair with low similarity has a smaller value of the corresponding element. For example, if the first device and the second device have high similarity in multiple characteristics, the element value at the corresponding position in the matrix may be close to 1; if the two devices differ greatly in most characteristics, the corresponding element value may be close to 0. The attribute semantic association matrix constructed through the above implementation process can effectively quantify the complex relationship between the first device and the second device, provide strong support for subsequent use of the matrix to map the energy consumption time sequence pattern message passing encoding vector of the first device to the energy consumption time sequence feature across the semantic space, and further improve the accuracy and reliability of device energy consumption anomaly detection in building intelligent operation and maintenance.

[0080] Correspondingly, considering that the abnormal propagation between devices often involves complex nonlinear association and mapping difficulties in heterogeneous semantic spaces. If simple linear weighting or statistical correlation analysis is used, it is difficult to model the non-uniform influence of air conditioner system wind pressure fluctuations on ventilation equipment through a specific topological path. For example, when the air conditioner host causes the static pressure to rise due to filter blockage, the influence degree on each floor ventilation equipment not only depends on the physical distance, but also closely related to the air dynamics attributes such as air duct branch angle, local resistance coefficient, etc. Due to the lack of dynamic mapping mechanism across semantic spaces, the existing technology often simplifies the association between devices as fixed weight transfer, resulting in that the decay law and superposition effect of air conditioner abnormal characteristics in the pipeline network cannot be accurately reflected when predicting the energy consumption of ventilation equipment. Based on this, the attribute semantic association matrix is used to map the energy consumption time sequence pattern message passing encoding vector of the first device to the energy consumption time sequence feature across the semantic space to obtain the second device energy consumption time sequence pattern reasoning encoding vector. In this way, after obtaining the second device energy consumption time sequence pattern reasoning encoding vector, the energy consumption of the second device can be predicted based on it. Since the encoding vector integrates the energy consumption features of the first device and the attribute semantic association information between devices, the prediction result can more comprehensively reflect the actual energy consumption of the second device, and has higher accuracy compared with the traditional method of predicting based on only the historical data of the second device itself. For example, when predicting the energy consumption of ventilation equipment, considering the energy consumption characteristics of air conditioner devices and their association relationship, the energy consumption of ventilation equipment under different working conditions can be more accurately predicted.

[0081] Specifically, in the embodiment of the present application, the step S142, the energy consumption time sequence feature decoding prediction of the second device energy consumption time sequence pattern inference encoding vector to obtain the generation time queue of the energy consumption data of the second device, comprises: using an RNN-based decoder to perform energy consumption time sequence feature decoding prediction on the second device energy consumption time sequence pattern inference encoding vector to obtain the generation time queue of the energy consumption data of the second device. It should be understood that the second device energy consumption time sequence pattern inference encoding vector essentially contains energy consumption feature information related to time series. RNN has a natural advantage in processing time series data, and its internal loop structure allows information to be transmitted between different time steps, which can effectively capture the time dependence in the data. In energy consumption data, the current energy consumption state of the device is often related to the past energy consumption state, for example, the preheating process of the device, the gradual change of the load, etc. RNN can well model this dynamic change, so it is suitable for decoding prediction of the encoding vector. In particular, the energy consumption pattern of a building device is usually complex, and there may be periodic, seasonal changes, and sudden abnormal fluctuations, etc. RNN can learn these complex patterns, remember the past information by constantly updating the hidden state, and make predictions based on the current input and the remembered information. The RNN-based decoder can decode the complex energy consumption features contained in the second device energy consumption time sequence pattern inference encoding vector, thereby more accurately restoring the time series of the energy consumption data.

[0082] Specifically, the process of using an RNN-based decoder to perform energy consumption time sequence feature decoding prediction on the second device energy consumption time sequence pattern inference encoding vector to obtain the generation time queue of the energy consumption data of the second device is as follows:

[0083] Before decoding prediction, it is necessary to ensure the accuracy and effectiveness of the second device energy consumption time sequence pattern inference encoding vector. The encoding vector is obtained by constructing an attribute semantic association matrix between the first device and the second device, and performing cross-semantic space mapping on the first device energy consumption time sequence pattern message passing encoding vector, which integrates the energy consumption features of the first device and the semantic association information between devices.

