Fire-fighting facility fault diagnosis method and system based on digital twinning

Abstract

The invention relates to the field of digital twinning, and particularly discloses a fire-fighting equipment fault diagnosis method and system based on digital twinning, which captures gradual change faults through statistical analysis based on EWMA, and deeply digs transient mode anomalies by using a sequence auto-encoder. A traditional static score fusion mode is abandoned, and a dual-channel score fusion and final confidence judgment mechanism is introduced. The mechanism can dynamically allocate decision weights for two analysis channels of trend and mode according to the real-time determinacy of the diagnosis result of each channel and in combination with fault feature priori knowledge of specific equipment, thereby effectively avoiding interference on overall judgment when information of a single channel is fuzzy, and improving the accuracy of judgment. Therefore, accurate and robust diagnosis of early-stage and hidden faults of fire-fighting equipment is realized, and the alarm reliability is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of digital twinning, and more specifically, to a fire-fighting facility fault diagnosis method and system based on digital twinning. BACKGROUND

[0002] Fire-fighting facilities, as the key defense line to protect people's lives and property safety in modern buildings and industrial sites, are of great importance in terms of long-term operation stability and reliability. These facilities, such as fire pumps, sprinkler systems, etc., are usually in a state of long-term standby or intermittent operation, and their internal components may experience performance degradation or hidden failures due to aging, corrosion or environmental factors. Traditional fire-fighting facility maintenance relies on regular on-site inspections and passive repair responses, which not only have high labor costs and low efficiency, but more importantly, they are difficult to effectively capture early signs of failure. Once an emergency such as a fire occurs, the fire-fighting facilities in a fault state will not be able to start normally or exert their full effectiveness, which may lead to irreparable disastrous consequences.

[0003] With the development of Internet of Things and digital twinning technology, intelligent operation and maintenance of fire-fighting facilities has become a new possibility. By deploying sensors on physical devices and mapping real-time data to digital twin models in virtual space, managers can achieve visual monitoring of facility status. However, existing digital twinning-based monitoring systems are still at a relatively early stage in terms of fault diagnosis, such as setting fixed static threshold alarms. The disadvantage of this method is that the working state of fire-fighting facilities (such as fire pumps) is dynamically changing, and their sensor parameters (such as outlet pressure) under different working conditions (such as standby, inspection, start-up, etc.) are completely different. A single, fixed alarm threshold cannot adapt to such changes, and is prone to generate a large number of false alarms when the device is normally started and stopped, or fail to detect real faults such as gradual pressure changes due to slow leaks in standby state, resulting in low alarm reliability.

[0004] Therefore, an optimized digital twinning-based fire-fighting facility fault diagnosis scheme is desired. SUMMARY

[0005] To solve the above technical problems, the present application is proposed. The embodiments of the present application provide a digital twinning-based fire-fighting facility fault diagnosis method and system.

[0006] According to one aspect of the present application, a digital twinning-based fire-fighting facility fault diagnosis method is provided, which comprises: contextual alignment of the obtained raw sensor data stream and device working condition data to obtain a raw sensor numerical sequence with working condition labels; performing EWMA-based statistical trend deviation analysis on the original sensor value sequence with the working condition label to obtain a trend deviation score; performing sequence autoencoder-based time series pattern reconstruction on the original sensor value sequence with the working condition label to obtain a pattern anomaly score; performing two-channel score fusion and final confidence decision on the trend deviation score and the pattern anomaly score to obtain a final confidence score; performing adaptive alarm generation based on the final confidence score to obtain an adaptive alarm; displaying the adaptive alarm on a display screen through the data twin module.

[0007] According to another aspect of the present application, a fire-fighting facility fault diagnosis system based on digital twinning is provided, which comprises: an original sensor data alignment module configured to perform context alignment on the obtained original sensor data stream and device working condition data to obtain an original sensor value sequence with a working condition label; a statistical trend deviation analysis module configured to perform EWMA-based statistical trend deviation analysis on the original sensor value sequence with the working condition label to obtain a trend deviation score; a time series pattern reconstruction module configured to perform sequence autoencoder-based time series pattern reconstruction on the original sensor value sequence with the working condition label to obtain a pattern anomaly score; a confidence score calculation module configured to perform two-channel score fusion and final confidence decision on the trend deviation score and the pattern anomaly score to obtain a final confidence score; an adaptive alarm module configured to perform adaptive alarm generation based on the final confidence score to obtain an adaptive alarm; a digital twin display module configured to display the adaptive alarm on a display screen through the data twin module.

[0008] Compared with the prior art, the fire-fighting facility fault diagnosis method and system based on digital twinning provided by the present application can capture gradual faults through EWMA-based statistical analysis, and simultaneously use a sequence autoencoder to deeply mine transient pattern anomalies. This discards the traditional static score fusion method, and instead uses a two-channel score fusion and final confidence decision mechanism. This mechanism can dynamically allocate decision weights to the trend and pattern analysis channels according to the real-time determination degree of the diagnosis results of each channel, and in combination with prior knowledge of the fault characteristics of specific devices, effectively avoiding the interference of single-channel information ambiguity on the overall judgment, thereby realizing accurate and robust diagnosis of early and hidden faults of fire-fighting facilities, and improving the reliability of alarms. BRIEF DESCRIPTION OF DRAWINGS

[0009] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings in which:

[0010] Figure 1 Flow chart of the fire-fighting facility fault diagnosis method based on digital twinning according to the embodiments of the present application; Figure 2 Data flow diagram of the fire-fighting facility fault diagnosis method based on digital twinning according to the embodiments of the present application; Figure 3 Flow chart of the EWMA-based statistical trend deviation analysis of the original sensor value sequence with the working condition label to obtain the trend deviation score of the fire-fighting facility fault diagnosis method based on digital twinning according to the embodiments of the present application; Figure 4 Flow chart of the sequence auto-encoder-based time series pattern reconstruction of the original sensor value sequence with the working condition label to obtain the pattern anomaly score of the fire-fighting facility fault diagnosis method based on digital twinning according to the embodiments of the present application; Figure 5 Block diagram of the fire-fighting facility fault diagnosis system based on digital twinning according to the embodiments of the present application. DETAILED DESCRIPTION

[0011] Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. It should be apparent to those skilled in the art that the described embodiments are merely a portion of the embodiments of the present application and thus are not limited to the described embodiments. Further, the scope of the present application should be defined by the claims rather than the described embodiments.

