Fire-fighting equipment fault prediction method based on big data analysis

Abstract

The invention discloses a fire-fighting equipment fault prediction method based on big data analysis, and the method comprises the following steps: collecting fire-fighting equipment parameters, carrying out the preprocessing, and generating a standardized input tensor; inputting the input tensor into a Mama encoder containing a hot pressure flow physical sensing layer, generating a physical modulation factor according to temperature gradient, pressure response and flow inertia, performing weighted correction on a state transition matrix, and outputting a time sequence implicit feature matrix; inputting the matrix into a Chronos model to generate a multi-step prediction trajectory, calibrating prediction distribution by using a physical modulation factor, adjusting a time causal attention weight, and outputting a prediction result subjected to physical calibration; and a health score and a fault probability are generated through a decoding module, and the prediction advance time and the early warning level are determined according to the sub-position cell interval. According to the invention, based on fusion of the Mama encoder and the Chronos model, high-precision, explainable and advanced prediction of fire-fighting equipment faults is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence data analysis technology, and in particular to a method for predicting fire equipment failures based on big data analysis. Background Technology

[0002] With the continuous development of building intelligence and urban safety systems, fire protection equipment is widely deployed in public buildings, industrial parks, and infrastructure. However, equipment such as fire pumps, sprinkler valve assemblies, pressure pipelines, and smoke detectors are susceptible to various factors during long-term operation, including ambient temperature, humidity, pipeline pressure fluctuations, and power supply status, leading to gradual performance degradation or the development of hidden faults.

[0003] Currently, the maintenance of fire-fighting equipment mainly relies on regular manual inspections or threshold-based alarm mechanisms. The system only triggers an alarm when signals such as temperature, pressure, or current exceed set ranges. This approach has the following shortcomings:

[0004] (1) Insufficient real-time performance: The inspection cycle is too long, making it difficult to capture early signs of faults in a timely manner;

[0005] (2) Low data utilization: The collected operational data failed to be uniformly modeled and multi-source fused, making it impossible to identify potential anomalies from long-term trends;

[0006] (3) High false alarm and false alarm rates: Fixed thresholds are difficult to adapt to complex environments and parameter drift caused by equipment aging, resulting in inaccurate early warnings;

[0007] (4) Lack of physical constraints: Most existing prediction models based on machine learning or deep learning are purely data-driven and fail to combine physical characteristics such as heat conduction, pressure response and flow inertia in fire protection systems, resulting in poor interpretability of prediction results.

[0008] In recent years, with the development of big data and deep time series modeling technology, researchers have attempted to use models such as recurrent neural networks and Transformers for equipment state prediction. However, these models generally suffer from high computational complexity, insufficient long-term dependency capture, and a lack of characterization of physical laws, making them difficult to apply stably in scenarios such as fire-fighting equipment that are multi-source, strongly coupled, and have high real-time requirements.

[0009] Therefore, how to provide a method for predicting fire equipment failures based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0010] One objective of this invention is to propose a method for predicting fire equipment faults based on big data analysis. This invention establishes a standardized input tensor by collecting temperature, pressure, flow rate, current, and environmental parameters of fire equipment to achieve unified modeling of multi-source data. A thermo-pressure-flow physical sensing layer is introduced into the Mamba time-series encoder. Physical modulation factors are generated based on temperature gradient, pressure response, and flow inertia to weight and correct the state transition matrix, enabling the model to reflect the physical characteristics of the fire system during feature extraction. A multi-step probability prediction of the implicit temporal features is performed using the Chronos time-series model, and the physical modulation factors are used to perform consistency calibration on the prediction distribution, dynamically adjusting the temporal causal attention weights to align physical laws with data predictions. During the decoding stage, a health score and fault probability distribution are output, and prediction lead time and warning level are generated based on the location confidence intervals. Simultaneously, prediction residuals are calculated based on equipment operation feedback, and physical modulation parameters and calibration weights are corrected online, achieving model self-learning and accuracy maintenance, thereby improving the accuracy, lead time, and interpretability of fire equipment fault prediction.

[0011] A method for predicting fire equipment failures based on big data analysis according to an embodiment of the present invention includes the following steps:

[0012] The temperature, pressure, flow rate, current and environmental parameters of the fire-fighting equipment are collected, and the data is synchronized in time, corrected for outliers and normalized to obtain a standardized input tensor.

[0013] The input tensor is input into the Mamba encoder, which includes a thermal-pressure-flow physical sensing layer. The physical modulation factor is generated based on the temperature gradient, pressure response and flow inertia, and a weighted correction is performed on the state transition matrix accordingly to output a temporal implicit feature matrix.

[0014] The time-series latent feature matrix is ​​input into the Chronos time series basic model, and the Chronos time series basic model generates a multi-step probability prediction trajectory within the target time window.

[0015] The physical modulation factor output by the thermal-pressure-flow physical sensing layer is used to perform consistency calibration on the Chronos prediction distribution. When the prediction result is inconsistent with the physical sensing features, the temporal causal attention weight is adjusted to obtain the physically calibrated prediction distribution.

[0016] The physically calibrated predicted distribution is input into the decoding module, which outputs a health score and a fault probability distribution.