[0084] After preparing the encoding vector, it can be input into the RNN-based decoder. RNN has a unique advantage in processing time series data, and its internal loop structure can allow information to be transmitted between different time steps, thereby effectively capturing the time dependence in the data, which is crucial for analyzing the energy consumption data of building devices over time.

[0085] Inside the RNN decoder, the hidden state is first initialized. The hidden state is like a "memory unit" of RNN, which saves the information of the previous time step for use in subsequent calculations. The initial hidden state is usually set to a zero vector or an initial value set according to experience.

[0086] As the encoding vectors are inputted one by one in time steps, the RNN updates the hidden state step by step. At each time step, the RNN combines the current input encoding vector and the hidden state of the last time step for calculation. Specifically, the input and the hidden state are linearly combined and transformed through a specific weight matrix and an activation function to obtain a new hidden state. For example, using an activation function such as tanh or ReLU, the result of linear combination is nonlinearly transformed, so that the RNN can learn more complex time series patterns.

[0087] While updating the hidden state, the RNN decoder also generates a predicted output based on the current hidden state. The predicted output is an estimate of the energy consumption value of the second device at the current time step. This predicted output is based on the energy consumption time series features learned by the RNN and the information of the previous time steps.

[0088] After multiple time steps of calculation, the RNN decoder gradually generates a series of predicted outputs, which are arranged in chronological order to form the generated time queue of the second device's energy consumption data. To improve the accuracy and stability of the prediction, a large amount of historical energy consumption data is used as training samples when training the RNN decoder. By continuously adjusting the weight parameters in the RNN, the predicted output of the decoder is made as close as possible to the actual energy consumption data. Common training methods include the backpropagation algorithm, which can calculate the prediction error and update the weights according to the gradient of the error, so that the RNN gradually learns the inherent rules of the energy consumption data.

[0089] In practical applications, the RNN-based decoder can also be optimized and extended according to specific needs. For example, increasing the number of RNN layers to form a deep RNN to learn more complex features, or using improved RNN structures such as LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit) to better handle long-term dependencies and avoid gradient vanishing or gradient explosion, thereby further improving the accuracy of energy consumption prediction.

[0090] In step S150, the energy consumption data time series distribution difference value between the time queue of the second device's energy consumption data and the generated time queue of the second device's energy consumption data is calculated. Accordingly, since the generated time queue of the second device's energy consumption data is predicted by the model, and the actual energy consumption data time queue is observed, the difference value between the two can directly reflect the deviation of the predicted results generated by the RNN-based decoder from the actual situation, thereby evaluating the accuracy of the energy consumption prediction model.

[0091] Specifically, first, a suitable calculation method is selected to measure the difference between the two. Commonly used methods include mean square error (MSE), mean absolute error (MAE), and dynamic time warping (DTW) distance, etc.

[0092] Taking the mean square error as an example, the calculation process is relatively intuitive. Subtract the energy consumption data at the corresponding time points in the two time queues to get the error value at each time point, then square the error values to amplify the impact of the error and eliminate the situation where positive and negative errors cancel each other out. Then, add up the squared errors of all time points and divide by the total number of time points to get the mean square error value. The larger this value, the greater the difference between the two time queues; otherwise, the smaller the difference.

[0093] The calculation of the mean absolute error is to directly calculate the absolute value of the difference between the energy consumption data at the corresponding time points of the two time queues, then add up all the absolute value errors and divide by the total number of time points. It can more intuitively reflect the average error between the predicted value and the actual value.

[0094] The calculation of the dynamic time warping distance is relatively complex, and it is suitable for the case where the two time series have local time shifts on the time axis. This method finds the best matching path between the two time series, calculates the cumulative distance along this path, and takes it as the difference value.

[0095] In actual operation, according to the specific needs and data characteristics, select the appropriate calculation method. After calculating the difference value, this value will be an important basis for judging whether the energy consumption of the second device is abnormal. It can quantify the deviation between the predicted data and the actual data, provide objective data support for subsequent judgment of whether the device energy consumption is abnormal based on the preset threshold, and help to discover the possible faults or abnormal operation state of the device in time.