[0012] As used in the present application and claims, the terms "a," "an," and "the" do not denote a singular entity, but include one or more entities unless the context clearly indicates otherwise. In general, the term "comprising" is used in the sense of "including", but not necessarily limited to, the listed steps and elements.

[0013] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or server. The modules are merely illustrative, and different aspects of the system and method can use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by systems according to embodiments of the present application. It should be understood that the operations in the figures do not necessarily have to be performed in the precise order shown. Rather, various steps can be handled in reverse order or simultaneously, as desired. Other operations can also be added or removed from the processes, or one or more steps can be modified.

[0015] In the following, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. It is obvious that the described embodiments are only some of the embodiments of the present application, and the present application is not limited to the described embodiments but can be implemented in various ways. Thus, the scope of the present application should be determined not by the described embodiments but by the claims and equivalents thereof.

[0016] The existing fire-fighting facility fault diagnosis scheme mainly faces two technical problems: one is that its alarm logic cannot adapt to the dynamic changes of the equipment under standby, inspection, start and other different working conditions, resulting in frequent false alarms and missed reports; the second is that when multiple analysis methods are integrated, the information fusion mechanism is too simple to reliably determine the final result, and the ambiguous signal of one channel can veto the clear fault warning of another channel. In view of the above technical problems, in the technical scheme of the present application, a fire-fighting facility fault diagnosis method based on digital twinning is proposed. The method first aligns the obtained original sensor data stream with the device working condition data in context, providing a standardized data basis with working condition tags for subsequent accurate analysis. Subsequently, the system synchronously sends the data sequence into two parallel analysis channels: one, using EWMA-based statistical trend deviation analysis to capture the gradual data drift relative to the normal baseline of a specific working condition caused by slow leakage and other faults; the second, using sequence autoencoder-based time series pattern reconstruction to identify the structural distortion of data in short form caused by sudden abnormality of components. In the decision fusion stage, instead of simply adding or multiplying the abnormal scores of the two channels, the certainty of each score is first quantified by information entropy theory, and then combined with the preset prior weight for different device type fault characteristics to dynamically generate the final fusion weight for the trend and pattern channels. This mechanism ensures that in the final decision, the system can intelligently focus on the more reliable and more indicative analysis results in the current scene, and finally calculates a robust final confidence score through weighted summation, and drives the digital twinning system to generate accurate adaptive alarms, thereby achieving high reliability diagnosis of early and hidden faults of fire-fighting facilities.

[0017] In the technical scheme of the present application, a fire-fighting facility fault diagnosis method based on digital twinning is proposed. Figure 1 Flowchart of the fire-fighting facility fault diagnosis method based on digital twinning according to an embodiment of the present application. Figure 2 Data flow schematic diagram of the fire-fighting facility fault diagnosis method based on digital twinning according to an embodiment of the present application. As shown inFigure 1 and Figure 2 As shown in the figure, the fire-fighting facility fault diagnosis method based on digital twinning according to the embodiment of the application comprises the steps of: S100, context alignment of acquired original sensor data stream and equipment working condition data to obtain original sensor numerical sequence with working condition label; S200, EWMA-based statistical trend deviation analysis on the original sensor numerical sequence with working condition label to obtain trend deviation score; S300, sequence auto-encoder-based time series pattern reconstruction on the original sensor numerical sequence with working condition label to obtain pattern anomaly score; S400, two-channel score fusion and final confidence score decision on the trend deviation score and the pattern anomaly score to obtain final confidence score; S500, adaptive alarm generation based on the final confidence score to obtain adaptive alarm; S600, display of the adaptive alarm on the display screen through the data twinning module.

[0018] Specifically, in step S100, the acquired original sensor data stream and equipment working condition data are context-aligned to obtain original sensor numerical sequence with working condition label. It should be understood that a key equipment such as a fire pump has multiple different working conditions, such as long standby, periodic inspection or emergency start, and the normal baseline value and fluctuation range of the outlet pressure under different working conditions are quite different. If the working conditions are not distinguished and the original pressure data is directly analyzed, the data change during normal start and stop of the equipment will be misjudged as a fault, or a small pressure drop caused by slow leakage in standby state cannot be detected, thereby causing a large number of false positives and false negatives of the diagnosis model. Therefore, in the technical solution of the present application, the acquired original sensor data stream and equipment working condition data are context-aligned to accurately label the equipment running background of each original sensor data point, and unordered and isolated original data is converted into information sequence with clear physical meaning and analysis value. It is worth mentioning that here, the original sensor data stream is a pressure sensor data stream collected by the outlet pressure sensor of the fire pump. By generating original sensor numerical sequence with working condition label, high-quality and unambiguous input is provided for the subsequent fault diagnosis model, ensuring that the diagnosis analysis is performed under comparable conditions, and the accuracy and reliability of the entire diagnosis scheme are fundamentally improved.

[0019] More specifically, in one specific example of the present application, the alignment process includes a series of operations such as data acquisition, timestamp synchronization, state interval division and data point mapping. First, the system acquires continuous pressure sensor data stream from the outlet pressure sensor of the fire pump in parallel, and acquires the device condition data log recording the start-stop, patrol and other actions of the device from the control system of the pump group (such as programmable logic controller PLC or supervisory control and data acquisition system SCADA). Then, the system normalizes and time-axis aligns the two heterogeneous data streams with timestamp as the key index, to ensure the consistency of the time reference. Subsequently, for any data point in the pressure sensor data stream, the system will retrieve and determine the running state interval to which the time belongs in the device condition data log according to its timestamp, such as standby pressure stabilization, periodic patrol or emergency start. Finally, the determined condition label and the corresponding original sensor value are bound to generate a structured original sensor value sequence with clear condition context, providing high-quality input for subsequent statistical trend deviation analysis and time series pattern reconstruction.