[0017] And based on the location information intervals, the forecast lead time and warning level are generated;

[0018] The system collects operational feedback data, calculates prediction residuals, and corrects the physical modulation parameters in the thermo-pressure-flow physical sensing layer and the calibration weights of the Chronos model based on the residuals.

[0019] Furthermore, the step of generating the normalized input tensor includes:

[0020] Raw operational data is collected by various types of sensors installed in fire pumps, sprinkler valve assemblies, pipeline nodes, and environmental monitoring points, and timestamp information is added to each set of sampled data.

[0021] Using a unified time reference signal as a reference, interpolation resampling is performed on sensor signals with different sampling periods;

[0022] Anomaly detection is performed on the acquired raw signals. Based on sliding window statistics and neighborhood density, abrupt changes or drift points are identified and corrected using the mean of adjacent windows.

[0023] The corrected signal is normalized by using a combination of minimum-maximum mapping and zero-mean normalization to standardize the data of each channel, forming an input matrix with uniform dimensions.

[0024] The input matrix is ​​concatenated and organized into a standardized input tensor along the time series dimension.

[0025] Furthermore, the generation step of the temporal latent feature matrix includes:

[0026] The standardized input tensor is mapped to an encoded input channel according to the device type and physical quantity dimension, and a sliding window and batch length are set in the time step dimension;

[0027] In the heat-pressure-flow physical sensing layer, temperature gradient, pressure response and flow inertia are calculated based on the numerical changes of adjacent time steps and statistics within the window. The temperature gradient is based on the time difference or local slope of the temperature sequence, the pressure response is based on the pulsation amplitude and rise / fall duration of the pressure sequence, and the flow inertia is based on the acceleration characteristics and hysteresis of the flow sequence.

[0028] The temperature gradient, pressure response, and flow inertia are combined according to the preset channel mapping relationship and the device type index to obtain the corresponding physical feature vector, and then the physical modulation factor is generated through linear transformation and normalization.

[0029] The state transition matrix of the Mamba encoder is weighted and corrected according to the physical modulation factor, including updating the matrix row and column weights channel by channel, and iteratively applying the corrected state transition coefficients at each time step to update the hidden state.

[0030] After completing the weighted correction and hidden state update, the hidden state outputs of each time step are collected in chronological order to form a temporal hidden feature matrix.

[0031] Furthermore, the generation of the physical modulation factor includes:

[0032] Within a set time sliding window, the mean and variance of the time difference of the temperature series are calculated respectively, and used as temperature gradient parameters;

[0033] Calculate the pulsation amplitude, rising edge, and falling edge duration of the pressure sequence as pressure response parameters;

[0034] Calculate the acceleration, hysteresis, and velocity fluctuation rate of the flow sequence as flow inertia parameters;

[0035] Temperature gradient parameters, pressure response parameters, and flow inertia parameters are input into a coupling function for nonlinear fusion. The coupling function is a bilinear mapping function.

[0036] The fusion results are normalized and subjected to exponential moving average processing to generate the physical modulation factor corresponding to the time step.

[0037] Furthermore, the generation step of the multi-step probabilistic prediction trajectory includes:

[0038] The temporal latent feature matrix is ​​expanded by time step, a context sequence of length L and a prediction window of length H are set, and a corresponding time index is constructed.

[0039] Add time and period codes to each time step, including intraday index, intraweek index and rhythmic period identifier, and concatenate them with device identifier embedding and channel identifier embedding to form Chronos input tensor;

[0040] Causal masking is applied to the input tensor, autoregressive decoding is used, and sliding window step size and overlap length are set. During the training phase, teacher-mandated and random masking strategies are used to generate conditional input sequences.

[0041] During the decoding phase, the prediction window is expanded step by step, and the conditional probability parameter set and the quantile prediction corresponding to the quantile set are output for each time step.

[0042] By using shared model parameters for different device channels and introducing channel bias embedding, the multi-channel outputs are stacked according to device and time step to obtain multi-step probability prediction trajectories within the target time window.

[0043] Furthermore, the step of generating the physically calibrated predicted distribution includes:

[0044] Obtain the physical modulation factor sequence output by the thermal-pressure-flow physical sensing layer at each time step, and normalize the modulation factors according to the time step and channel dimension.

[0045] Calculate the difference measure between the mean trajectory of the predicted distribution output by the Chronos model at each time step and the trend of the modulation factor change, the difference measure including time-series correlation difference and amplitude offset difference;

[0046] A physical calibration weight matrix is ​​generated based on the difference measure, and then time-weighted smoothed according to the importance of the channel.

[0047] In the temporal causal attention layer of the Chronos model, the physical calibration weight matrix is ​​embedded as a modulation coefficient before the attention score is calculated, and the attention score is adjusted element-wise.

[0048] The adjusted attention scores are re-normalized and forward propagation is performed. The attention weights are updated and the output distribution is recalculated to obtain the physically calibrated prediction distribution.