[0096] In step S160, based on the comparison between the time-series distribution difference value of the energy consumption data and a preset threshold, it is confirmed whether there is an anomaly in the energy consumption of the second device. It should be understood that the preset threshold is a reference value set based on factors such as the device's historical energy consumption data, operating characteristics, and relevant standards. The time-series distribution difference value of the energy consumption data reflects the degree of deviation between predicted and actual energy consumption. By comparing it with the preset threshold, the judgment of energy consumption anomalies can be quantified, avoiding the uncertainty of subjective judgment and providing an objective and measurable evaluation standard. Specifically, the energy consumption of a device usually exhibits certain regularity and stability. When the time-series distribution difference value of the energy consumption data exceeds the preset threshold, it indicates a significant deviation between actual and expected energy consumption. This is likely caused by factors such as device malfunction, changes in the operating environment, or improper human operation. Timely comparison can identify these potential problems as early as possible, allowing for appropriate measures to be taken to prevent further deterioration. By confirming whether there are any abnormalities in the energy consumption of the second device, potential faults or other problems can be detected in a timely manner, so that maintenance personnel can carry out timely inspections and maintenance, ensuring the normal operation of the equipment, avoiding adverse consequences such as production interruptions and service quality degradation caused by equipment failures, and ensuring the stability and reliability of the entire system.

[0097] In summary, the intelligent building operation and maintenance method based on a large model, as described in this application, is explained. It employs data processing and analysis algorithms based on large model artificial intelligence. This technical concept constructs an energy consumption inference mechanism for dynamic relationships between devices using a large model. First, the energy consumption time series of the first device is subjected to one-dimensional convolutional encoding to extract local temporal correlation features of its operating modes. Then, temporal propagation encoding is used to capture the dynamic evolution of these features along the time axis. Subsequently, based on the semantic association matrix of device attributes, the energy consumption features of the first device are nonlinearly mapped to the semantic space of the second device, enabling the energy consumption prediction of the second device to incorporate the influence of the operating status of upstream devices. Finally, by comparing the temporal distribution differences between the predicted sequence and the actual monitoring data, the implicit energy consumption fluctuations of the second device indirectly caused by the anomalies of the first device are identified. This solution breaks through the limitations of traditional isolated modeling. By transmitting dynamic features across the semantic space of devices, it accurately captures the energy transfer anomalies caused by the failure of the first device to the second device. It integrates the originally scattered threshold alarm events into a complete hidden fault propagation chain tracing analysis, which significantly reduces false alarms and missed alarms caused by the lack of modeling of the coupling of operating conditions between devices, improves the accuracy and timeliness of anomaly detection, and reduces the workload of manual verification.

[0098] Figure 5 This is a block diagram of a building intelligent operation and maintenance system based on a large model, according to an embodiment of this application. Figure 5As shown, the big model-based building intelligent operation and maintenance system 100 according to the embodiment of the present application comprises: a first device energy consumption data acquisition module 110, configured to acquire a time queue of energy consumption data of a first device in a building; a second device energy consumption data acquisition module 120, configured to acquire a time queue of energy consumption data of a second device in the building; a first device energy consumption time sequence encoding module 130, configured to perform local time sequence-based recursive message passing encoding on the time queue of energy consumption data of the first device to obtain a first device energy consumption time sequence pattern message passing encoding vector; a second device energy consumption time sequence encoding module 140, configured to perform energy consumption time sequence encoding and decoding on the time queue of energy consumption data of the second device based on the first device energy consumption time sequence pattern message passing encoding vector to obtain a generated time queue of energy consumption data of the second device; an energy consumption data difference calculation module 150, configured to calculate an energy consumption data time sequence distribution difference value between the time queue of energy consumption data of the second device and the generated time queue of energy consumption data of the second device; and an abnormality judgment module 160, configured to confirm whether the energy consumption of the second device is abnormal based on a comparison between the energy consumption data time sequence distribution difference value and a preset threshold value.

[0099] Here, those skilled in the art can understand that the specific operations of each step in the above big model-based building intelligent operation and maintenance system have been described in detail above with reference to the description of the big model-based building intelligent operation and maintenance method of Figures 1 to 4 Therefore, the repeated description thereof will be omitted.