[0020] Specifically, in step S200, the original sensor value sequence with condition label is subjected to EWMA-based statistical trend deviation analysis to obtain a trend deviation score. It should be understood that due to some early failures of the fire facility, such as a small leak in the fire pump pipe network or a valve leak, they do not manifest as a sharp transient anomaly on the sensor data, but rather as a slow but continuous statistical drift of the key parameters (such as outlet pressure) compared to their normal baseline in a particular condition. This gradual feature is easily overlooked by traditional threshold alarm methods. Therefore, in the technical solution of the present application, the original sensor value sequence with condition label is further subjected to EWMA-based statistical trend deviation analysis to accurately quantify the deviation degree and persistence of the statistical center of the sensor value sequence compared to its historical normal level. In this way, it can effectively capture and evaluate those hidden failure precursors characterized by gradual deterioration, and convert the imperceptible performance degradation into a clear and quantifiable trend deviation score, providing a key basis for early warning.

[0021] Figure 3 A flowchart of the EWMA-based statistical trend deviation analysis of the original sensor value sequence with condition label according to the digital twin-based fire facility fault diagnosis method of the embodiments of the present application. As shown in FIG. 2, the process includes the following steps: Figure 3As shown, step S200 includes: S210, extracting historical mean and historical standard deviation from the preset model repository with the working condition label as the query keyword; S220, calculating upper control limit and lower control limit based on the historical mean and the historical standard deviation; S230, performing time series smoothing on the original sensor value sequence based on the smoothing coefficient to obtain a time series smoothed sensor value sequence; S240, performing instantaneous out-of-limit deviation sequence quantification on the time series smoothed sensor value sequence based on the upper control limit and the lower control limit to obtain a sensor value deviation sequence; and S250, performing deviation aggregation and normalization on the sensor value deviation sequence to obtain the trend deviation score.

[0022] More specifically, in one specific example of the present application, the statistical trend deviation analysis process first extracts the historical mean and the historical standard deviation of the fire pump outlet pressure under the specific working condition from the historical data model pre-constructed and stored in the model repository with the accompanying working condition label as the query index. In particular, the repository solidifies the pressure data statistical characteristics formed by the long-term stable operation of the fire pump under various known working conditions (such as standby, patrol), and matches each working condition label with its corresponding historical mean and historical standard deviation. Subsequently, after obtaining the two reference parameters, the system calculates the upper control limit and the lower control limit for defining the normal pressure fluctuation range under the working condition based on the extracted historical mean and historical standard deviation, which clearly defines the mathematical boundary of the normal fluctuation range. At the same time, the system uses the exponential weighted moving average (EWMA) algorithm to perform iterative smoothing processing on the input original sensor value sequence, where the smoothed value at the current time is the weighted sum of the current original value and the smoothed value at the previous time, and the weight is determined by the smoothing coefficient, to filter out random noise and highlight the inherent trend changes, thereby obtaining a time series smoothed sensor value sequence. Then, each data point in the smoothed sequence is compared with the previously calculated upper and lower control limits to quantify the instantaneous out-of-limit deviation of all data points that exceed the normal range, forming a sensor value deviation sequence. Finally, the system accumulates and aggregates all non-zero values in the deviation sequence within a predetermined time window to reflect the cumulative effect of the trend deviation, and maps the aggregation result to the interval of 0 to 1 through a normalization function, finally outputting a standardized trend deviation score that can intuitively reflect the severity of the statistical trend anomaly.

[0023] In particular, in the embodiment of the present application, step S220 includes calculating the upper control limit and the lower control limit based on the historical mean and the historical standard deviation using the following formula: ； ； wherein, is a smoothing coefficient, whose value ranges between is a control limit width coefficient, is a control limit width coefficient, and is a historical mean and a historical standard deviation, is an upper control limit, is a lower control limit.

[0024] Specifically, in step S300, the original sensor value sequence with the working condition label is subjected to sequence autoencoder-based time series pattern reconstruction to obtain a pattern anomaly score. It should be understood that due to another type of failure of the fire-fighting facility, such as instantaneous jamming of the fire pump impeller, rapid abnormal opening and closing of the valve or electrical system interference, it is manifested as short-term and severe waveform distortion or aperiodic oscillation on the sensor data. Such pattern-level anomalies cannot be effectively captured by trend analysis methods that focus on long-term statistical drift. Therefore, in the technical solution of the present application, the original sensor value sequence with the working condition label is further subjected to sequence autoencoder-based time series pattern reconstruction to deeply learn and master the fine time series pattern of the fire pump under a specific working condition, and to identify any structural deviation from the normal pattern. In this way, those sudden failures characterized by instantaneous and morphological abnormalities can be accurately identified and quantified as a pattern anomaly score, thereby complementing the trend deviation analysis and building a more comprehensive fault diagnosis view.

[0025] Figure 4 is a flowchart of the sequence autoencoder-based time series pattern reconstruction of the original sensor value sequence with the working condition label in the fire-fighting facility fault diagnosis method based on digital twinning according to the embodiment of the present application. As shown in Figure 4 step S300, it includes: S310, inputting the original sensor value sequence into the encoder of the sequence autoencoder to obtain a pressure time series pattern implicit encoding vector; S320, performing time series coding robust optimization on the pressure time series pattern implicit encoding vector to obtain a pressure time series pattern robust implicit encoding vector; S330, inputting the pressure time series pattern robust implicit encoding vector into the decoder of the sequence autoencoder to obtain a reconstructed original sensor value sequence; S340, determining a pattern anomaly score based on the reconstruction error between the original sensor value sequence and the reconstructed original sensor value sequence.

[0026] Specifically, in step S310, the original sensor value sequence is input into the encoder of the sequence autoencoder to obtain a pressure time sequence pattern implicit coding vector. It should be understood that, since the original sensor value sequence is a high-dimensional, noisy and complex time-dependent data stream, directly performing pattern comparison or anomaly identification on the original data has high computational complexity and is difficult to capture the inherent and abstract dynamic structure. Therefore, in the technical solution of the present application, the original sensor value sequence is further input into the encoder of the sequence autoencoder to perform nonlinear dimensionality reduction and core feature extraction on the original time sequence data, and to map it from a complex data space to a low-dimensional, information-intensive feature space. In this way, a pressure time sequence pattern implicit coding vector can be generated, which is a compact and essential representation of the original sequence pattern, and not only removes redundant information and noise interference, but also provides a high-quality and standardized feature basis for subsequent robust optimization and accurate reconstruction.

