A method for health monitoring of a battery stack

Abstract

The application discloses a kind of health monitoring methods for battery stack, comprising: obtaining the electrochemical characteristic data of battery stack in preset timing interval, and obtaining the mechanical vibration intensity data synchronized with the electrochemical characteristic data;According to the mechanical vibration intensity data, the electrochemical characteristic data is dynamically affected weight suppression, attenuates or shields the feature weight of the timing segment with transient mechanical vibration noise, obtains the feature after inhibition;Obtain the spatial physical environment characteristics representing the three-dimensional space deployment position where the battery stack is currently located, and generate diagnostic decision sensitivity adjustment factor according to the spatial physical environment characteristics;Using the analysis model configured with the diagnostic decision sensitivity adjustment factor, the inhibited feature is processed, and the health state estimation result of the battery stack is generated and output.The application improves the robustness and accuracy of battery stack health state monitoring.

Description

Technical Field

[0001] This invention relates to the field of battery stack management technology, and more specifically, to a method for health monitoring of battery stacks. Background Technology

[0002] Health assessment of battery stacks typically relies on collecting electrochemical data (such as voltage, current, and internal resistance) during operation and inputting it into analytical models to identify degradation characteristics or potential failures. This conventional monitoring logic is based on the implicit assumption that the electrochemical signals acquired by sensors can purely and stably map the true chemical state inside the battery.

[0003] However, in actual industrial or outdoor physical deployment environments, this implicit assumption often proves difficult to uphold. On the one hand, mechanical vibrations accompanying equipment operation or external intervention directly affect the battery body and test circuit, superimposing non-stationary transient disturbances into the electrochemical signal, causing the monitoring data to deviate from its inherent electrochemical characteristics. On the other hand, the gradual degradation rate inside the battery is profoundly affected by the thermodynamic conditions of its microenvironment; ventilation conditions, physical shielding, and temperature differences at different three-dimensional spatial locations result in the same type of battery exhibiting drastically different degradation evolution benchmarks at different deployment locations.

[0004] Existing health monitoring solutions typically treat electrochemical signals as isolated input dimensions for static analysis. Because transient physical disturbances are not isolated from the monitoring signals, the system is prone to misinterpreting vibration-induced signal distortion as electrical anomalies. Furthermore, since environmental heterogeneity is not incorporated into the diagnostic logic, static models struggle to adapt to baseline shifts under varying heat dissipation conditions, leading to missed early degradation signals in harsh environments or false alarms in favorable environments. This approach of directly applying ideal models to complex physical realities results in distorted battery stack health status estimates, failing to meet the practical requirements for highly robust and accurate monitoring of battery stack health status under complex physical environments. Summary of the Invention

[0005] The purpose of this invention is to provide a health monitoring method for battery stacks, which solves the problem of distorted health status assessment of battery stacks in the prior art due to the inability to isolate transient mechanical vibration disturbance noise and the failure to consider the heterogeneity of the spatial environment, and meets the requirements of high robustness and high precision monitoring of battery stack health status in complex physical environments.

[0006] To achieve the above-mentioned objectives, embodiments of the present invention provide the following technical solutions:

[0007] On one hand, the present invention provides a health monitoring method for battery stacks, comprising:

[0008] Acquire electrochemical characteristic data of the battery stack within a preset time interval, and acquire mechanical vibration intensity data synchronized with the electrochemical characteristic data;

[0009] Based on the mechanical vibration intensity data, the electrochemical characteristic data is subjected to dynamic influence weight suppression to obtain suppressed features; wherein, the suppression intensity of the dynamic influence weight suppression increases with the increase of the mechanical vibration intensity data, so as to attenuate or shield the feature weights of time segments with transient mechanical vibration noise in at least the multiple time segments divided within the preset time interval;

[0010] Acquire spatial physical environment characteristics that characterize the current three-dimensional deployment location of the battery stack;

[0011] Based on the characteristics of the space physical environment, a diagnostic sensitivity adjustment factor is generated to control the sensitivity to progressive degradation signals; wherein, the progressive degradation signal is an electrochemical characteristic change signal characterizing the gradual deterioration of the battery stack over time; and

[0012] The suppressed features are processed using an analysis model configured with the diagnostic sensitivity adjustment factor. The diagnostic sensitivity adjustment factor is used to scale the hidden features extracted by the analysis model, and the health status estimation result of the battery stack is generated and output.

[0013] In one possible embodiment, the method further includes:

[0014] The cumulative charge capacity is obtained by integrating the current data of the charging cycle in the electrochemical characteristic data over time, and a capacity-voltage curve is generated by combining it with the synchronous voltage data.

[0015] By performing a differential operation on the capacity-voltage curve, the data interval in which the absolute value of the voltage-capacity differential is less than a preset threshold for a consecutive preset number of sampling points is identified as a voltage plateau region, and the corresponding plateau region voltage and plateau region width are extracted; wherein, the plateau region width is the capacity change corresponding to the voltage plateau region, and the plateau region voltage is the average voltage within the voltage plateau region.

[0016] In one possible embodiment, the step of dynamically weighting the electrochemical characteristic data based on the mechanical vibration intensity data to obtain the suppressed features includes:

[0017] Within the time interval, the mechanical vibration intensity data is divided into multiple consecutive time segments, and the mechanical vibration intensity metric value within each time segment is calculated.

[0018] Based on the mechanical vibration intensity metric value within the time segment, a one-dimensional attention suppression mask is generated;

[0019] Based on the mechanical vibration intensity data and the electrochemical feature data, correlation feature calculation is performed to generate an original attention weight matrix characterizing the correlation between each time segment and the electrochemical feature;

[0020] Using the one-dimensional attention suppression mask, the weights of the time-series segments containing the transient mechanical vibration noise in the original attention weight matrix are attenuated or masked to obtain the suppressed attention weight matrix.

[0021] The suppressed attention weight matrix is ​​fused with the value matrix obtained by encoding the electrochemical feature data and copying and expanding it according to a preset time correspondence, and a fused feature matrix is ​​obtained. Based on the one-dimensional attention suppression mask, the fused feature matrix is ​​subjected to masked global average pooling along the time dimension to obtain a fused feature vector, which is used as the suppressed feature.

[0022] In one possible embodiment, generating a one-dimensional attention suppression mask based on the mechanical vibration intensity metric within the time segment includes:

[0023] Calculate the root mean square value of the mechanical vibration intensity data within each time segment, and arrange the root mean square values ​​according to time sequence;

[0024] Using the natural constant The inverse logistic decay function with base is used to calculate a sequence of suppression coefficients based on the root mean square value, a preset suppression threshold, and a smoothing factor to control the steepness of the transition, thus forming the one-dimensional attention suppression mask; wherein, the expression for the inverse logistic decay function is:

[0025] in, For the first The suppression coefficient of each time segment, For the first The root mean square value of each time segment The preset suppression threshold, A smoothing factor to control the steepness of the transition; when When approaching zero, A value close to 1 indicates that the feature weights of the corresponding time segment are fully preserved; when Significantly greater than hour, A value close to 0 indicates that the feature weights of the corresponding time segment are masked.