[0100] As described above, the big model-based building intelligent operation and maintenance system 100 according to the embodiment of the present disclosure can be implemented in various wireless terminals, such as a server with a big model-based building intelligent operation and maintenance algorithm, etc. In one possible implementation, the big model-based building intelligent operation and maintenance system 100 according to the embodiment of the present disclosure can be integrated into a wireless terminal as a software module and / or a hardware module. For example, the big model-based building intelligent operation and maintenance system 100 can be a software module in the operating system of the wireless terminal, or can be an application program developed for the wireless terminal; of course, the big model-based building intelligent operation and maintenance system 100 can also be one of the many hardware modules of the wireless terminal.

[0101] Alternatively, in another example, the big model-based building intelligent operation and maintenance system 100 and the wireless terminal can also be separate devices, and the big model-based building intelligent operation and maintenance system 100 can be connected to the wireless terminal through a wired and / or wireless network, and transmit interactive information in an agreed data format.

[0102] In light of the above, it is to be understood that the above description is intended to be illustrative only and not restrictive. It is to be understood that the above examples are to be construed as merely illustrative of the principles and operation of the application, and that modifications and variations thereto can occur to those skilled in the art.

Claims

1. A large model-based building intelligent operation and maintenance method, characterized in that, The method comprises the following steps: obtaining a time queue of energy consumption data of a first device in a building; obtaining a time queue of energy consumption data of a second device in the building; performing local time sequence-based recursive message passing coding on the time queue of energy consumption data of the first device to obtain a first device energy consumption time sequence pattern message passing coding vector, comprising: performing one-dimensional convolution coding-based energy consumption time sequence pattern feature extraction on the time queue of energy consumption data of the first device to obtain a sequence distribution of first device energy consumption local time domain correlation feature vectors; and performing time sequence propagation recursive structured coding on the sequence distribution of first device energy consumption local time domain correlation feature vectors to obtain the first device energy consumption time sequence pattern message passing coding vector; performing energy consumption time sequence coding on the time queue of energy consumption data of the second device based on the first device energy consumption time sequence pattern message passing coding vector to obtain a generated time queue of energy consumption data of the second device, comprising: performing energy consumption time sequence mapping inference coding on the time queue of energy consumption data of the second device based on the first device energy consumption time sequence pattern message passing coding vector to obtain a second device energy consumption time sequence pattern inference coding vector; and performing energy consumption time sequence feature decoding prediction on the second device energy consumption time sequence pattern inference coding vector to obtain the generated time queue of energy consumption data of the second device; calculating an energy consumption data time sequence distribution difference value between the time queue of energy consumption data of the second device and the generated time queue of energy consumption data of the second device; based on a comparison between the energy consumption data time sequence distribution difference value and a preset threshold, confirming whether the energy consumption of the second device is abnormal.

2. The large model-based building intelligentization operation and maintenance method according to claim 1, characterized in that, The method comprises the following steps: performing sequence coding on the sequence distribution of first device energy consumption local time domain correlation feature vectors based on a recurrent neural network to obtain a sequence distribution of sequence transmission first device energy consumption local time domain correlation initial feature vectors; performing message passing based on space-time credibility adjustment on the sequence distribution of sequence transmission first device energy consumption local time domain correlation initial feature vectors to obtain the first device energy consumption time sequence pattern message passing coding vector.

3. The large model-based building intelligentization operation and maintenance method according to claim 2, characterized in that, The method comprises the following steps: calculating a first device energy consumption local time domain correlation feature time domain credibility adjustment coefficient of each sequence transmission first device energy consumption local time domain correlation initial feature vector in the sequence distribution of sequence transmission first device energy consumption local time domain correlation initial feature vectors; calculating a first device energy consumption local time domain correlation feature space domain credibility adjustment coefficient of each sequence transmission first device energy consumption local time domain correlation initial feature vector in the sequence distribution of sequence transmission first device energy consumption local time domain correlation initial feature vectors; Based on the first device energy consumption local time domain correlation feature temporal reliability adjustment coefficient and the first device energy consumption local time domain correlation feature spatial reliability adjustment coefficient of the first device energy consumption local time domain correlation initial feature vector transmitted in each sequence, the first device energy consumption local time domain correlation feature message transmission spatiotemporal coordination adjustment coefficient of the first device energy consumption local time domain correlation initial feature vector transmitted in each sequence is constructed; Based on the first device energy consumption local time domain association initial feature vector of each sequence transmission, the first device energy consumption local time domain association initial feature vector of each sequence transmission is subjected to message transmission aggregation based on time-conditioning mechanism to obtain the first device energy consumption time sequence pattern message transmission encoding vector.