[0027] More specifically, in one specific example of the present application, the process inputs a predetermined length of original sensor value sequence with working condition label as input unit into an encoder composed of a recurrent neural network, which can specifically adopt a long short-term memory network (LSTM) or a gated recurrent unit (GRU) and the like which is good at processing sequence data. In the processing process, the encoder reads each data point in the sequence in time step order. After reading each point, the internal hidden state is updated according to the current input and the hidden state at the previous time, thereby gradually accumulating and encoding the time dependence and dynamic pattern information of the entire sequence. When all data points of the input sequence are processed, the final hidden state vector output by the encoder at the last time step is defined as the pressure time sequence pattern implicit coding vector of the time sequence data, which condenses the core dynamic features of the entire input sequence.

[0028] Specifically, in step S320, the pressure time sequence pattern implicit encoding vector is time sequence encoding robust optimized to obtain a pressure time sequence pattern robust implicit encoding vector. It should be understood that, due to the initial pressure time sequence pattern implicit encoding vector generated by the encoder, although the time sequence information is compressed, it gives equal weight to the information of all time points in the sequence in the characterization process, and fails to distinguish the key instantaneous pressure mutation (such as valve impact or pump instantaneous jamming induced peak) from the random noise in the smooth period, resulting in that the weight of the key fault feature may be improperly weakened due to the time elapse, affecting the robustness of pattern recognition. Therefore, in the technical solution of the present application, the pressure time sequence pattern implicit encoding vector is further time sequence encoding robust optimized to construct a non-uniform time sequence information decay model dynamically modulated by event saliency. The core of the model is to couple the importance of information with its time effectiveness, to identify the high saliency events in the pressure sequence and slow down the decay rate of the information weight, so as to highlight the persistent influence of key fault signs in information aggregation. In this way, a more intelligent and robust pressure time sequence pattern robust implicit encoding vector can be generated, which not only encodes the dynamic process of the sequence, but also more accurately reflects the degree to which the key pressure event is important so far, providing a key information enhanced and more anti-interference feature representation for subsequent decoding reconstruction and abnormal judgment.

[0029] More specifically, in the embodiment of the present application, the pressure time sequence pattern implicit encoding vector is time sequence encoding robust optimized to obtain a pressure time sequence pattern robust implicit encoding vector, including: performing local pattern segment division on the pressure time sequence pattern implicit encoding vector to obtain a sequence distribution of pressure local time sequence pattern implicit encoding vectors; extracting the maximum eigenvalue of each pressure local time sequence pattern implicit encoding vector in the sequence distribution of pressure local time sequence pattern implicit encoding vectors as a pressure time sequence saliency factor; calculating the time span of each pressure local time sequence pattern implicit encoding vector based on the time stamp of each pressure local time sequence pattern implicit encoding vector in the sequence distribution of pressure local time sequence pattern implicit encoding vectors; and performing time sequence decay propagation on the sequence distribution of pressure local time sequence pattern implicit encoding vectors based on the time span and the pressure time sequence saliency factor of each pressure local time sequence pattern implicit encoding vector in the sequence distribution of pressure local time sequence pattern implicit encoding vectors to obtain the pressure time sequence pattern robust implicit encoding vector.

[0030] That is, specifically, the pressure time sequence pattern implicit encoding vector is time sequence encoding robust optimized to obtain a pressure time sequence pattern robust implicit encoding vector, and the steps are as follows: Firstly, the pressure time-series pattern implicit encoding vector is locally segmented into a sequence distribution of pressure local time-series pattern implicit encoding vectors. It should be understood that, since the pressure time-series pattern implicit encoding vector generated by the encoder is a holistic and compressed representation of the entire time series, the pressure pattern information at different time points has been fused together in the vector. This single and macroscopic vector structure cannot provide discrete operation units for subsequent time-series decay analysis that needs to distinguish the importance of events at different times. Therefore, in the technical solution of the present application, the pressure time-series pattern implicit encoding vector is further locally segmented into a sequence distribution of multiple smaller-granularity implicit encoding vectors representing local time window features. In this way, independent analysis objects with their own time stamps can be provided for subsequent analysis steps, so that the system can extract the significance factor of each local mode segment and calculate its time span, thereby laying the necessary data structure foundation for realizing the dual-weighted information propagation mechanism based on event importance and timeliness.

[0031] Next, the maximum eigenvalue of each pressure local time-series pattern implicit encoding vector in the sequence distribution of pressure local time-series pattern implicit encoding vectors is extracted as a pressure time-series significance factor, which is represented by the following formula: ; wherein, is the time stamp of the i-th pressure local time-series pattern implicit encoding vector in the sequence distribution of pressure local time-series pattern implicit encoding vectors, is the i-th pressure local time-series pattern implicit encoding vector in the sequence distribution of pressure local time-series pattern implicit encoding vectors, is the maximum value, is the i-th pressure local time-series pattern implicit encoding vector, is the maximum value, is the i-th pressure local time-series pattern implicit encoding vector, is the pressure time-series significance factor corresponding to the i-th pressure local time-series pattern implicit encoding vector.

[0032] It should be understood that, since the information value contained in each fragment is not equal in the sequence distribution containing multiple local pattern fragments, some fragments may correspond to key events such as pressure spikes or sudden changes in state, while others only represent smooth background fluctuations. The system needs a quantitative indicator to distinguish this intrinsic importance. Therefore, in the technical solution of the present application, the maximum eigenvalue of each vector in the sequence distribution of the implicit coding vector of the pressure local temporal pattern is further extracted to assign a scalar called the pressure temporal significance factor to each pressure local temporal pattern fragment. This factor serves as an effective proxy for the signal activation intensity or pattern prominence in the time snapshot. In this way, the system can have the ability to identify key stress events, and provide a core basis for the subsequent temporal attenuation propagation stage to achieve differentiated information attenuation modulation based on content importance, ensuring that historical events with high significance can retain a higher weight in the final coding representation even over time.