[0026] In one possible embodiment, the step of performing correlation feature calculation based on the mechanical vibration intensity data and the electrochemical feature data to generate an original attention weight matrix characterizing the correlation between each time segment and the electrochemical feature includes: encoding the mechanical vibration intensity data through a one-dimensional convolutional neural network to obtain a mechanical vibration time-series feature matrix, which serves as a query matrix for cross-attention; copying and expanding the encoded electrochemical feature data according to a preset time correspondence to obtain a bond matrix and a value matrix; performing a scaled dot product operation between the query matrix and the bond matrix and then normalizing it using Softmax to generate the original attention weight matrix.

[0027] The step of using the one-dimensional attention suppression mask to attenuate or mask the weights of time segments with transient mechanical vibration noise in the original attention weight matrix to obtain a suppressed attention weight matrix includes: broadcasting and expanding the one-dimensional attention suppression mask along the column direction, and then performing a Hadamard product with the original attention weight matrix to attenuate or mask the corresponding attention weights of time segments with transient mechanical vibration noise during polymerization electrochemical characteristics, thereby obtaining the suppressed attention weight matrix.

[0028] In one possible embodiment, obtaining the spatial physical environment characteristics characterizing the current three-dimensional deployment location of the battery stack includes:

[0029] A request containing the spatial location identifier of the battery stack is sent to a remote building information modeling platform via a data interface;

[0030] The system receives the spatial physical environment features returned by the building information modeling platform based on the call request; wherein the spatial physical environment features include at least an estimated ambient temperature, a ventilation condition status code, a physical obstruction coefficient, and a spatial relative position scalar.

[0031] In one possible embodiment, generating a diagnostic sensitivity adjustment factor based on the characteristics of the space physical environment to control the sensitivity to progressive degradation signals includes:

[0032] The spatial physical environment features are input into the environmental encoder to generate the diagnostic judgment sensitivity adjustment factor; wherein, the diagnostic judgment sensitivity adjustment factor is a scalar.

[0033] Among the multiple encoder layers included in the analysis model, a set of candidate encoder layers is determined whose layer numbers fall within the range of one-third to two-thirds of the total number of layers. When the endpoints of the range correspond to non-integer layer numbers, the lower limit is rounded up and the upper limit is rounded down. The layer with the middle number is selected from the set of candidate encoder layers as the target injection layer. When the set of candidate encoder layers contains an even number of layer numbers, the middle layer number that is slightly later is selected.

[0034] In the target injection layer, the diagnostic judgment sensitivity adjustment factor is used to perform element-wise scaling modulation on the hidden state feature tensor output by the multi-head self-attention sublayer of the target injection layer, so as to control the judgment sensitivity to the progressive degradation signal.

[0035] In one possible embodiment, the step of processing the suppressed features using an analytical model configured with the diagnostic judgment sensitivity adjustment factor to generate and output the health status estimation result of the battery stack includes:

[0036] The suppressed features are copied and extended along the time dimension to form a time sequence. The one-dimensional attention suppression mask is downsampled to the number of sampling times of the electrochemical feature data and used as a gating weight. The electrochemical feature data after moving average smoothing is subjected to element-wise multiplication weighting and then concatenated with the time sequence along the feature dimension to construct a multi-channel input tensor.

[0037] The multi-channel input tensor is mapped to an initial hidden sequence through an input linear projection layer, and after encoding at the superposition position, it is input into the multi-level cascaded encoder contained in the analysis model.

[0038] Features are extracted layer by layer by the multi-level cascaded encoder, and in the target injection layer, the hidden state feature tensor output by the multi-head self-attention sub-layer of the layer is subjected to element-wise scaling modulation using the diagnostic judgment sensitivity adjustment factor.

[0039] Global average pooling is performed on the feature sequence output by the multi-stage cascaded encoder to obtain a comprehensive feature vector.

[0040] The comprehensive feature vector is input into the regression head and classification head of the analysis model respectively to generate a comprehensive prediction result containing continuous numerical values ​​of battery stack health status and probability distribution of fault categories, which is used as the health status estimation result of the battery stack.

[0041] In one possible embodiment, the battery stack type includes a lead-acid battery stack or a lithium-ion battery stack;

[0042] When the battery stack is a lead-acid battery stack, the progressive degradation signal includes at least one of the following detection items:

[0043] DC internal resistance increasing trend: By calling the local or cloud database through the data interface, a long periodic sequence composed of multiple consecutive historical internal resistance measurements is obtained. The slope of internal resistance change is calculated by linear regression. When the slope is greater than zero and the cumulative increase in internal resistance exceeds the preset internal resistance growth threshold, it is determined that a DC internal resistance increasing trend degradation signal has appeared.

[0044] Abnormal deviation of platform voltage: Extract the platform voltage of the current charging cycle and compare it with the pre-calibrated reference platform voltage. When the deviation of the platform voltage from the reference platform voltage exceeds a preset deviation threshold, it is determined that an abnormal deviation degradation signal of platform voltage has occurred.

[0045] When the battery stack is a lithium-ion battery stack, the progressive degradation signal includes at least one of the following detection items:

[0046] Coulomb efficiency continuously decreasing trend: By calling the local or cloud database through the data interface, a long period sequence composed of coulomb efficiencies of multiple consecutive historical charge-discharge cycles is obtained. The slope of coulomb efficiency change is calculated by linear regression. When the slope is less than zero and the cumulative decrease in coulomb efficiency exceeds the preset decrease threshold, it is determined that a degradation signal of continuous decrease in coulomb efficiency has occurred.

[0047] Abnormal deviation of platform voltage: Extract the platform voltage of the current charging cycle and compare it with the pre-calibrated reference platform voltage. When the deviation of the platform voltage from the reference platform voltage exceeds a preset deviation threshold, it is determined that an abnormal deviation degradation signal of platform voltage has occurred.

[0048] In one possible embodiment, the method further includes a fault-tolerant degradation step:

[0049] During normal system operation, the spatial physical environment features successfully acquired each time are persistently stored locally, and the corresponding cache timestamp is recorded.

[0050] When the call request sent to the building information modeling platform times out or the connection is abnormal, the latest cached spatial physical environment features are read locally, and the estimated environmental temperature in the cache is replaced with the measured temperature of the environmental temperature sensor pre-deployed locally on the battery stack to obtain the fused and corrected spatial physical environment features.

[0051] The fused and corrected spatial physical environment features are input into the environment encoder to generate the diagnostic judgment sensitivity adjustment factor, maintaining the continuous operation of the diagnostic process; and a degradation mark and the cache timestamp are added to the health status estimation result.

[0052] In one possible embodiment, the method further includes a step of dynamically verifying the type of the battery stack:

[0053] The current feature vector, composed of the platform region voltage and the platform region width, is normalized according to preset calibration parameters, and then the normalized Euclidean distance is calculated with the pre-stored standard battery operation feature vector.

[0054] When the normalized Euclidean distance exceeds the consistency tolerance threshold for a certain number of consecutive preset times, a battery type verification inconsistency alarm signal is triggered.