[0049] Furthermore, the steps for generating the prediction lead time and warning level include:

[0050] The physically calibrated prediction distribution is grouped by time step and device channel, and then input into the probability mapping layer of the decoding module;

[0051] In the probability mapping layer, parameter decoupling and reshaping are performed on the prediction distribution at each time step to calculate the corresponding health state vector;

[0052] The health status vector is input into the linear transformation layer and the normalization layer to generate a health score for each device channel;

[0053] The continuous rate of change over time is calculated based on the health score sequence, and the corresponding failure probability distribution is generated by combining the variance and confidence boundary of the predicted distribution.

[0054] Quantile analysis is performed on the fault probability distribution to calculate the upper and lower limits of the confidence interval, and the prediction lead time is determined based on the intersection of different quantile boundaries.

[0055] The warning levels are divided according to the prediction lead time and the failure probability level, and the corresponding warning results and confidence interval parameters are output.

[0056] Furthermore, the step of generating the predicted residual includes:

[0057] Obtain the runtime feedback data corresponding to the prediction time window, and perform time alignment and format conversion on the feedback data;

[0058] The feedback data and prediction results are compared at the same time step to calculate the health score residual and the failure probability residual, and a residual sequence is formed.

[0059] Perform moving average and outlier filtering on the residual sequence to obtain a smoothed residual signal;

[0060] The physical modulation parameters in the thermal-pressure-flow physical sensing layer are updated based on the smoothed residual signal, and the temperature gradient, pressure response and flow inertia weights are corrected.

[0061] Simultaneously, the physical calibration weights of the Chronos model are updated based on the mean square deviation of the residuals of each channel, and the attention weight matrix is ​​incrementally adjusted.

[0062] After the update is complete, the forward propagation is re-executed to generate a new prediction distribution.

[0063] The beneficial effects of this invention are:

[0064] This invention collects temperature, pressure, flow rate, current, and environmental parameters of fire-fighting equipment, and performs time synchronization, anomaly correction, and normalization processing to construct a standardized input tensor. This enables a unified representation of data from different devices and at different sampling frequencies, thereby improving data integrity and modeling accuracy.

[0065] A thermal-pressure-flow physical sensing layer is set up inside the Mamba time encoder. A physical modulation factor is generated based on the temperature gradient, pressure response and flow inertia, and a weighted correction is performed on the state transition matrix accordingly. This enables the model to dynamically reflect the thermodynamic and fluid characteristics of the fire protection system during the time modeling process, and realize feature extraction under physical constraints.

[0066] This invention utilizes the Chronos model to perform multi-step probabilistic prediction of time-series latent features encoded by Mamba. It introduces time encoding and periodic encoding at the input end to adapt to the periodic patterns of fire-fighting equipment operation, thereby simultaneously capturing short-term fluctuations and long-term trends, and improving the timeliness and stability of prediction.

[0067] The Chronos model predicts distribution consistency by using the physical modulation factor output by the thermal-pressure-fluid physical sensing layer. When the prediction results do not match the physical sensing features, the temporal causal attention weight is dynamically adjusted to achieve adaptive alignment between the predicted distribution and physical laws, thereby reducing the accumulation of errors caused by environmental disturbances.

[0068] In the decoding stage, this invention performs quantile analysis on the physically calibrated predicted distribution, calculates confidence intervals, and generates health scores and fault probability distributions, thereby achieving quantitative expression of fault risk and accurate classification of early warning levels.

[0069] By collecting feedback data from equipment operation, calculating prediction residuals, and using these residuals to correct the physical modulation parameters in the thermo-pressure-flow physical sensing layer and the calibration weights of the Chronos model, an online adaptive update mechanism is constructed, enabling the model to maintain prediction accuracy and environmental adaptability during long-term operation. Attached Figure Description

[0070] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0071] Figure 1 This is a flowchart of a fire equipment fault prediction method based on big data analysis proposed in this invention;

[0072] Figure 2 This is a schematic diagram of the thermal pressure flow physical sensing layer structure of a fire equipment fault prediction method based on big data analysis proposed in this invention.

[0073] Figure 3 This is a Chronos multi-step prediction flowchart for a fire equipment fault prediction method based on big data analysis proposed in this invention. Detailed Implementation

[0074] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0075] refer to Figure 1 - Figure 3 A method for predicting fire equipment failures based on big data analysis includes the following steps:

[0076] The temperature, pressure, flow rate, current and environmental parameters of the fire-fighting equipment are collected, and the data is synchronized in time, corrected for outliers and normalized to obtain a standardized input tensor.

[0077] The input tensor is input into the Mamba encoder, which includes a thermal-pressure-flow physical sensing layer. The physical modulation factor is generated based on the temperature gradient, pressure response and flow inertia, and a weighted correction is performed on the state transition matrix accordingly to output a temporal implicit feature matrix.

[0078] The time-series latent feature matrix is ​​input into the Chronos time series basic model, and the Chronos time series basic model generates a multi-step probability prediction trajectory within the target time window.

[0079] The physical modulation factor output by the thermal-pressure-flow physical sensing layer is used to perform consistency calibration on the Chronos prediction distribution. When the prediction result is inconsistent with the physical sensing features, the temporal causal attention weight is adjusted to obtain the physically calibrated prediction distribution.

[0080] The physically calibrated predicted distribution is input into the decoding module, which outputs a health score and a fault probability distribution.