4. The large model-based building intelligentization operation and maintenance method according to claim 3, characterized in that, Based on the first device energy consumption local time domain correlation initial feature vector transmitted in each sequence, the first device energy consumption local time domain correlation feature message transmission spatiotemporal coordination adjustment coefficient is used to perform time-based mechanism-based message transmission aggregation on the first device energy consumption local time domain correlation initial feature vector transmitted in each sequence to obtain the first device energy consumption time sequence pattern message transmission encoding vector, including: Curvature-driven spatiotemporal multimodal fusion geometric regularization is applied to the first device energy consumption local temporal domain correlation initial feature vector of each sequence to obtain the first device energy consumption local temporal domain correlation feature message transmission spatiotemporal collaborative optimization adjustment coefficient of each sequence to obtain the first device energy consumption local temporal domain correlation initial feature vector of each sequence. Based on the first device energy consumption local temporal domain association initial feature vector of each sequence, the first device energy consumption local temporal domain association feature message transmission spatiotemporal collaborative optimization adjustment coefficient is used to perform position-weighted sum calculation on the first device energy consumption local temporal domain association initial feature vector of each sequence to obtain the first device energy consumption time sequence pattern message transmission encoding vector.

5. The large model-based building intelligentization operation and maintenance method according to claim 4, characterized in that, Based on the energy consumption time-series pattern message passing encoding vector of the first device, energy consumption time-series mapping inference encoding is performed on the time queue of the energy consumption data of the second device to obtain the energy consumption time-series pattern inference encoding vector of the second device, including: Construct an attribute semantic association matrix between the first device and the second device; The attribute semantic association matrix is ​​used to perform energy consumption time-series feature cross-semantic space mapping on the first device energy consumption time-series pattern message passing encoding vector to obtain the second device energy consumption time-series pattern inference encoding vector.

6. The large model-based building intelligentization operation and maintenance method according to claim 5, characterized in that, The process of performing energy consumption time-series feature decoding and prediction on the energy consumption time-series pattern inference encoding vector of the second device to obtain the generation time queue of energy consumption data of the second device includes: using an RNN-based decoder to perform energy consumption time-series feature decoding and prediction on the energy consumption time-series pattern inference encoding vector of the second device to obtain the generation time queue of energy consumption data of the second device.

7. A large model-based building intelligent operation and maintenance system, characterized in that, include: The first equipment energy consumption data acquisition module is used to acquire the time queue of energy consumption data of the first equipment in the building; The energy consumption data acquisition module for the second equipment is used to acquire the time queue of energy consumption data of the second equipment in the building. The first device energy consumption time-series encoding module is used to perform recursive message passing encoding based on local time sequence on the time queue of the energy consumption data of the first device to obtain the first device energy consumption time-series pattern message passing encoding vector, including: performing energy consumption time-series pattern feature extraction based on one-dimensional convolutional encoding on the time queue of the energy consumption data of the first device to obtain the sequence distribution of the first device energy consumption local time-domain associated feature vector; and performing time-series propagation recursive structured encoding on the sequence distribution of the first device energy consumption local time-domain associated feature vector to obtain the first device energy consumption time-series pattern message passing encoding vector. The second device energy consumption time-series encoding module is used to perform energy consumption time-series encoding and decoding on the time queue of the energy consumption data of the second device based on the first device energy consumption time-series mode message passing encoding vector to obtain the generation time queue of the energy consumption data of the second device. This includes: performing energy consumption time-series mapping inference encoding on the time queue of the energy consumption data of the second device based on the first device energy consumption time-series mode message passing encoding vector to obtain the second device energy consumption time-series mode inference encoding vector; and performing energy consumption time-series feature decoding prediction on the second device energy consumption time-series mode inference encoding vector to obtain the generation time queue of the energy consumption data of the second device. The energy consumption data difference calculation module is used to calculate the energy consumption data time sequence distribution difference value between the time queue of the energy consumption data of the second device and the generation time queue of the energy consumption data of the second device. The anomaly detection module is used to confirm whether there is an anomaly in the energy consumption of the second device based on the comparison between the time-series distribution difference value of the energy consumption data and a preset threshold.

Images