[0055] In one possible embodiment, when the battery type mismatch alarm signal is triggered, the method further includes performing a hardware-level forced protection step:

[0056] Use the battery type inconsistency alarm signal or hardware-level voltage abnormality signal as the protection trigger source;

[0057] In response to the activation of the protection trigger source, the main charging and discharging circuit is forcibly disconnected via a hardware latching circuit independent of the upper-level software control signal; wherein, the hardware latching circuit has a self-holding characteristic to continuously lock the disconnected state during power supply maintenance; and

[0058] In response to receiving a matching on-site authorization verification command and a system self-test pass signal, the self-holding characteristic of the hardware latch circuit is released to restore the controllable state of the charging and discharging main circuit.

[0059] The beneficial effects of this invention are as follows:

[0060] This invention provides a health monitoring method for battery stacks. By introducing synchronized mechanical vibration intensity data and spatial physical environment characteristics, a dynamic diagnostic mechanism is established that combines filtering of underlying physical interference with macroscopic environmental perception. The resulting technical effects are as follows:

[0061] First, this invention utilizes mechanical vibration intensity data to perform dynamic influence weight suppression on electrochemical characteristic data, establishing an adaptive adjustment logic between physical disturbances and electrochemical characteristic analysis. This mechanism, without altering the underlying electrochemical sampling hardware, attenuates or shields data distortion fragments caused by transient mechanical vibrations at the data characteristic level, blocking the transmission of physical noise to the analysis model, restoring the purity of the electrochemical characteristics used for health estimation, and effectively reducing the system's false alarm rate.

[0062] Secondly, this invention generates a diagnostic sensitivity adjustment factor by acquiring the spatial physical environment characteristics of the battery stack's three-dimensional deployment location, transforming the original static analysis model into an adaptive evaluation system for the deployment environment. This mechanism bridges the logical gaps caused by the lack of environmental context in traditional methods, enabling the model to dynamically calibrate the judgment scale for progressive degradation signals based on the actual boundary conditions faced by the battery, such as heat dissipation and shading. The combination of these two aspects allows this method to output objective and accurate battery stack health status estimates even under conditions of continuous physical disturbances and highly heterogeneous spatial environments.

[0063] In summary, this invention addresses two problems mentioned in the background art: transient mechanical vibration contamination of electrochemical signals and distortion of the baseline due to the heterogeneity of the deployment space environment. These problems are solved by dynamic influence weight suppression driven by mechanical vibration intensity data and adaptive adjustment of diagnostic sensitivity driven by space physical environment characteristics, respectively. Both mechanisms share the same analytical model, jointly optimizing during the training phase and working synergistically during the inference phase. This achieves the technical effect of reducing false alarm and false negative rates and improving the accuracy of health status estimation, thus realizing the invention's objective. Attached Figure Description

[0064] Figure 1 A general flowchart of a health monitoring method for battery stacks provided in an embodiment of the present invention;

[0065] Figure 2 This is a schematic diagram of the temporal segmentation of mechanical vibration intensity data provided in an embodiment of the present invention;

[0066] Figure 3 This is a schematic diagram illustrating the cross-modal cross-attention fusion and masking application provided in an embodiment of the present invention;

[0067] Figure 4 Hardware-level physical latching logic circuit diagram provided for embodiments of the present invention;

[0068] Figure 5 This is a schematic diagram showing the comparison of the timing waveforms of mechanical vibration intensity, root mean square value, and one-dimensional attention suppression mask provided in an embodiment of the present invention.

[0069] Explanation of reference numerals in the attached diagram: 510, hardware voltage comparator; 515, OR gate; 520, main control microcontroller; 530, RS flip-flop; 540, relay drive circuit; 550, solid-state relay. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the present invention clearer, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0071] Terminology: BIM – Building Information Modeling; SOH – Battery Stack Health Status; RMS – Root Mean Square Value; BMS – Battery Management System; MEMS – Microelectromechanical System; NTC – Negative Temperature Coefficient Thermistor; SSR – Solid State Relay; GUID – Globally Unique Identifier; MHSA – Multi-Head Self-Attention Network; FFN – Feedforward Network; PE – Location Code; FC – Fully Connected Network; API – Application Programming Interface.

[0072] In actual engineering operations within complex construction scenarios such as intelligent building, battery stack monitoring data frequently exhibits transient abrupt changes unrelated to the actual state of the batteries. Furthermore, battery stacks of the same batch and model, deployed in areas with varying ventilation, shielding, and temperature conditions within battery swapping stations, show significant differences in their degradation processes. Mechanical vibration is not merely a measurement error, but rather a non-stationary transient disturbance directly superimposed on the electrochemical signal. Without separating this at the feature level, any degradation identification relying on the assumption of a pure signal will be distorted. The spatial environment is not merely peripheral information for diagnosis, but a key variable determining the degradation baseline; ignoring environmental differences at deployment locations inevitably leads to inappropriate diagnostic sensitivity settings. Based on these understandings, this invention combines "dynamic influence weight suppression driven by mechanical vibration intensity" with "adaptive adjustment of diagnostic sensitivity driven by the spatial physical environment," and performs joint training and collaborative inference within the same analytical model. The following embodiments are based on the above concepts and are used to fully implement and verify this technical concept.

[0073] In some embodiments, please refer to Figure 1 This invention provides a method for monitoring the health of battery stacks, applied to a hybrid mobile energy storage battery swapping station in an underground parking lot under construction. The station is equipped with both lead-acid and lithium iron phosphate battery stacks (a type of lithium-ion battery stack), respectively providing emergency backup power and battery swapping services for electric construction machinery in the construction area, deployed within a smart construction scenario. The deployment area of ​​this battery swapping station experiences continuous background mechanical vibration, and due to nearby piling and compaction operations, occasional instantaneous impact vibration acceleration peaks can reach 1.8g; the ambient temperature fluctuates widely, and ventilation conditions in some enclosed areas are extremely poor. The technical problem solved by this invention is to achieve highly robust monitoring of the battery stack's health status under these complex working conditions. The method includes:

[0074] S100: Obtain electrochemical characteristic data and synchronous mechanical vibration intensity data for the preset time interval of the battery stack.

[0075] S200. Based on the mechanical vibration intensity data, the electrochemical characteristic data are dynamically affected by weight suppression to obtain the suppressed characteristics.

[0076] S300: Obtain spatial physical environment characteristics that characterize the current three-dimensional spatial deployment location of the battery stack.

[0077] S400: Generate a diagnostic sensitivity adjustment factor based on the characteristics of the space physical environment.

[0078] S500: Using an analysis model configured with the diagnostic judgment sensitivity adjustment factor, the suppressed features are processed to generate and output the health status estimation results of the battery stack.