[0081] And based on the location information intervals, the forecast lead time and warning level are generated;

[0082] The system collects operational feedback data, calculates prediction residuals, and corrects the physical modulation parameters in the thermo-pressure-flow physical sensing layer and the calibration weights of the Chronos model based on the residuals.

[0083] In this embodiment, the step of generating the normalized input tensor includes:

[0084] Raw operational data is collected by various types of sensors installed in fire pumps, sprinkler valve assemblies, pipeline nodes, and environmental monitoring points, and timestamp information is added to each set of sampled data.

[0085] Using a unified time reference signal as a reference, interpolation resampling is performed on sensor signals with different sampling periods. Linear interpolation algorithm is used for linearly changing signals, and spline interpolation algorithm is used for signals with obvious nonlinear fluctuations.

[0086] Anomaly detection is performed on the acquired raw signals. Based on sliding window statistics and neighborhood density, abrupt changes or drift points are identified and corrected using the mean of adjacent windows. The window length for anomaly detection is adjusted according to the sampling frequency.

[0087] The corrected signal is normalized by using a combination of minimum-maximum mapping and zero-mean normalization to standardize the data of each channel, forming an input matrix with uniform dimensions. The minimum-maximum mapping is used to eliminate the dimensional differences between different physical quantities, and the zero-mean normalization is used to maintain the stationarity of the time series.

[0088] The input matrix is ​​concatenated along the time series dimension and organized into a standardized input tensor. During the concatenation process, the channel order is kept consistent, and the time indices are numbered consecutively.

[0089] In this embodiment, the step of generating the temporal hidden feature matrix includes:

[0090] The standardized input tensor is mapped to encoded input channels according to device type and physical quantity dimension. During the mapping process, the arrangement order and weight initialization method of the input channels are determined according to the signal characteristics of different devices. A sliding window and batch length are set in the time step dimension. The window length is set according to the signal sampling period.

[0091] In the heat-pressure-flow physical sensing layer, the temperature gradient, pressure response, and flow inertia are calculated based on the numerical changes of adjacent time steps and the statistics within the window. The temperature gradient is obtained through time difference or local linear fitting. The pressure response is calculated based on the pulsation amplitude of the pressure sequence and the duration of the rising and falling edges. The flow inertia is obtained through the rate of change of velocity and the hysteresis term, and is processed by moving average in the time dimension to eliminate transient noise.

[0092] Temperature gradient, pressure response, and flow inertia are combined according to a preset channel mapping relationship and device type index to obtain the corresponding physical feature vector. Then, through linear transformation and standardization, different physical quantities are mapped to a unified feature space, thereby generating a physical modulation factor that reflects the comprehensive physical state of multiple channels.

[0093] The state transition matrix of the Mamba encoder is weighted and corrected according to the physical modulation factor. During the weighting process, the row and column weights of the state matrix are updated channel by channel. The corrected state transition coefficients are iteratively applied at each time step to update the hidden state. A residual preservation mechanism is introduced in the update to prevent the state gradient from decaying.

[0094] After completing the weighted correction and hidden state update, the hidden state outputs of each time step are collected in chronological order to form a temporal hidden feature matrix. This matrix corresponds to the dynamic characteristics of various physical quantities in the channel dimension and records the state evolution sequence in the time dimension, providing input basis for the Chronos time series basic model.

[0095] In this embodiment, the generation of the physical modulation factor includes:

[0096] Within a set time sliding window, the mean and variance of the time difference of the temperature sequence are calculated respectively as temperature gradient parameters. The mean of the time difference is used to characterize the trend of temperature change, and the variance is used to measure the stability of temperature fluctuation. Together, they reflect the dynamic characteristics of fire-fighting equipment in the heat conduction process.

[0097] The pulsation amplitude, rising edge, and falling edge duration of the pressure sequence are calculated as pressure response parameters. The pulsation amplitude is obtained by calculating the peak-to-valley difference of the pressure signal within a window, and the rising edge and falling edge duration are determined by the length of the monotonic interval of the detection signal, which are used to describe the dynamic response rate of the system pressure.

[0098] The acceleration, hysteresis, and velocity fluctuation rate of the flow sequence are calculated as flow inertial parameters. The acceleration is calculated by the rate of change of velocity between adjacent time steps, the hysteresis is determined by the time difference between the velocity and pressure responses, and the velocity fluctuation rate reflects the stability of fluid flow and is used to capture the inertial characteristics of the system flow.

[0099] Temperature gradient parameters, pressure response parameters, and flow inertia parameters are input into a coupling function for nonlinear fusion. The coupling function is a bilinear mapping function. During the fusion process, learnable weights are assigned to different physical quantities to achieve dynamic coupling between heat, pressure, and flow.

[0100] The fusion results are normalized and subjected to exponential moving average processing. Normalization is used to maintain the numerical stability of the fusion output, while exponential moving average is used to smooth transient fluctuations and preserve long-term trends, thereby generating the physical modulation factor corresponding to each time step.

[0101] In this embodiment, the physical consistency calibration step includes:

[0102] The physical modulation factor sequence output by the thermo-pressure flow physical sensing layer at each time step is obtained, and the modulation factor is normalized according to the time step and channel dimension. The normalization adopts a channel-by-channel linear scaling method to distribute the modulation values ​​of different physical quantities within a uniform range, so as to ensure the numerical stability of the calibration calculation.