[0033] Then, based on the timestamps of the respective pressure local time series pattern implicit coding vectors in the sequence distribution of the pressure local time series pattern implicit coding vectors, the time span of the respective pressure local time series pattern implicit coding vectors is calculated, which is expressed by the following formula: ； in, is the timestamp corresponding to the implicit encoding vector of the current pressure local temporal pattern, The local temporal pattern of pressure implicitly encodes the time span corresponding to the vector.

[0034] ； in, , is a learnable decay coefficient that controls the rate of time decay (which is optimized by backpropagation), is the neighbor set of the implicit encoding vector of the current pressure local temporal pattern, For the The pressure time series significance factor corresponding to the implicit coding vector of the pressure local time series pattern is used to enhance the contribution of important neighbors. is the logarithmic function value with the natural constant e as the base, is the Gaussian decay kernel (decays more slowly for recent features), is the timing attenuation factor.

[0035] It should be understood that the timeliness of historical events cannot be measured by the pressure time series significance factor alone, and a complete time series information decay model must consider both the importance of the event and the distance from its occurrence time. If it only relies on discrete sequence positions, it cannot accurately simulate the natural law that information decays with the passage of physical time. Therefore, in the technical solution of the present application, the time span of each local pattern fragment relative to the current time point is accurately calculated based on the timestamp of each vector in the sequence distribution of the implicit coding vector of the local pressure time series pattern. In this way, the abstract sequence position is converted into a continuous time interval with clear physical meaning, providing a core physical measurement basis for the subsequent time series decay propagation. In this way, a feature-time coupled modulation framework can be constructed, in which the time span of each calculated implicit coding vector of the local pressure time series pattern is directly used as the input of the Gaussian decay kernel, and together with the pressure time series significance factor, the weight of the historical information is determined, so that the significant features can retain a higher influence even over time. At the same time, by introducing a learnable attenuation coefficient, the system can adaptively adjust the sensitivity of the attenuation curve to the time span, thereby realizing a more robust pressure-time span dynamic information attenuation mechanism that is more suitable for specific equipment scenarios.

[0036] Finally, based on the time span of each pressure local temporal pattern implicit coding vector in the sequence distribution of the pressure local temporal pattern implicit coding vector and the pressure time series significance factor, the sequence distribution of the pressure local temporal pattern implicit coding vector is subjected to time attenuation propagation to obtain the pressure time series pattern robust implicit coding vector, which is expressed by the following formula: ； ; in, is the sequence distribution of the temporal decay encoding vector of the local temporal pattern of pressure, are the first, second, and third in the sequence distribution of the temporal attenuation coding vector of the local temporal pattern of pressure. and A pressure local temporal pattern temporal attenuation encoding vector, for encoder, is the robust implicit encoding vector of the pressure timing pattern.

[0037] It should be appreciated that, since the system has respectively acquired the stress temporal significance factor representing the importance of the event and the time span representing the timeliness of the event, but the two are independent and heterogeneous information dimensions, if not effectively fused, the system will not be able to distinguish between an old but critical stress event and a recent but unimportant fluctuation when aggregating information, so that the final stress temporal pattern encoding vector cannot accurately reflect the true risk state of the sequence. Therefore, in the technical solution of the present application, further based on the time span and the stress temporal significance factor, the sequence distribution of the stress local temporal pattern implicit encoding vector is propagated in time decay, so as to perform a feature-time coupled weighted aggregation process, which takes the time span as the basic input of the Gaussian decay kernel, and uses the stress temporal significance factor as the modulator of the decay amplitude. In this way, a final stress temporal pattern robust implicit encoding vector can be generated, which is not only the sum of sequence information, but also a compact representation reflecting which important stress events occur at what time and their lasting influence, so that significant failure features can still retain a higher weight over time, and the decay rate is dynamically adjusted by the learnable parameter to adapt to the time sensitivity needs of different working condition scenarios, thereby greatly improving the robustness and accuracy of pattern representation.

[0038] Specifically, in step S330, the stress temporal pattern robust implicit encoding vector is input into the decoder of the sequence autoencoder to obtain the reconstructed original sensor numerical sequence. It should be appreciated that, since the stress temporal pattern robust implicit encoding vector is a feature-enhanced, low-dimensional abstract representation, it cannot be directly compared with the original high-dimensional sensor numerical sequence to evaluate the degree of abnormality, it only represents the core pattern information of the sequence. Therefore, in the technical solution of the present application, the robust implicit encoding vector is further input into the decoder part of the sequence autoencoder, so as to perform a mapping operation inverse to the encoding process, that is, according to the highly condensed pattern information, try to restore the original time series data. In this way, a reconstructed original sensor numerical sequence can be generated, which represents the best fitting and reproduction of the input information based on the normal pattern knowledge learned by the model, providing a directly comparable object for subsequent quantification of pattern abnormalities by calculating reconstruction error.

[0039] More specifically, in one specific example of the present application, the decoding process first inputs the pressure time-series pattern robustly encoded vector as the initial hidden state of the decoder. The decoder is symmetric in structure to the encoder, also adopting a recurrent neural network structure such as long short-term memory network (LSTM) or gated recurrent unit (GRU). After receiving the initial hidden state, the decoder starts the iterative generation. In the first time step, it generates the first reconstructed data point of the sequence according to the initial state, and takes the data point as the input of the next time step while updating its internal hidden state. This process is repeated until a sequence of the same length as the original input sequence is generated. At each time step, the decoder predicts the value of the current time step according to its current hidden state and the output of the previous time step. Finally, all the values generated at each time step are combined in order to form the complete reconstructed original sensor value sequence.