[0079] In some embodiments, the data acquisition process specifically includes: setting up voltage acquisition circuits on the positive and negative terminals of the battery stack to acquire the total voltage of the battery stack and the voltage of each individual cell at a sampling frequency of 1Hz; connecting a Hall current sensor in series in the main circuit to acquire the charging and discharging current at a sampling frequency of 1Hz; arranging an NTC thermistor temperature sensor on the surface of the battery stack to acquire the battery surface temperature at a sampling frequency of 1Hz; setting up an independent ambient temperature sensor near the battery stack deployment location to acquire the ambient temperature at a sampling frequency of 1Hz; measuring the DC internal resistance every 10 seconds using a pulse discharge method, and extending the internal resistance value to each 1Hz sampling moment using a zero-order hold method; installing a triaxial MEMS accelerometer at the bottom of the battery swapping station to acquire vibration acceleration signals in real time at a sampling frequency of 500Hz as mechanical vibration intensity data. Since the sampling frequency (1Hz) of the electrochemical characteristic data and the sampling frequency (500Hz) of the mechanical vibration intensity data differ by orders of magnitude, this invention aligns the two data using the same time base and achieves alignment and fusion using a time-series segmentation method. In this embodiment, the preset time interval is 100 seconds.

[0080] It should be noted that the internal resistance measurement frequency is once every 10 seconds, and approximately 10 internal resistance measurement values ​​are obtained within a 100-second time interval. To construct 100-point four-channel timing data synchronized with voltage, current, and temperature, the sampling time between two adjacent internal resistance measurements is extended using zero-order hold (i.e., taking the most recent internal resistance measurement value), so that the internal resistance channel is aligned with the voltage, current, and temperature channels in the time dimension.

[0081] In some embodiments, the dynamic influence weight suppression of electrochemical characteristic data specifically includes:

[0082] S210, Calculation of time-series segmentation and mechanical vibration intensity metrics. Please refer to [link / reference]. Figure 2 The vibration acceleration signal (500Hz sampling rate) collected within a preset time interval (100 seconds in this embodiment) is divided into 200 consecutive time segments according to time sequence. Each time segment corresponds to 0.5 seconds of mechanical vibration data (250 sampling points). For each time segment... ( ), calculate the root mean square value of the vibration acceleration signal within this segment, and use it as a measure of the mechanical vibration intensity of this segment:

[0083] in, The number of sampling points for this segment (in this embodiment) ), For the first The first segment The acceleration values ​​(in g) at each sampling point. It should be noted that in this embodiment, the 200 vibrational time segments, each approximately 0.5 seconds long, correspond exactly to a total time interval of 100 seconds. This is not a one-to-one correspondence with the 100 electrochemical feature sampling points, each lasting 1 second; rather, each electrochemical sampling moment corresponds to two adjacent vibrational time segments. In subsequent steps, the suppression and fusion results of the 200 time segments are converged into a single fusion feature vector representing the whole, and then expanded and aligned.

[0084] S220, One-dimensional attention suppression mask generation. (See also:) Figure 5 According to each time segment Using the natural constant The suppression coefficient is calculated using the inverse logistic decay function with base 0. :

[0085] in, The preset suppression threshold, A smoothing factor is used to control the steepness of the transition. This embodiment determines the smoothing factor through a comparative experiment on a mechanical vibration table. (This corresponds to the statistical upper limit of continuous background vibration at the construction site, as opposed to occasional instantaneous peak impact.) (to concentrate the inhibition transition) Within a very narrow area, ensure high steepness.

[0086] It should be noted that the aforementioned 1.8g vibration peak value is the instantaneous impact acceleration peak value caused by pile driving and other operations near the construction site, which is an outlier sample with an extremely short duration; while The 0.5g value is the upper limit of RMS statistics for normal background vibrations. The two have different statistical calibers but are not contradictory. It is precisely when a transient strong vibration impact of 1.8g occurs that the corresponding time segment... It will significantly exceed After calculation by formula When the value approaches 0, the feature weight of the time segment is effectively suppressed.

[0087] Under these parameter calibration conditions, when When it approaches 0 (extremely weak mechanical vibration). This indicates that the mechanical vibration interference in this time segment is minimal, and the attention weights corresponding to the electrochemical characteristics are almost completely preserved; when Significantly greater than hour, The noise rapidly decays and approaches zero to filter out mechanical vibration noise features. Thus, from the RMS sequence of 200 time segments, a suppression coefficient sequence of length 200 is generated, forming a one-dimensional attention suppression mask. .

[0088] As an example of parameter calibration: apply excitation covering the typical background vibration spectrum and impact spectrum of the construction site on a mechanical vibration table, statistically analyze the RMS distribution of the vibration acceleration signal, and calibrate by taking the statistical upper limit of the background vibration RMS (such as the 95th percentile). ; Selected through grid search, the inhibition coefficient is within Nearby completion by to The steep transition.

[0089] S230, Cross-modal Cross-Attention Fusion and Masking Applications. See also: Figure 3 :

[0090] Step 1: Input the vibration acceleration signals of 200 time segments into a one-dimensional convolutional neural network, and the output shape is as follows: The mechanical vibration time-series feature matrix is ​​used as the query matrix. As an example, this one-dimensional convolutional neural network consists of sequentially connected convolutional layers, batch normalization layers, ReLU activation layers, and pooling layers, mapping each vibration time segment containing 250 sampling points into 64-dimensional features.

[0091] Step 2: Encode the electrochemical feature data (voltage, current, temperature, and internal resistance at 100 sampling times) into a (100, 64) electrochemical time-series feature matrix using a fully connected network (as an example, this encoding network consists of two fully connected layers, mapping the 4-dimensional electrochemical features at each sampling time to 64-dimensional features). Then, perform linear transformations to map the features into bond and value matrices respectively, and copy the features at each sampling time twice along the time dimension (the first copy is the second copy). The electrochemical sampling time corresponds to the first The and the first (Several vibrational time segments) yielded bond matrices of shape (200, 64). Sum matrix This ensures that the key and value positions correspond one-to-one with 200 vibration time segments.

[0092] Step 3: Perform scaled dot product attention and output the attention weight matrix. (shape(200,200)), where elements Indicates the first The vibration timing segment affects the first The correlation of extended electrochemical feature sites.

[0093] Step 4: Apply the suppression mask (Length 200) Broadcast along the column direction to expand into a mask matrix of shape (200, 200). To ensure that the attenuation weights are applied to the attention distribution of the transient mechanical vibration noise time series segment when querying associated electrochemical features, element-wise suppression is performed through Hadamard product to obtain the suppressed attention weight matrix. .

[0094] Step 5: Perform matrix multiplication (shape(200,64)), then based on the mask right Global average pooling is performed along the time dimension, meaning the denominator is the sum of the effective suppression coefficients. Instead of simply using a fixed number of 200 steps, a 64-dimensional fused feature vector is obtained. This is a suppressed feature after adaptive filtering of mechanical vibration noise.

[0095] S300. Obtain the spatial physical environment characteristics representing the three-dimensional spatial deployment location. Specifically, this includes sending a call request containing the battery stack GUID to the remote Building Information Modeling (BIM) platform via API (i.e., data interface). The GUID is a unique identifier of the battery stack used to query the three-dimensional spatial deployment location of the battery stack in the BIM platform. The BIM platform returns four-dimensional spatial physical environment characteristics: an estimated ambient temperature (34℃ in this embodiment, provided by the BIM platform based on environmental monitoring and simulation data of this location), a ventilation condition status code (0 for airtightness in this embodiment, with a value range of [0,1], where 0 represents airtightness and 1 represents complete ventilation), a physical shading coefficient (0.7 in this embodiment, with a value range of [0,1], where 0 represents no shading and 1 represents complete shading), and a spatial relative position scalar (a normalized value representing the spatial distance of the battery stack relative to key locations such as ventilation openings and heat sources, with a value range of [0,1]).