[0103] The difference measure between the mean trajectory of the predicted distribution output by the Chronos model at each time step and the trend of the modulation factor change is calculated. The difference measure includes temporal correlation difference and amplitude offset difference. The temporal correlation difference is obtained by calculating the correlation coefficient between the predicted mean change rate and the physical modulation trend. The amplitude offset difference is calculated by the difference between the means and the variance of the two, and is used to reflect the degree of deviation between the prediction result and the physical change.

[0104] A physical calibration weight matrix is ​​generated based on the aforementioned difference measure, and time-weighted smoothing is applied to it according to the importance of the channel. The time-weighted smoothing adopts a recursive weighted average strategy, so that the calibration weight can change dynamically over time without abrupt changes, thereby maintaining the continuity of the calibration process.

[0105] In the temporal causal attention layer of the Chronos model, before embedding the physical calibration weight matrix as modulation coefficients into the attention score calculation, element-wise multiplication is performed on the weight coefficients of each time step during the attention score calculation stage, so that the model can simultaneously consider the constraint information of physical quantity changes when focusing on time dependence.

[0106] The adjusted attention score is re-normalized and forward propagation is completed. After updating the attention weights, the output distribution is recalculated to obtain a physically calibrated predicted distribution that satisfies the physical laws of thermo-pressure flow.

[0107] In this embodiment, the step of generating the multi-step probability prediction trajectory includes:

[0108] The temporal latent feature matrix is ​​expanded by time step to set a context sequence and a prediction window of fixed length, and a corresponding time index is established between the two;

[0109] Time and period codes are added to each time step, including intraday index, intraweek index and rhythmic period identifier. After the time code is generated, it is concatenated with the device identifier and channel identifier embedded in the feature dimension to construct the input tensor of the Chronos model.

[0110] Causal masking is applied to the input tensor, an autoregressive decoding method is adopted, and the sliding window step size and overlap length are set. During the training phase, the model is guided to use real historical values ​​as decoding input through a teacher-forced strategy. At the same time, a random masking strategy is combined to mask some input features to enhance the generalization ability of the model and generate conditional input sequences.

[0111] During the decoding stage, the prediction window is expanded step by step. For each time step, the decoding output of the previous time step is received sequentially as the current input. The conditional probability parameters are calculated and the corresponding quantile prediction values ​​are generated. The prediction results are sequentially superimposed to form a prediction set of continuous time series.

[0112] Shared model parameters are used for different device channels and channel bias embedding is introduced. After decoding, the prediction results of each channel are stacked in dimensions and aligned with the index to generate a multi-step probability prediction trajectory within the target time window.

[0113] In this embodiment, the step of generating the physically calibrated predicted distribution includes:

[0114] The sequence of physical modulation factors output by the thermal-pressure-flow physical sensing layer at each time step is obtained, and each group of modulation factors is read in the order of time step index. They are arranged into a two-dimensional matrix by channel dimension. Linear scaling and zero-mean normalization operations are performed on the data of each channel of the matrix to achieve two-dimensional normalization processing of time step and channel.

[0115] The difference measure between the mean trajectory of the predicted distribution output by the Chronos model at each time step and the trend of the modulation factor is calculated. The calculation of the difference measure includes two parts: first, the correlation coefficient between the time difference sequence of the predicted distribution mean and the time difference sequence of the modulation factor is calculated to obtain the temporal correlation difference; then, the mean difference and variance difference between the two at the same time step are calculated to obtain the amplitude offset difference.

[0116] A physical calibration weight matrix is ​​generated based on the aforementioned difference measure. During the generation process, weight coefficients are assigned to relevant differences and offset differences respectively. After the matrix is ​​constructed, the weights of each channel are subjected to time-weighted smoothing based on the importance index of the channel. The smoothing method adopts a recursive weighted average to avoid abrupt changes in calibration parameters.

[0117] In the temporal causal attention layer of the Chronos model, before embedding the physical calibration weight matrix as a modulation coefficient into the attention score calculation, an element-wise multiplication operation is performed on the score vector of each time step and the physical calibration weight matrix when performing the dot product operation of the attention score, so that the temporal causal attention is affected by physical constraints during the calculation stage.

[0118] The adjusted attention scores are re-normalized and forward propagation is completed. During the forward propagation phase, the output distribution is recalculated using the updated attention weights, gradually replacing the uncalibrated prediction results, and finally obtaining the physically calibrated prediction distribution, which provides input for subsequent health scoring and fault probability calculation.

[0119] In this embodiment, the steps for generating the prediction lead time and the early warning level include:

[0120] The physically calibrated prediction distribution is grouped by time step and device channel. During the grouping process, a mapping table is established based on the device number and channel index. The prediction results of multiple channels of the same device are merged into a set of input data and input into the probability mapping layer of the decoding module. During the input stage, the time step order is kept continuous to ensure the temporal consistency of the prediction results.

[0121] In the probability mapping layer, parameter decoupling and reorganization are performed on the predicted distribution at each time step. By independently extracting and reorganizing the mean, variance and skewness parameters of the predicted distribution, a corresponding set of feature parameters is generated, and the corresponding health state vector is calculated.