[0040] Specifically, in step S340, the pattern anomaly score is determined based on the reconstruction error between the original sensor value sequence and the reconstructed original sensor value sequence. It should be understood that since the sequence autoencoder is trained based on a large amount of fire pump pressure data under normal working conditions only, its core capability lies in learning and accurately reproducing the internal structure of the normal time-series pattern. When the input original sensor value sequence contains abnormal patterns (such as pressure spikes or oscillations caused by sudden failures), the model cannot effectively reconstruct due to the lack of prior knowledge of such patterns, resulting in significant deviation between the reconstructed sequence and the original sequence. Therefore, in the technical solution of the present application, the reconstruction error between the original sensor value sequence and the reconstructed original sensor value sequence is further calculated to accurately quantify the degree of inconsistency between the time-series pattern of the current input sequence and the learned normal pattern. In this way, this invisible pattern difference can be converted into a clear and measurable pattern anomaly score, and the size of the score directly reflects the severity of the sudden failure, providing a decisive basis for system identification of transient anomalies.

[0041] More specifically, in the embodiments of the present application, the mode anomaly score is determined based on the reconstruction error between the original sensor value sequence and the reconstructed original sensor value sequence, including: calculating the root mean square error between the original sensor value sequence and the reconstructed original sensor value sequence as the reconstruction error; and normalizing the reconstruction error to obtain the mode anomaly score. That is, specifically, the process of determining the mode anomaly score first compares the original sensor value sequence with the reconstructed original sensor value sequence output by the decoder point by point. The system calculates the square of the difference between the original pressure value and the reconstructed pressure value at each time step, then accumulates all these square values in the entire sequence and takes the arithmetic mean, and finally takes the square root of the mean, thereby obtaining a root mean square error that can comprehensively reflect the overall difference between the two sequences. This error is defined as the reconstruction error. Subsequently, to eliminate the dimensional effect and make the score have uniform interpretability and comparability, the system normalizes the reconstruction error value, for example, by a pre-set normalization function to linearly or nonlinearly map it to a standardized interval (such as 0 to 1), and the normalized value obtained finally is the mode anomaly score.

[0042] Specifically, in step S400, the trend deviation score and the mode anomaly score are fused and finally the confidence score is determined. It should be understood that since the trend deviation score and the mode anomaly score are used to diagnose the running state of the fire-fighting equipment from two different dimensions, corresponding to slow-changing faults and sudden faults respectively, if a simple static fusion method such as addition or multiplication is used, the intelligent decision-making according to the fault type, equipment characteristics and real-time signal quality cannot be made, and it is easy for the fuzzy or uncertain signal of one channel to mistakenly veto the clear and accurate fault warning from the other channel, thereby resulting in insufficient robustness and reliability of the diagnosis system. Therefore, in the technical solution of the present application, the trend deviation score and the mode anomaly score are further fused and finally the confidence score is determined, so as to construct an adaptive dynamic weight fusion mechanism based on information uncertainty quantification and equipment prior knowledge. This mechanism can simulate the expert diagnosis logic, i.e., first evaluating the reliability of each diagnostic evidence, and then dynamically allocating decision weights in combination with the prior knowledge of equipment fault modes. In this way, the risk of misjudgment caused by the uncertainty of single channel information can be effectively avoided, so that the strong signal of one channel cannot be overwhelmed by the fuzzy signal of the other channel, and finally a robust final confidence score highly related to the real fault risk is generated.

[0043] More specifically, in the embodiments of the present application, the trend deviation score and the pattern anomaly score are fused in two channels and finally the confidence score is determined, including: the trend deviation score and the pattern anomaly score are quantified to obtain the trend certainty and the pattern certainty; based on the device prior weight, the trend certainty and the pattern certainty are dynamically fused based on the prior weight to generate the final trend weight and the final pattern weight; based on the final trend weight and the final pattern weight, the weighted sum of the trend deviation score and the pattern anomaly score is calculated as the final confidence score.

[0044] That is, the traditional two-channel alarm score fusion mechanism usually adopts a direct multiplication method to synthesize the final confidence score. This method is simple, but it has its inherent technical defects. Specifically, the mechanism lacks recognition of information quality in evaluation, and gives static and equal weights to signals from different channels. When the signal quality of one channel is low or its judgment is ambiguous, the simple multiplication operation will improperly suppress the clear and strong real fault signal that the other channel may emit. This risk of one vote veto and the indiscriminate treatment of device fault features make it difficult to perform high-reliability early warning tasks in complex industrial environments.

[0045] Based on this, the embodiments of the present application propose a set of adaptive dynamic weight fusion decision mechanism based on information entropy and device prior knowledge. Instead of simply multiplying the alarm score, the mechanism simulates the diagnosis logic of experts through a series of precise mathematical steps: first, evaluate the credibility of each piece of evidence, i.e. the trend deviation score and the pattern anomaly score, then dynamically allocate the right to speak in combination with the prior knowledge of the device fault mode, and finally make a comprehensive judgment.

[0046] Specifically, first, the trend deviation score and the pattern anomaly score are quantified. The process is represented by the formula: ； ； Wherein, is the trend deviation score, is the pattern anomaly score, is the trend certainty, is the pattern certainty.

[0047] Accordingly, considering that the original anomaly score itself does not fully reflect the certainty of diagnosis; an intermediate score has much more inherent uncertainty than an extreme score. Therefore, it is intended to use the tools of information theory to explicitly quantify this invisible uncertainty. Accordingly, a degree of certainty is given to the diagnosis result of each channel, which directly affects its component in the final decision.

[0048] Specifically, the above formula is applied to the original score from the trend channel (trend deviation score) and the pattern channel (pattern anomaly score), so as to obtain their respective degrees of certainty (trend certainty) and ( pattern certainty). After performing this quantization, the system obtains two new dimensions, realizing the upgrade from signal strength to judgment certainty. The closer a score is to the two poles (abnormal or normal), the higher its certainty; on the contrary, the closer it is to the ambiguous middle ground, the lower the certainty, thereby providing a solid foundation for weight allocation.

[0049] Then, based on the device prior weight, the trend certainty and the pattern certainty are dynamically fused based on the prior weight to generate the final trend weight and the final pattern weight. This process can be represented by the formula: ； ； wherein, is the device prior weight, is the final trend weight, is the final pattern weight.