[0096] S400, generation of diagnostic judgment sensitivity adjustment factor. This incorporates four-dimensional spatial physical environment characteristics. The input environment encoder (consisting of two fully connected layers, with ReLU activation between the layers) outputs a 128-dimensional environment embedding vector. ;Will After one linear transformation (weight) bias Mapped to scalar Then through the Sigmoid function Normalized to the [0,1] interval, and finally passed through the mapping function Will Mapping to the [0.5, 1.5] interval yields the diagnostic sensitivity adjustment factor. In this embodiment, the four-dimensional features of the sealed high-temperature zone correspond to... After calculation, the output is... .in, This indicates an increased sensitivity to the detection of progressive degradation signals. This indicates a reduction in the sensitivity of the judgment.

[0097] In the analysis model, the third encoder layer located in the middle section is selected as the target injection layer, and its hidden state feature tensor is analyzed. Perform element-wise scaling modulation: In this embodiment, the analysis model is a 6-layer cascaded encoder, with a total of [number of layers missing]. The interval from one-third to two-thirds corresponds to [⌈6 / 3⌉, ⌊2×6 / 3⌋]=[2,4], i.e., layers 2 to 4, with layer 3 located within this interval. This scaling modulation is jointly backpropagated with the loss functions of the regression head and classification head during the training phase: the training samples are historical running data and their labeled SOH true values ​​and fault category labels; the training objective is to minimize the weighted sum of regression loss (e.g., mean squared error loss) and classification loss (e.g., cross-entropy loss); the parameters of the environment encoder, analysis model, regression head, and classification head are jointly updated during training, making... The value of adapts to the characteristics of the deployment environment. As an example, training uses the Adam optimizer, with an initial learning rate set to . The batch size is set to 64, and the number of training rounds is determined based on the convergence of the validation set loss. Fault category labels in the training samples include, but are not limited to: abnormal DC internal resistance, abnormal Coulomb efficiency, voltage shift in the plateau region, over-temperature, and over / under-voltage. Each encoder layer in the analysis model consists of a multi-head self-attention sublayer and a feedforward network sublayer, employing residual connections and layer normalization. In this embodiment, the number of attention heads is, for example, 8, and the hidden dimension of the feedforward network is, for example, 256. Position encoding uses sinusoidal position encoding or learnable position encoding. Therefore, in environments with characteristics such as high temperature and confined spaces... Under higher deployment conditions, the model becomes more sensitive to progressive degradation signals, enabling adaptive calibration to address the differences in degradation baselines across different deployment environments.

[0098] S500, suppressing the features The sequence is copied 100 times along the time dimension (64-dimensional) to form a sequence of shape (100, 64). The one-dimensional attention suppression mask is downsampled to 100 sampling points and used as a gating weight vector. Element-wise multiplication is then performed on the voltage, current, temperature, and internal resistance time-series data at the 100 sampling times after smoothing by moving average. Then, the purified four-channel time-series data is concatenated with the sequence along the feature dimension to construct a multi-channel input tensor. (A 68-dimensional multi-channel input tensor is formed by concatenating 64-dimensional suppressed features with 4-dimensional electrochemical feature data along the feature dimension). After being mapped by the input linear projection layer and superimposed with position encoding, it is input into a 6-layer cascaded encoder; the 3rd layer utilizes... Injection modulation is performed. After global average pooling, the output sequence is used by the regression head and classification head to generate health status estimation results (SOH values ​​and fault category probability distributions).

[0099] It should be noted that the electrochemical feature time-series data incorporated here is pre-processed data that has undergone smoothing, filtering out high-frequency physical glitches while preserving the original temporal resolution information; combined with the suppressed features extracted through cross-modal adaptive extraction. Together, they completely block the transmission of transient mechanical vibration noise in the multi-channel input tensor.

[0100] In some embodiments, in conjunction with the aforementioned S100 to S500, a complete implementation of the analysis model is as follows: the analysis model is a time-series analysis model based on a Transformer encoder, and the input is a multi-channel input tensor. The output is a continuous numerical value of the battery stack's State of Health (SOH) and a probability distribution of fault categories. The model sequentially includes an input linear projection layer, a position encoding layer, a 6-layer cascaded encoder, a global average pooling layer, a regression head, and a classification head. The initial hidden sequence mapped to (100, 128) (hidden dimension) This is followed by a sinusoidal positional encoding or a learnable positional encoding with a dimension of 128. Each encoder layer consists of a multi-head self-attention sublayer and a feedforward network sublayer, with residual connections and layer normalization used in all sublayers. The number of attention heads in the multi-head self-attention sublayer is, for example, 8, and the dimension of each attention head is, for example, 16 (i.e., 128 / 8). The hidden dimension of the feedforward network sublayer is, for example, 256, and the activation function is ReLU. The third layer is the target injection layer: the hidden state feature tensor output by the multi-head self-attention sublayer in this layer... Execute first The element-wise scaling modulation is then input into the feedforward network sublayer.

[0101] Model output: Global average pooling is performed on the feature sequence output by the 6th layer encoder to obtain a comprehensive feature vector; the regression head consists of a fully connected layer and outputs continuous SOH values; the classification head consists of at least one fully connected layer and a Softmax layer, used to output the probability distribution of fault categories; fault categories include but are not limited to: abnormal DC internal resistance, abnormal Coulomb efficiency, plateau voltage deviation, over-temperature, and over-voltage / under-voltage.

[0102] Training method: Training samples are extracted from historical operating data using a sliding window (window length 100 sampling points corresponding to 100 seconds, step size 50 sampling points), covering different vibration intensity conditions and different spatial environment regions. Each sample contains multi-channel electrochemical time-series data, synchronized mechanical vibration intensity data, spatial physical environment characteristics at the corresponding time, and labeled SOH true values ​​and fault category labels. Before training, z-score normalization is performed on each input channel based on the mean and standard deviation of the training set statistics, and calibration parameters are deployed along with the model. The loss function is a weighted sum of regression loss (mean squared error) and classification loss (cross entropy), for example, both weights are set to 1; the Adam optimizer is used, with an initial learning rate of... With a batch size of 64, training is stopped early based on the convergence of the validation set loss.

[0103] Inference Deployment: During inference, real-time data is captured using a sliding window, and the same data preprocessing as during training is performed; the model is deployed on edge computing devices after INT8 fixed-point quantization, and the third layer still executes... Injection modulation outputs SOH values ​​and fault category probability distributions. The network layer number, hidden dimension, number of attention heads, and other parameters mentioned above are merely examples; those skilled in the art can adjust them according to battery type, sampling rate, and computing resources.