[0122] The health status vector is input into the linear transformation layer and the normalization layer. In this stage, a linear weighting operation is performed on the status vector of each channel to compress the dimension. Then, the normalization operation is used to eliminate the dimensional differences between channels of different devices to obtain a standardized health score matrix.

[0123] The continuous rate of change over time is calculated based on the health score sequence. The sliding window method is used to perform difference operations on the health scores of adjacent time steps to obtain the rate of change sequence. Combined with the variance and confidence boundary of the prediction distribution, the failure probability value of each time step is calculated through the uncertainty reflected by the variance and the dispersion of the confidence boundary, thus forming a time-seriesd failure probability distribution.

[0124] Quantile analysis is performed on the fault probability distribution. At each time step, the high quantile and low quantile values ​​are calculated to determine the upper and lower limits of the confidence interval. The prediction lead time is determined based on the intersection of different quantile boundaries. The position where the high quantile boundary first crosses the low quantile boundary between adjacent time steps is marked as the prediction lead point.

[0125] Early warning levels are divided based on the prediction lead time and the probability level of failure. Multiple thresholds are set to correspond to different risk levels. The prediction lead time and risk level are combined to form an early warning label table, and the corresponding early warning results and confidence interval parameters are output.

[0126] In this embodiment, the step of generating the prediction residual includes:

[0127] Obtain operational feedback data corresponding to the prediction time window, collect real-time monitoring values ​​such as temperature, pressure, flow and current from the fire protection system monitoring terminal, organize them according to equipment number and time index, perform time alignment and format conversion on the feedback data, use a unified time reference to interpolate and resample signals of different sampling periods, and format the data of each channel into a tensor structure that is the same as the prediction result.

[0128] The feedback data and prediction results are compared at the same time step. For each time step, the health score residual and the failure probability residual are calculated. The residual is the difference between the predicted value and the actual value. The residual matrix is ​​generated by stacking the data according to the device channel dimension and then expanding it along the time dimension to form a residual sequence.

[0129] The residual sequence is subjected to moving average and outlier filtering. A fixed-length sliding window is set to calculate the local mean. Outliers that deviate from the window mean by more than a threshold are replaced by median or interpolation to obtain a smooth residual signal.

[0130] The physical modulation parameters in the thermal-pressure-flow physical sensing layer are updated based on the smoothed residual signal. The modulation coefficients of temperature gradient, pressure response and flow inertia are recalculated at each time step. The residual trend is mapped to the weight gain factor and the parameter table is updated to achieve dynamic correction of physical parameters.

[0131] Meanwhile, the physical calibration weights of the Chronos model are updated based on the mean square deviation of the residuals of each channel, and the weight matrix of the time-causal attention layer is adjusted incrementally element by element for each channel, so that the weights of high residual channels are appropriately increased to strengthen their calibration constraint capabilities.

[0132] After the update is completed, forward propagation is re-executed, using the corrected physical modulation parameters and calibration weights as inputs, to recalculate the prediction distribution and generate new output results. The updated prediction results are then compared with the previous results to complete the prediction residual generation and model correction process.

[0133] Example 1:

[0134] To verify the feasibility of this invention in practice, it was applied to a fault prediction scenario for the fire protection system of a high-rise complex. The building has 18 floors, and the fire protection system includes 4 fire pumps, 2 sets of sprinkler valves, 6 pressure monitoring nodes, 12 flow sensors, 24 smoke detectors, and several environmental temperature and humidity monitoring points. The system deployment period is 90 days of continuous operation, with a sampling interval of 30 seconds, and a total data volume of approximately 1.2 × 10⁻⁶. 8 All sensor signals are aggregated through an edge computing gateway and then transmitted to the "Fault Prediction Platform Based on Mamba-Chronos Fusion Model" described in this invention for online learning and predictive inference.

[0135] The main parameters collected in the experiment included temperature (°C), pressure (MPa), flow rate (L / min), current (A), and ambient humidity (%). Fire pump P1, sprinkler valve assembly V1, and pressure node N3 were used as typical test objects. After time synchronization, outlier correction, and normalization, the data were converted into a standardized input tensor and input to the thermo-pressure-flow physical sensing layer of the Mamba encoder. The sensing layer calculated the temperature gradient, pressure response, and flow inertia based on the numerical changes of adjacent time steps, generating a physical modulation factor. The Chronos model performed multi-step predictions on the latent features and performed consistency calibration between the predicted distribution and the physical modulation factor. The output results included two indicators: health score and failure probability. Table 1 shows some prediction results within the selected typical sampling period (day 25 to day 30).