[0050] Accordingly, considering that different types of fire-fighting equipment have different typical signs of failure, the sensitivity to trends and the sensitivity to patterns should be emphasized. For example, the slow leakage of a water pump shows strong trend, while the blade crack of a fan shows strong pattern. The purpose of this step is to solidify this expert-level device prior knowledge into the model, and combine it with the real-time certainty obtained in the previous step to generate truly dynamic and intelligent decision weights.

[0051] It should be understood that the final say of a channel is determined by its innate importance and the confidence of the judgment / . When diagnosing a water pump with trend anomaly as the main fault feature, even if the certainty of the trend channel is slightly lower than that of the pattern channel, its weight Still dominated by high values. More importantly, when any one channel falls into uncertainty values approach zero, its weight will automatically collapse, and the system will decisively rely on another more certain channel, thus autonomously circumventing information pollution.

[0052] After the above-mentioned dynamic allocation of weights is completed, a weighted sum of the trend deviation score and the mode anomaly score is calculated as the final confidence score based on the final trend weight and the final mode weight. It should be understood that the final confidence score is no longer a fragile and easily vetoed judgment, but a robust decision that has undergone uncertainty evaluation, priori knowledge calibration and dynamic weight adjustment. This way of weighted sum makes the strong signal of one channel not be overwhelmed by the ambiguous signal of another channel, but occupies its due decision proportion according to the real-time scene. In this way, key failure precursors that the original mechanism would miss can be accurately captured, and a robust confidence score highly related to the real failure risk is output.

[0053] Specifically, in steps S500 and S600, adaptive alarm generation is performed based on the final confidence score to obtain an adaptive alarm, and the adaptive alarm is displayed on the display screen through a data twin module. It should be understood that since the final confidence score is a comprehensive numerical value quantifying the failure risk of the fire-fighting facility, it cannot be directly converted into an operation instruction that can be understood and executed by the operation and maintenance personnel. If only presented in numerical form, it cannot directly reveal the emergency degree and spatial position of the failure, thereby reducing the efficiency and accuracy of failure response. Therefore, in the technical solution of the present application, adaptive alarm generation is further performed based on the final confidence score, and the adaptive alarm is displayed on the display screen through a data twin module, so as to translate the abstract risk score into an alarm event with clear levels and content, and accurately associate the alarm information with the physical entity of the facility through three-dimensional visualization means. In this way, an intuitive, immersive and information-rich failure diagnosis interface can be provided for the operation and maintenance personnel, so that they can immediately master the severity level, exact location and related data of the failure, thereby realizing rapid decision-making and precise intervention, and significantly improving the operation and maintenance level and safety protection capability of the fire-fighting system.

[0054] ​More specifically, in one specific example of the present application, the adaptive alarm generation and display process first compares the calculated final confidence score with a pre-set multi-level alarm threshold system. The system defines different score intervals corresponding to different alarm levels, such as pre-alarm, general alarm and serious alarm, etc. According to the interval that the final confidence score falls into, the system determines the alarm level of the current event. Then, the system generates a structured adaptive alarm information object based on the alarm level, which contains the alarm level, the final confidence score, the event timestamp and a fault description text automatically generated according to the score sources (i.e. the weight contribution of the trend deviation score and the pattern anomaly score). Finally, the alarm information object is passed to the data twin module for display, which first locates the corresponding fire pump digital twin model according to the device identifier contained in the alarm information in the three-dimensional visualization scene, and then applies different visual rendering effects to the model according to the alarm level, such as highlighting with yellow, orange or red color, while popping up an information panel next to or above the model to clearly show all the detailed contents in the adaptive alarm information object.

[0055] In summary, the fire-fighting facility fault diagnosis method based on digital twin according to the embodiments of the present application is illustrated, which captures gradual faults based on EWMA statistical analysis, and at the same time uses sequence autoencoder to deeply mine transient pattern anomalies. This abandons the traditional static score fusion method, and introduces a dual-channel score fusion and final confidence decision mechanism. The mechanism can dynamically allocate decision weights for the trend and pattern analysis channels according to the real-time determination degree of each channel diagnosis result, combined with the prior knowledge of the fault characteristics of specific devices, effectively avoiding the interference of single channel information ambiguity on the overall judgment, thereby realizing accurate and robust diagnosis of early and hidden faults of fire-fighting facilities, and improving the reliability of the alarm.

[0056] Further, a fire-fighting facility fault diagnosis system based on digital twin is also provided.

[0057] Figure 5 A block diagram of the fire-fighting facility fault diagnosis system based on digital twin according to the embodiments of the present application is shown in FIG. 1. As shown in the figure, the system comprises a data collection module 101, a data preprocessing module 102, a trend analysis module 103, a pattern analysis module 104, a final confidence score calculation module 105, an adaptive alarm generation and display module 106, and a digital twin display module 107. Figure 5As shown, the fire protection facility fault diagnosis system 100 based on digital twin according to the embodiment of the present application includes: a raw sensor data alignment module 110, which is used to perform context alignment on the acquired raw sensor data stream and equipment operating condition data to obtain a raw sensor value sequence with an attached operating condition label; a statistical trend deviation analysis module 120, which is used to perform a statistical trend deviation analysis based on EWMA on the raw sensor value sequence with the attached operating condition label to obtain a trend deviation score; a time series pattern reconstruction module 130, which is used to perform a time series pattern reconstruction based on a sequence autoencoder on the raw sensor value sequence with the attached operating condition label to obtain a pattern anomaly score; a confidence score calculation module 140, which is used to perform dual-channel score fusion and final confidence judgment on the trend deviation score and the pattern anomaly score to obtain a final confidence score; an adaptive alarm module 150, which is used to perform adaptive alarm generation based on the final confidence score to obtain an adaptive alarm; and a digital twin display module 160, which is used to display the adaptive alarm on a display screen through the data twin module.

[0058] As described above, the firefighting facility fault diagnosis system 100 based on digital twins according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with a firefighting facility fault diagnosis algorithm based on digital twins. In one possible implementation, the firefighting facility fault diagnosis system 100 based on digital twins according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the firefighting facility fault diagnosis system 100 based on digital twins can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the firefighting facility fault diagnosis system 100 based on digital twins can also be one of the many hardware modules of the wireless terminal.