[0104] S600. Dynamic Verification of Battery Type. In some embodiments, the type of the battery stack is dynamically verified: a capacity-voltage (QU) curve is constructed with the cumulative charge capacity obtained by integrating the charging current over time during the charging cycle as the horizontal axis and the corresponding terminal voltage as the vertical axis. The curve is then differentiated to identify voltage plateau regions, and the voltage of these plateau regions is extracted. and platform area width The voltage plateau region is where the absolute value of the differential of voltage with respect to capacity is continuous on the capacity-voltage curve. One sampling point (in this embodiment) For intervals where all values ​​are less than a preset threshold, the width of the plateau region is the corresponding capacity change, and the voltage of the plateau region is the average voltage within that interval. If multiple plateau regions meet the conditions, the one with the largest width is selected. The extracted plateau region voltage and width are used to construct a two-dimensional feature vector. Compared with the pre-stored standard battery operating feature vector After normalization according to preset calibration parameters; wherein, the standard battery operating feature vector is the standard feature vector corresponding to the current registration type of the battery stack, the normalized Euclidean distance between the two is calculated:

[0105] = ;

[0106] in, , , , These are preset calibration parameters. When... ( When the consistency tolerance threshold is reached for a preset number of consecutive times, a battery type verification inconsistency alarm signal is triggered. As an example, the preset threshold for identifying the voltage plateau region can be calibrated based on the noise level of voltage and capacity measurements; the consistency tolerance threshold... It can be set to 0.2 to 0.5, and the number of continuous monitoring times can be set to 3 to 5 times. The specific values ​​are calibrated according to the battery type and the measurement noise level.

[0107] In some embodiments, S700, when the battery type verification inconsistency alarm signal is triggered, the method further includes performing a hardware-level forced protection step to construct a low-level security defense independent of upper-level software control. Specifically, this step includes: using the battery type verification inconsistency alarm signal or a hardware-level voltage abnormality signal as a protection trigger source; in response to the activation of the protection trigger source, forcibly disconnecting the charging and discharging main circuit through a hardware latching circuit independent of the upper-level software control signal; wherein the hardware latching circuit has a self-holding characteristic to continuously lock the disconnected state during power supply maintenance; and in response to receiving a matching field authorization verification command and a system self-test pass signal, releasing the self-holding characteristic of the hardware latching circuit to restore the controllable state of the charging and discharging main circuit.

[0108] The beneficial effect of this design is that conventional battery management systems rely heavily on software commands to disconnect the main circuit. When the system crashes, restarts abnormally, or experiences severe electromagnetic interference, the purely software-based protection logic is prone to failure, leading to the relay being accidentally re-closed. Addressing this hidden safety vulnerability, this embodiment constructs a low-level physical interlocking mechanism with dual-path triggering (software alarm / hardware limit violation) and complete independence from software control. Once an anomaly is triggered, the low-level hardware latching circuit directly takes over the disconnected state of the main circuit and maintains it. At this point, even if the upper-level diagnostic model or control software crashes or resets, the safe disconnection state of the main circuit will not be unexpectedly revoked. This interlocking state requires strict on-site manual authorization to be released, thus establishing a low-level hardware safety defense for the battery stack system in complex physical deployment environments.

[0109] As a specific circuit implementation of the above-mentioned hardware-level mandatory protection steps, please refer to [link / reference]. Figure 4 The underlying hardware control logic includes a hardware voltage comparator 510, a main control microcontroller 520, an RS flip-flop 530 (as the core component of the hardware latching circuit), a relay drive circuit 540, and a solid-state relay 550.

[0110] S710, Forcefully disconnect the main charging and discharging circuit. When a battery type mismatch alarm signal is triggered, the main control microcontroller 520 outputs a protection trigger signal; the hardware voltage comparator 510 compares the total voltage of the battery stack or the voltage of a single cell with a preset threshold, and outputs a high-level signal when a voltage abnormality is detected; the above protection trigger signal and voltage abnormality signal are respectively input to the two input terminals of the OR gate 515, and after the OR operation, the output terminal of the OR gate 515 outputs a high-level signal to the set terminal (S terminal) of the RS flip-flop 530, that is, the set is completed as long as either of the two trigger sources is valid. The OR gate 515 is used to perform an OR operation on the protection trigger signal and the voltage abnormality signal, and the result of the OR operation is used as the set trigger signal of the RS flip-flop 530; the RS flip-flop 530 adopts a high-level set and high-level reset latch circuit, and its S terminal and R terminal are not simultaneously high to avoid the inhibited state. The output terminal (Q terminal) of the RS flip-flop 530 outputs a high level, which drives the solid-state relay 550 to forcibly disconnect the main circuit through the relay drive circuit 540. For the main contactor scheme, the relay drive circuit 540 is replaced with the contactor coil drive circuit, and the main contactor forms a self-holding circuit through auxiliary contacts.

[0111] S720 enters the hardware-level physical latching state. After the main circuit is disconnected, the RS flip-flop 530 forms a latching path with self-holding characteristics, and its output state is maintained during power supply maintenance. Even if the upper-level software crashes or restarts, the solid-state relay remains disconnected;

[0112] S730 Unlocking Conditions. Simultaneously, the following conditions must be met: On-site maintenance personnel must complete verification on the local operation interface with valid on-site authorization information and generate a matching on-site authorization verification command. After the system self-test passes, the main control microcontroller 520 sends a high-level release pulse to the reset terminal (R terminal) of the RS flip-flop 530, causing the RS flip-flop 530 to flip its state, thus restoring the normal controllable state of the solid-state relay 550. After release, the main control microcontroller 520 re-sends a normal control signal, driving the solid-state relay 550 to close or the main contactor to engage, restoring the main circuit conduction.

[0113] The results are synchronized to the cloud and the display interface, while the underlying layer maintains independent security monitoring.

[0114] In some embodiments, control experiments show that, based on three months of operational data (including periods of impact vibration such as piling and compaction) collected in the same battery swapping station deployment environment, the false alarm rate of the complete solution of the present invention is 2.3%, the false alarm rate is 1.5%, and the mean absolute error of SOH is 2.8%. The control solution (without vibration suppression and without BIM environmental perception) has a false alarm rate of 11.7%, a false alarm rate of 8.2%, and a mean absolute error of SOH of 7.4%. Here, the false alarm rate refers to the proportion of faults misjudged during fault-free periods, and the false alarm rate refers to the proportion of faults not identified during actual fault periods.

[0115] plan False alarm rate underreporting rate SOH mean absolute error Comparison scheme (no vibration suppression, no BIM awareness) 11.7% 8.2% 7.4% Vibration suppression only 5.1% 4.8% 4.6% BIM environment perception only 7.3% 5.5% 5.2% Complete solution of the present invention 2.3% 1.5% 2.8%

[0116] In some embodiments, equivalent alternatives may be used for generating mechanical vibration intensity metrics and masks:

[0117] Alternative Solution 1 (Peak Acceleration Method): Using the maximum absolute value of acceleration within a time segment. As a measure of mechanical vibration intensity, the corresponding suppression formula is:

[0118] Experimental determination , .