[0136] Table 1 Comparison of Key Parameters and Predicted Results for Firefighting Equipment

[0137] Time (h) Equipment Number Temperature (°C) Pressure (MPa) Flow rate (L / min) Current (A) Health score S_h Failure probability P_f Actual state 00:00 P1 28.3 0.86 162 3.2 0.92 0.04 normal 04:00 P1 31.7 0.89 157 3.4 0.87 0.09 normal 08:00 P1 36.4 0.92 154 3.6 0.81 0.18 slight fluctuations 12:00 P1 42.8 0.97 146 3.9 0.72 0.33 Mild abnormality 16:00 P1 48.6 1.01 138 4.2 0.61 0.52 Warning triggered 20:00 P1 53.2 1.05 131 4.5 0.54 0.66 Anomaly Confirmation 00:00 V1 27.5 0.83 168 2.8 0.94 0.03 normal 08:00 V1 29.9 0.88 161 2.9 0.89 0.08 normal 16:00 V1 34.5 0.95 149 3.2 0.76 0.27 Potential anomalies 00:00 N3 25.1 0.79 173 2.4 0.96 0.02 normal 08:00 N3 27.8 0.85 167 2.5 0.90 0.06 normal 16:00 N3 39.1 0.94 158 2.8 0.74 0.29 Warning triggered

[0138] As shown in Table 1, during the continuous sampling period, the temperature of device P1 gradually increased from day 25 to day 30, while the pressure and flow rate showed a clear inverse trend. The Mamba encoder's thermo-pressure-flow physical sensing layer detected a continuous increase in temperature gradient, pressure response delay, and flow inertia shift during this stage. Its overall modulation factor decreased by approximately 35%, leading to an automatic contraction of the weights in the state transition matrix and a significant increase in the hidden state change rate. The Chronos model captured the deviation between the predicted distribution and the physical trend during the calibration phase, automatically adjusting the attention weights and shifting the prediction focus from high-frequency noise signals to low-frequency physical trends. Finally, at 16:00 on day 28, the system predicted a failure probability of device P1 exceeding 0.5 and issued a level-two warning, approximately 32 hours in advance of the actual excessive temperature rise and pressure deviation alarm events.

[0139] For valve group V1 and node N3, the model maintained a health score S_h greater than 0.85 within the stable temperature and pressure range, and the predicted results were consistent with the measured conditions, with no false alarms. On the night of day 29, the flow signal at node N3 suddenly dropped by about 8%. The model increased the failure probability to 0.29 in the following two time steps, triggering the warning indicator but not reaching the alarm threshold. Field verification showed that this was a short-term fluctuation in water supply, indicating that the model has a good balance between fault tolerance and sensitivity.

[0140] Experimental results show that this invention can accurately identify subtle abnormal trends in the operation of fire-fighting equipment and achieve early fault prediction through the combined effect of the thermo-pressure-flow physical sensing layer and the Chronos time-series prediction model. During 90 consecutive days of operation, the model's average prediction lead time was 26.4 hours, with an accuracy rate of 94.7% and a false alarm rate of less than 3%. Without increasing additional hardware costs, the system can monitor the status of fire-fighting equipment in real time, achieving intelligent early warning and dynamic maintenance decision-making, thus verifying the feasibility and practical value of this invention.

[0141] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fire extinguishing equipment failure prediction method based on big data analysis, characterized by, The method comprises the following steps: Collecting temperature, pressure, flow, current and environmental parameters of the fire-fighting equipment, time-synchronizing, abnormal point correcting and normalizing the data to obtain a standardized input tensor; Inputting the input tensor into a Mamba encoder, the Mamba encoder comprising a heat-pressure-flow physical perception layer, generating a physical modulation factor according to a temperature gradient, a pressure response and a flow inertia, and performing weighted correction on a state transition matrix according to the physical modulation factor to output a time-series hidden feature matrix; Inputting the time-series hidden feature matrix into a Chronos time-series basic model to generate a multi-step probability prediction trajectory within a target time window by the Chronos time-series basic model; Performing consistency calibration on a Chronos prediction distribution by using the physical modulation factor output by the heat-pressure-flow physical perception layer, adjusting time-causal attention weights when the prediction result is inconsistent with the physical perception feature to obtain a physically calibrated prediction distribution; Inputting the physically calibrated prediction distribution into a decoding module to output a health score and a failure probability distribution, and generating a prediction lead time and a warning level according to a position information interval; Collecting equipment operation feedback data, calculating a prediction residual, and correcting physical modulation parameters in the heat-pressure-flow physical perception layer and calibration weights of the Chronos model according to the residual.

2. The fire-fighting equipment failure prediction method based on big data analysis according to claim 1, characterized in that, The generation step of the standardized input tensor comprises: Collecting original operation data by multiple types of sensors installed on fire-fighting pumps, spray valve groups, pipe network nodes and environmental monitoring points, and adding timestamp information to each group of sampling data; Taking a unified time reference signal as a reference, performing interpolation resampling on the sensor signals with different sampling periods; Performing abnormality detection on the collected original signals, identifying mutation points or drift points based on sliding window statistics and neighborhood density judgment, and correcting by using the mean value of adjacent windows; Performing normalization processing on the corrected signals, and combining minimum-maximum mapping and zero-mean normalization to standardize the numerical values of each channel data to form an input matrix with unified dimensions; Splicing and arranging the input matrix according to the time sequence dimension to form a standardized input tensor.