[0059] While various embodiments of the present disclosure have been described above, the above descriptions are illustrative, non-exhaustive, and not intended to be limiting of the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or improvements to existing technologies, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A fire protection facility fault diagnosis method based on digital twin, characterized in that: include: Contextually align the acquired raw sensor data stream and equipment operating condition data to obtain a raw sensor value sequence with an attached operating condition label; Performing an EWMA-based statistical trend deviation analysis on the original sensor value sequence with the working condition label to obtain a trend deviation score; Reconstructing the time series pattern of the original sensor value sequence with the working condition label based on the sequence autoencoder to obtain a pattern anomaly score; Perform dual-channel score fusion and final confidence judgment on the trend deviation score and pattern anomaly score to obtain the final confidence score; Performing adaptive alarm generation based on the final confidence score to obtain an adaptive alarm; The adaptive alarm is displayed on the display screen through the data twin module.

2. The fire protection facility fault diagnosis method based on digital twin according to claim 1 is characterized in that: The original sensor data stream is a pressure sensor data stream collected by the outlet pressure sensor of the fire main pump.

3. The fire protection facility fault diagnosis method based on digital twin according to claim 1 is characterized in that: Performing an EWMA-based statistical trend deviation analysis on the original sensor value sequence with the working condition label to obtain a trend deviation score includes: Using the working condition label as the query keyword, the historical mean and historical standard deviation are extracted from the preset model repository; Calculate upper and lower control limits based on historical mean and historical standard deviation; Based on the smoothing coefficient, time series smoothing is performed on the original sensor value sequence to obtain a time series smoothed sensor value sequence; Based on the upper control limit and the lower control limit, quantizing the instantaneous limit-crossing deviation sequence of the time-series smoothed sensor value sequence to obtain a sensor value deviation sequence; Deviation aggregation and normalization are performed on the sensor value deviation sequence to obtain the trend deviation score.

4. The fire protection facility fault diagnosis method based on digital twin according to claim 3 is characterized in that: Calculating the upper control limit and the lower control limit based on the historical mean and the historical standard deviation includes: calculating the upper control limit and the lower control limit based on the historical mean and the historical standard deviation using the following formula, wherein the formula is: ； ； in, is the smoothing coefficient, is the control limit width coefficient, and are the historical mean and historical standard deviation, is the upper control limit, is the lower control limit.

5. The fire protection facility fault diagnosis method based on digital twin according to claim 1 is characterized in that: The original sensor value sequence with the working condition label is reconstructed based on the sequence autoencoder to obtain a pattern anomaly score, including: Inputting the original sensor value sequence into an encoder of a sequence autoencoder to obtain an implicit encoding vector of a pressure time series pattern; Performing temporal coding robust optimization on the pressure temporal pattern implicit coding vector to obtain a pressure temporal pattern robust implicit coding vector; Inputting the robust implicit coding vector of the pressure time series pattern into the decoder of the sequence autoencoder to obtain a reconstructed original sensor value sequence; A pattern anomaly score is determined based on a reconstruction error between the original sensor value sequence and the reconstructed original sensor value sequence.

6. The fire protection facility fault diagnosis method based on digital twin according to claim 5 is characterized in that: Performing temporal coding robust optimization on the pressure temporal pattern implicit coding vector to obtain a pressure temporal pattern robust implicit coding vector includes: Dividing the pressure time series pattern implicit coding vector into local pattern segments to obtain a sequence distribution of the pressure local time series pattern implicit coding vector; Extracting the maximum eigenvalue of each pressure local time series pattern implicit coding vector in the sequence distribution of the pressure local time series pattern implicit coding vector as the pressure time series significance factor; Calculating the time span of each pressure local time series pattern implicit code vector based on the timestamp of each pressure local time series pattern implicit code vector in the sequence distribution of the pressure local time series pattern implicit code vector; Based on the time span of each pressure local time series pattern implicit coding vector and the pressure time series significance factor in the sequence distribution of the pressure local time series pattern implicit coding vector, the sequence distribution of the pressure local time series pattern implicit coding vector is subjected to time attenuation propagation to obtain the pressure time series pattern robust implicit coding vector.

7. The fire protection facility fault diagnosis method based on digital twin according to claim 6 is characterized in that: Based on the reconstruction error between the original sensor value sequence and the reconstructed original sensor value sequence, a pattern anomaly score is determined, including: Calculate the root mean square error between the original sensor value sequence and the reconstructed original sensor value sequence as the reconstruction error; The reconstruction error is normalized to obtain the pattern anomaly score.

8. The fire protection facility fault diagnosis method based on digital twin according to claim 1 is characterized in that: Perform dual-channel score fusion and final confidence judgment on the trend deviation score and pattern anomaly score to obtain the final confidence score, including: The channel score uncertainty is quantified for the trend deviation score and pattern anomaly score to obtain the trend certainty and pattern certainty; Based on the device prior weight, the trend certainty and the pattern certainty are dynamically fused based on the prior weight to obtain the final trend weight and the final pattern weight; Based on the final trend weight and the final pattern weight, a weighted sum of the trend deviation score and the pattern anomaly score is calculated as the final confidence score.

9. A fire protection facility fault diagnosis system based on digital twins, characterized in that: include: A raw sensor data alignment module is used to contextually align the acquired raw sensor data stream and equipment operating condition data to obtain a raw sensor value sequence with an attached operating condition label; A statistical trend deviation analysis module, configured to perform an EWMA-based statistical trend deviation analysis on the original sensor value sequence with the working condition label to obtain a trend deviation score; A time series pattern reconstruction module, configured to reconstruct the time series pattern of the original sensor value sequence with the working condition label based on a sequence autoencoder to obtain a pattern anomaly score; The confidence score calculation module is used to perform dual-channel score fusion and final confidence judgment on the trend deviation score and pattern anomaly score to obtain the final confidence score; An adaptive alarm module, configured to generate an adaptive alarm based on the final confidence score to obtain an adaptive alarm; The digital twin display module is used to display the adaptive alarm on the display screen through the data twin module.

Images