[0119] Alternative Solution 2 (Frequency Domain Energy Method): Calculate the integral value of the power spectral density within a preset frequency band (e.g., 1Hz to 100Hz). As a measure of mechanical vibration intensity, the corresponding suppression formula is:

[0120] parameter , (like , Through actual measurement and calibration, all of the above schemes can maintain a high kurtosis characteristic that is almost completely preserved when there is no vibration.

[0121] In some embodiments, the correspondence between progressive degradation signals and battery types is as follows: For lead-acid battery stacks, progressive degradation signals include an increasing trend in DC internal resistance (slope greater than 0 and cumulative internal resistance increase exceeding a preset internal resistance growth threshold) and abnormal voltage shift in the plateau region (manifested as an abnormal rise in the voltage plateau at the end of charging); for lithium-ion battery stacks, progressive degradation signals include a continuously decreasing trend in coulombic efficiency (slope less than 0 and cumulative coulombic efficiency decrease exceeding a preset decrease threshold) and abnormal voltage shift in the plateau region (manifested as an abnormal drop in the plateau region voltage). Here, the cumulative increase in internal resistance refers to the increase in current internal resistance relative to the initial internal resistance of the long-period sequence; the cumulative decrease in coulombic efficiency refers to the decrease in current coulombic efficiency relative to the initial coulombic efficiency of the sliding window; and the reference plateau region voltage in the abnormal voltage shift in the plateau region is the initially calibrated reference plateau region voltage. As an example, the internal resistance growth threshold can be calibrated by determining the statistical distribution of the internal resistance change slope through aging experiments with batteries of the same model; the cumulative decrease in coulombic efficiency threshold can be set to 3%–5%; and the plateau region voltage shift threshold can be set to 1%–2% of the reference plateau region voltage.

[0122] The correspondence between progressive degradation signals and fault categories output by the classification head is shown in the table below:

[0123] Progressive degradation signal Corresponding fault category DC internal resistance increasing trend DC internal resistance abnormality Coulomb efficiency continues to decline Coulomb efficiency anomaly abnormal voltage deviation in the platform area Platform voltage offset Non-progressive abnormalities Over-temperature, over-voltage / under-voltage

[0124] In some embodiments, the inference process deployment scheme of the analysis model is as follows: Running data is captured using a sliding time window (window length 100 sampling points corresponding to 100 seconds, sliding step size 50 sampling points). The model is deployed on an edge computing device after INT8 fixed-point quantization. When the BIM platform disconnects, a local cache fault tolerance mechanism is activated. The latest cached spatial physical environment characteristics are read locally, and the estimated environmental temperature in the cache is replaced with the actual temperature measured by the local environmental temperature sensor to maintain degraded operation. Simultaneously, a degraded flag and a cache timestamp are appended to the output health status estimation result. If there are no historical cached records locally, the system runs with preset default environmental parameters and a degraded flag is appended.

[0125] The steps S100 to S700 described above together constitute a complete health status diagnosis inference data flow. The data shape, generation method and interrelationship involved in each step have been described in detail in the corresponding embodiments above, and will not be repeated here.

[0126] The technical features of the embodiments described above can be combined arbitrarily, and as long as the combination does not contradict each other, it should be considered within the scope of this specification. For those skilled in the art, various modifications and improvements can be made without departing from the concept of this invention, and these all fall within the protection scope of this invention.

Claims

1. A health monitoring method for battery stacks, characterized in that, include: Acquire electrochemical characteristic data of the battery stack within a preset time interval, and acquire mechanical vibration intensity data synchronized with the electrochemical characteristic data; Based on the mechanical vibration intensity data, the electrochemical characteristic data is subjected to dynamic influence weight suppression to obtain suppressed features; wherein, the suppression intensity of the dynamic influence weight suppression increases with the increase of the mechanical vibration intensity data, so as to attenuate or shield the feature weights of time segments with transient mechanical vibration noise in at least the multiple time segments divided within the preset time interval; Acquire spatial physical environment characteristics that characterize the current three-dimensional deployment location of the battery stack; Based on the characteristics of the space physical environment, a diagnostic sensitivity adjustment factor is generated to control the sensitivity to progressive degradation signals; wherein, the progressive degradation signal is an electrochemical characteristic change signal characterizing the gradual deterioration of the battery stack over time; and The suppressed features are processed using an analysis model configured with the diagnostic sensitivity adjustment factor. The diagnostic sensitivity adjustment factor is used to scale the hidden features extracted by the analysis model, and the health status estimation result of the battery stack is generated and output.

2. The health monitoring method for battery stacks according to claim 1, characterized in that, The method further includes: The cumulative charge capacity is obtained by integrating the current data of the charging cycle in the electrochemical characteristic data over time, and a capacity-voltage curve is generated by combining it with the synchronous voltage data. By performing a differential operation on the capacity-voltage curve, the data interval in which the absolute value of the voltage-capacity differential is less than a preset threshold for a consecutive preset number of sampling points is identified as a voltage plateau region, and the corresponding plateau region voltage and plateau region width are extracted; wherein, the plateau region width is the capacity change corresponding to the voltage plateau region, and the plateau region voltage is the average voltage within the voltage plateau region.

3. The health monitoring method for battery stacks according to claim 1, characterized in that, The step of dynamically weighting and suppressing the electrochemical characteristic data based on the mechanical vibration intensity data to obtain the suppressed features includes: Within the time interval, the mechanical vibration intensity data is divided into multiple consecutive time segments, and the mechanical vibration intensity metric value within each time segment is calculated. Based on the mechanical vibration intensity metric value within the time segment, a one-dimensional attention suppression mask is generated; Based on the mechanical vibration intensity data and the electrochemical feature data, correlation feature calculation is performed to generate an original attention weight matrix characterizing the correlation between each time segment and the electrochemical feature; Using the one-dimensional attention suppression mask, the weights of the time-series segments containing the transient mechanical vibration noise in the original attention weight matrix are attenuated or masked to obtain the suppressed attention weight matrix. The suppressed attention weight matrix is ​​fused with the value matrix obtained by encoding the electrochemical feature data and copying and expanding it according to a preset time correspondence to obtain a fused feature matrix. Based on the one-dimensional attention suppression mask, the fused feature matrix is ​​subjected to masked global average pooling along the time dimension to obtain a fused feature vector, which is used as the suppressed feature. The step of generating a one-dimensional attention suppression mask based on the mechanical vibration intensity metric within the time segment includes: Calculate the root mean square value of the mechanical vibration intensity data within each time segment, and arrange the root mean square values ​​according to time sequence; Using the natural constant The inverse logistic decay function with base is used to calculate a sequence of suppression coefficients based on the root mean square value, a preset suppression threshold, and a smoothing factor to control the steepness of the transition, thus forming the one-dimensional attention suppression mask; wherein, the expression for the inverse logistic decay function is: ； in, For the first The suppression coefficient of each time segment, For the first The root mean square value of each time segment The preset suppression threshold, A smoothing factor to control the steepness of the transition; when When approaching zero, A value close to 1 indicates that the feature weights of the corresponding time segment are fully preserved; when Significantly greater than hour, A value close to 0 indicates that the feature weights of the corresponding time segment are masked.