3. The fire-fighting equipment failure prediction method based on big data analysis according to claim 1, characterized in that, The generation step of the time-series hidden feature matrix comprises: Mapping the standardized input tensor into encoding input channels according to the device type and physical quantity dimension, and setting a sliding window and a batch length in the time step dimension; In the heat-pressure-flow physical perception layer, calculating a temperature gradient, a pressure response and a flow inertia according to the numerical change of adjacent time steps and the statistical quantity in the window, wherein the temperature gradient is based on the time difference or local slope of the temperature sequence, the pressure response is based on the fluctuation amplitude and rising / falling edge duration of the pressure sequence, and the flow inertia is based on the acceleration feature and lag quantity of the flow sequence; Combining the temperature gradient, the pressure response and the flow inertia according to a preset channel mapping relationship and a device type index to obtain a corresponding physical feature vector, and generating a physical modulation factor by linear transformation and normalization; The state transition matrix of the Mamba encoder is weighted and corrected according to the physical modulation factor, including updating the weight of the matrix row and column channel by channel, and iteratively applying the corrected state transition coefficient to update the hidden state at each time step; After the weighting correction and hidden state update are completed, the hidden state outputs of each time step are collected in chronological order to form a time sequence hidden feature matrix.

4. The fire-fighting equipment failure prediction method based on big data analysis according to claim 3, characterized in that, The generation of the physical modulation factor includes: The time difference mean and variance of the temperature sequence are calculated respectively within the set time sliding window as the temperature gradient parameter; The pulsation amplitude, rising edge and falling edge duration of the pressure sequence are calculated as the pressure response parameter; The acceleration, hysteresis and flow fluctuation rate of the flow sequence are calculated as the flow inertia parameter; The temperature gradient parameter, pressure response parameter and flow inertia parameter are input into the coupling function for nonlinear fusion, and the coupling function is a bilinear mapping function; The fusion result is normalized and exponentially smoothed to generate the physical modulation factor corresponding to the time step.

5. The fire-fighting equipment failure prediction method based on big data analysis according to claim 1, characterized in that, The generation steps of the multi-step probability prediction trajectory include: The time sequence hidden feature matrix is unfolded by time step, a fixed length context sequence and a prediction window are set, and a corresponding time index is established between them; A time coding and a period coding are added to each time step, including an intra-day index, a week index and a rhythm period identifier, and are spliced with device identifier embedding and channel identifier embedding to form a Chronos input tensor; The input tensor is applied with a causal mask, an autoregressive decoding method is used, and a sliding window step and an overlap length are set, and a teacher forced and random mask strategy is used in the training stage to generate a conditional input sequence; In the decoding stage, the prediction window is unfolded step by step, and the quantile prediction corresponding to the conditional probability parameter set and the quantile set is output for each time step; Shared model parameters are used for different device channels, and channel bias embedding is introduced, and the multi-channel output is stacked by device and time step to obtain the multi-step probability prediction trajectory in the target time window.

6. The fire-fighting equipment failure prediction method based on big data analysis according to claim 1, characterized in that, The generation steps of the physically calibrated prediction distribution include: The physical modulation factor sequence output by the heat-pressure-flow physical perception layer at each time step is obtained, and the modulation factor is normalized by time step and channel dimension; The difference measure between the mean trajectory of the prediction distribution output by the Chronos model at each time step and the change trend of the modulation factor is calculated, and the difference measure includes a time sequence correlation difference and an amplitude offset difference; A physical calibration weight matrix is generated based on the difference measure, and it is time-weighted and smoothed according to the importance of the channel; In the time-causal attention layer of the Chronos model, the physical calibration weight matrix is embedded as a modulation coefficient before calculating the attention score, and the attention score is weighted and adjusted element by element; The adjusted attention score is re-normalized and forward propagated to update the attention weight, and the output distribution is recalculated to obtain the physically calibrated prediction distribution.

7. The fire-fighting equipment failure prediction method based on big data analysis as claimed in claim 1, wherein, The generation steps of the prediction lead time and warning level include: grouping the physically calibrated prediction distribution by time step and device channel, and inputting into the probability mapping layer of the decoding module; performing parameter decoupling and reorganization on the prediction distribution of each time step in the probability mapping layer, and calculating the corresponding health state vector; inputting the health state vector into the linear transformation layer and the normalization layer to generate the health score of each device channel; calculating the time continuous change rate according to the health score sequence, and combining the variance and confidence boundary of the prediction distribution to generate the corresponding failure probability distribution; performing quantile analysis on the failure probability distribution, calculating the upper and lower limits of the confidence interval, and determining the prediction lead time according to the intersection point of different quantile boundaries; dividing the warning level according to the prediction lead time and the failure probability level, and outputting the corresponding warning result and confidence interval parameters.

8. The fire-fighting equipment failure prediction method based on big data analysis according to claim 1, characterized in that, The generation step of the prediction residual includes: obtaining the operation feedback data corresponding to the prediction time window, and performing time alignment and format conversion on the feedback data; comparing the feedback data and the prediction result at the same time step, calculating the health score residual and the failure probability residual, and forming the residual sequence; performing sliding average and outlier filtering on the residual sequence to obtain the smoothed residual signal; updating the physical modulation parameters in the heat-pressure-flow physical perception layer according to the smoothed residual signal, and correcting the temperature gradient, pressure response and flow inertia weight; at the same time, updating the physical calibration weight of the Chronos model according to the mean square deviation of each channel residual, and performing incremental adjustment on the attention weight matrix; after the update is completed, re-performing the forward propagation to generate a new prediction distribution.

Images