4. The health monitoring method for a battery stack according to claim 3, characterized in that, The step of performing correlation feature calculation based on the mechanical vibration intensity data and the electrochemical feature data to generate an original attention weight matrix characterizing the correlation between each time segment and the electrochemical feature includes: encoding the mechanical vibration intensity data through a one-dimensional convolutional neural network to obtain a mechanical vibration time-series feature matrix, which serves as a query matrix for cross-attention; copying and expanding the encoded electrochemical feature data according to a preset time correspondence to obtain a bond matrix and a value matrix; performing a scaled dot product operation between the query matrix and the bond matrix and then normalizing it using Softmax to generate the original attention weight matrix. The step of using the one-dimensional attention suppression mask to attenuate or mask the weights of time segments with transient mechanical vibration noise in the original attention weight matrix to obtain a suppressed attention weight matrix includes: broadcasting and expanding the one-dimensional attention suppression mask along the column direction, and then performing a Hadamard product with the original attention weight matrix to attenuate or mask the corresponding attention weights of time segments with transient mechanical vibration noise during polymerization electrochemical characteristics, thereby obtaining the suppressed attention weight matrix.

5. A health monitoring method for a battery stack according to claim 3 or 4, characterized in that, The acquisition of spatial physical environment features characterizing the current three-dimensional deployment location of the battery stack includes: A request containing the spatial location identifier of the battery stack is sent to a remote building information modeling platform via a data interface; The system receives the spatial physical environment features returned by the building information modeling platform based on the call request; wherein the spatial physical environment features include at least an estimated ambient temperature, a ventilation condition status code, a physical obstruction coefficient, and a spatial relative position scalar.

6. A health monitoring method for a battery stack according to claim 5, characterized in that, The step of generating a diagnostic sensitivity adjustment factor based on the characteristics of the space physical environment to control the sensitivity to progressive degradation signals includes: The spatial physical environment features are input into the environmental encoder to generate the diagnostic judgment sensitivity adjustment factor; wherein, the diagnostic judgment sensitivity adjustment factor is a scalar. Among the multiple encoder layers included in the analysis model, a set of candidate encoder layers is determined whose layer numbers fall within the range of one-third to two-thirds of the total number of layers. When the endpoints of the range correspond to non-integer layer numbers, the lower limit is rounded up and the upper limit is rounded down. The layer with the middle number is selected from the set of candidate encoder layers as the target injection layer. When the set of candidate encoder layers contains an even number of layer numbers, the middle layer number that is slightly later is selected. In the target injection layer, the diagnostic judgment sensitivity adjustment factor is used to perform element-wise scaling modulation on the hidden state feature tensor output by the multi-head self-attention sublayer of the target injection layer, so as to control the judgment sensitivity to the progressive degradation signal.

7. A health monitoring method for a battery stack according to claim 6, characterized in that, The step of using an analytical model configured with the diagnostic sensitivity adjustment factor to process the suppressed features and generate and output the health status estimation result of the battery stack includes: The suppressed features are copied and expanded along the time dimension to form a temporal sequence; The number of sampling times at which the one-dimensional attention suppression mask is downsampled to the electrochemical feature data is used as the gating weight; Using the gating weights, the electrochemical characteristic data after moving average smoothing is subjected to element-wise multiplication weighting. The weighted electrochemical feature data is concatenated with the time series along the feature dimension to construct a multi-channel input tensor; The multi-channel input tensor is mapped to an initial hidden sequence through an input linear projection layer, and after encoding at the superposition position, it is input into the multi-level cascaded encoder contained in the analysis model. Features are extracted layer by layer using the multi-level cascaded encoder; In the target injection layer, the element-wise scaling modulation is performed on the hidden state feature tensor output by the multi-head self-attention sub-layer using the diagnostic judgment sensitivity adjustment factor; Global average pooling is performed on the feature sequence output by the multi-stage cascaded encoder to obtain a comprehensive feature vector. The comprehensive feature vector is input into the regression head and classification head of the analysis model respectively to generate a comprehensive prediction result containing continuous numerical values ​​of battery stack health status and probability distribution of fault categories, which is used as the health status estimation result of the battery stack.

8. A health monitoring method for a battery stack according to claim 2, characterized in that, The battery stack is either a lead-acid battery stack or a lithium-ion battery stack. When the battery stack is a lead-acid battery stack, the progressive degradation signal includes at least one of the following detection items: DC internal resistance increasing trend: By calling the local or cloud database through the data interface, a long periodic sequence composed of multiple consecutive historical internal resistance measurements is obtained. The slope of internal resistance change is calculated by linear regression. When the slope is greater than zero and the cumulative increase in internal resistance exceeds the preset internal resistance growth threshold, it is determined that a DC internal resistance increasing trend degradation signal has appeared. Abnormal deviation of platform voltage: Extract the platform voltage of the current charging cycle and compare it with the pre-calibrated reference platform voltage. When the deviation of the platform voltage from the reference platform voltage exceeds a preset deviation threshold, it is determined that an abnormal deviation degradation signal of platform voltage has occurred. When the battery stack is a lithium-ion battery stack, the progressive degradation signal includes at least one of the following detection items: Coulomb efficiency continuously decreasing trend: By calling the local or cloud database through the data interface, a long period sequence composed of coulomb efficiencies of multiple consecutive historical charge-discharge cycles is obtained. The slope of coulomb efficiency change is calculated by linear regression. When the slope is less than zero and the cumulative decrease in coulomb efficiency exceeds the preset decrease threshold, it is determined that a degradation signal of continuous decrease in coulomb efficiency has occurred. Abnormal deviation of platform voltage: Extract the platform voltage of the current charging cycle and compare it with the pre-calibrated reference platform voltage. When the deviation of the platform voltage from the reference platform voltage exceeds a preset deviation threshold, it is determined that an abnormal deviation degradation signal of platform voltage has occurred.

9. A health monitoring method for a battery stack according to claim 6, characterized in that, The method also includes a fault-tolerant degradation step: During normal system operation, the spatial physical environment features successfully acquired each time are persistently stored locally, and the corresponding cache timestamp is recorded. When the call request sent to the building information modeling platform times out or the connection is abnormal, the latest cached spatial physical environment features are read locally, and the estimated environmental temperature in the cache is replaced with the measured temperature of the environmental temperature sensor pre-deployed locally on the battery stack to obtain the fused and corrected spatial physical environment features. The fused and corrected spatial physical environment features are input into the environment encoder to generate the diagnostic judgment sensitivity adjustment factor, thereby maintaining the continuous operation of the diagnostic process. A downgrade flag and the cache timestamp are appended to the health status estimation result.

10. A health monitoring method for a battery stack according to claim 2, characterized in that, The method further includes a step of dynamically verifying the type of the battery stack: The current feature vector, composed of the platform region voltage and the platform region width, is normalized according to preset calibration parameters, and then the normalized Euclidean distance is calculated with the pre-stored standard battery operation feature vector. When the normalized Euclidean distance exceeds the consistency tolerance threshold for a certain number of consecutive preset times, a battery type verification inconsistency alarm signal is triggered.

Images